mirrors
/
linux-stable
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
1
0
Fork 0
Code Issues Releases Wiki Activity
linux-stable/include/linux/atomic/atomic-arch-fallback.h

4694 lines
126 KiB
C
Raw Normal View History

locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
// SPDX-License-Identifier: GPL-2.0
// Generated by scripts/atomic/gen-atomic-fallback.sh
// DO NOT MODIFY THIS FILE DIRECTLY
#ifndef _LINUX_ATOMIC_FALLBACK_H
#define _LINUX_ATOMIC_FALLBACK_H
#include <linux/compiler.h>
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_xchg)
#define raw_xchg arch_xchg
#elif defined(arch_xchg_relaxed)
#define raw_xchg(...) \
__atomic_op_fence(arch_xchg, __VA_ARGS__)
#else
extern void raw_xchg_not_implemented(void);
#define raw_xchg(...) raw_xchg_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_xchg_acquire)
#define raw_xchg_acquire arch_xchg_acquire
#elif defined(arch_xchg_relaxed)
#define raw_xchg_acquire(...) \
__atomic_op_acquire(arch_xchg, __VA_ARGS__)
#elif defined(arch_xchg)
#define raw_xchg_acquire arch_xchg
#else
extern void raw_xchg_acquire_not_implemented(void);
#define raw_xchg_acquire(...) raw_xchg_acquire_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_xchg_release)
#define raw_xchg_release arch_xchg_release
#elif defined(arch_xchg_relaxed)
#define raw_xchg_release(...) \
__atomic_op_release(arch_xchg, __VA_ARGS__)
#elif defined(arch_xchg)
#define raw_xchg_release arch_xchg
#else
extern void raw_xchg_release_not_implemented(void);
#define raw_xchg_release(...) raw_xchg_release_not_implemented()
#endif
#if defined(arch_xchg_relaxed)
#define raw_xchg_relaxed arch_xchg_relaxed
#elif defined(arch_xchg)
#define raw_xchg_relaxed arch_xchg
#else
extern void raw_xchg_relaxed_not_implemented(void);
#define raw_xchg_relaxed(...) raw_xchg_relaxed_not_implemented()
#endif
#if defined(arch_cmpxchg)
#define raw_cmpxchg arch_cmpxchg
#elif defined(arch_cmpxchg_relaxed)
#define raw_cmpxchg(...) \
__atomic_op_fence(arch_cmpxchg, __VA_ARGS__)
#else
extern void raw_cmpxchg_not_implemented(void);
#define raw_cmpxchg(...) raw_cmpxchg_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg_acquire)
#define raw_cmpxchg_acquire arch_cmpxchg_acquire
#elif defined(arch_cmpxchg_relaxed)
#define raw_cmpxchg_acquire(...) \
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_op_acquire(arch_cmpxchg, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg)
#define raw_cmpxchg_acquire arch_cmpxchg
#else
extern void raw_cmpxchg_acquire_not_implemented(void);
#define raw_cmpxchg_acquire(...) raw_cmpxchg_acquire_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg_release)
#define raw_cmpxchg_release arch_cmpxchg_release
#elif defined(arch_cmpxchg_relaxed)
#define raw_cmpxchg_release(...) \
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_op_release(arch_cmpxchg, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg)
#define raw_cmpxchg_release arch_cmpxchg
#else
extern void raw_cmpxchg_release_not_implemented(void);
#define raw_cmpxchg_release(...) raw_cmpxchg_release_not_implemented()
#endif
#if defined(arch_cmpxchg_relaxed)
#define raw_cmpxchg_relaxed arch_cmpxchg_relaxed
#elif defined(arch_cmpxchg)
#define raw_cmpxchg_relaxed arch_cmpxchg
#else
extern void raw_cmpxchg_relaxed_not_implemented(void);
#define raw_cmpxchg_relaxed(...) raw_cmpxchg_relaxed_not_implemented()
#endif
#if defined(arch_cmpxchg64)
#define raw_cmpxchg64 arch_cmpxchg64
#elif defined(arch_cmpxchg64_relaxed)
#define raw_cmpxchg64(...) \
__atomic_op_fence(arch_cmpxchg64, __VA_ARGS__)
#else
extern void raw_cmpxchg64_not_implemented(void);
#define raw_cmpxchg64(...) raw_cmpxchg64_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg64_acquire)
#define raw_cmpxchg64_acquire arch_cmpxchg64_acquire
#elif defined(arch_cmpxchg64_relaxed)
#define raw_cmpxchg64_acquire(...) \
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_op_acquire(arch_cmpxchg64, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg64)
#define raw_cmpxchg64_acquire arch_cmpxchg64
#else
extern void raw_cmpxchg64_acquire_not_implemented(void);
#define raw_cmpxchg64_acquire(...) raw_cmpxchg64_acquire_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg64_release)
#define raw_cmpxchg64_release arch_cmpxchg64_release
#elif defined(arch_cmpxchg64_relaxed)
#define raw_cmpxchg64_release(...) \
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_op_release(arch_cmpxchg64, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg64)
#define raw_cmpxchg64_release arch_cmpxchg64
#else
extern void raw_cmpxchg64_release_not_implemented(void);
#define raw_cmpxchg64_release(...) raw_cmpxchg64_release_not_implemented()
#endif
#if defined(arch_cmpxchg64_relaxed)
#define raw_cmpxchg64_relaxed arch_cmpxchg64_relaxed
#elif defined(arch_cmpxchg64)
#define raw_cmpxchg64_relaxed arch_cmpxchg64
#else
extern void raw_cmpxchg64_relaxed_not_implemented(void);
#define raw_cmpxchg64_relaxed(...) raw_cmpxchg64_relaxed_not_implemented()
#endif
#if defined(arch_cmpxchg128)
#define raw_cmpxchg128 arch_cmpxchg128
#elif defined(arch_cmpxchg128_relaxed)
#define raw_cmpxchg128(...) \
__atomic_op_fence(arch_cmpxchg128, __VA_ARGS__)
#else
extern void raw_cmpxchg128_not_implemented(void);
#define raw_cmpxchg128(...) raw_cmpxchg128_not_implemented()
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg128_acquire)
#define raw_cmpxchg128_acquire arch_cmpxchg128_acquire
#elif defined(arch_cmpxchg128_relaxed)
#define raw_cmpxchg128_acquire(...) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
__atomic_op_acquire(arch_cmpxchg128, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg128)
#define raw_cmpxchg128_acquire arch_cmpxchg128
#else
extern void raw_cmpxchg128_acquire_not_implemented(void);
#define raw_cmpxchg128_acquire(...) raw_cmpxchg128_acquire_not_implemented()
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_cmpxchg128_release)
#define raw_cmpxchg128_release arch_cmpxchg128_release
#elif defined(arch_cmpxchg128_relaxed)
#define raw_cmpxchg128_release(...) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
__atomic_op_release(arch_cmpxchg128, __VA_ARGS__)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_cmpxchg128)
#define raw_cmpxchg128_release arch_cmpxchg128
#else
extern void raw_cmpxchg128_release_not_implemented(void);
#define raw_cmpxchg128_release(...) raw_cmpxchg128_release_not_implemented()
#endif
#if defined(arch_cmpxchg128_relaxed)
#define raw_cmpxchg128_relaxed arch_cmpxchg128_relaxed
#elif defined(arch_cmpxchg128)
#define raw_cmpxchg128_relaxed arch_cmpxchg128
#else
extern void raw_cmpxchg128_relaxed_not_implemented(void);
#define raw_cmpxchg128_relaxed(...) raw_cmpxchg128_relaxed_not_implemented()
#endif
#if defined(arch_try_cmpxchg)
#define raw_try_cmpxchg arch_try_cmpxchg
#elif defined(arch_try_cmpxchg_relaxed)
#define raw_try_cmpxchg(...) \
__atomic_op_fence(arch_try_cmpxchg, __VA_ARGS__)
#else
#define raw_try_cmpxchg(_ptr, _oldp, _new) \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg((_ptr), ___o, (_new)); \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg_acquire)
#define raw_try_cmpxchg_acquire arch_try_cmpxchg_acquire
#elif defined(arch_try_cmpxchg_relaxed)
#define raw_try_cmpxchg_acquire(...) \
__atomic_op_acquire(arch_try_cmpxchg, __VA_ARGS__)
#elif defined(arch_try_cmpxchg)
#define raw_try_cmpxchg_acquire arch_try_cmpxchg
#else
#define raw_try_cmpxchg_acquire(_ptr, _oldp, _new) \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg_acquire((_ptr), ___o, (_new)); \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg_release)
#define raw_try_cmpxchg_release arch_try_cmpxchg_release
#elif defined(arch_try_cmpxchg_relaxed)
#define raw_try_cmpxchg_release(...) \
__atomic_op_release(arch_try_cmpxchg, __VA_ARGS__)
#elif defined(arch_try_cmpxchg)
#define raw_try_cmpxchg_release arch_try_cmpxchg
#else
#define raw_try_cmpxchg_release(_ptr, _oldp, _new) \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg_release((_ptr), ___o, (_new)); \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg_relaxed)
#define raw_try_cmpxchg_relaxed arch_try_cmpxchg_relaxed
#elif defined(arch_try_cmpxchg)
#define raw_try_cmpxchg_relaxed arch_try_cmpxchg
#else
#define raw_try_cmpxchg_relaxed(_ptr, _oldp, _new) \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg_relaxed((_ptr), ___o, (_new)); \
asm-generic/atomic: Add try_cmpxchg() fallbacks Only x86 provides try_cmpxchg() outside of the atomic_t interfaces, provide generic fallbacks to create this interface from the widely available cmpxchg() function. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Will Deacon <will@kernel.org> Link: https://lore.kernel.org/r/159870621515.1229682.15506193091065001742.stgit@devnote2
2020-08-29 13:03:35 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg64)
#define raw_try_cmpxchg64 arch_try_cmpxchg64
#elif defined(arch_try_cmpxchg64_relaxed)
#define raw_try_cmpxchg64(...) \
__atomic_op_fence(arch_try_cmpxchg64, __VA_ARGS__)
#else
#define raw_try_cmpxchg64(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg64((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg64_acquire)
#define raw_try_cmpxchg64_acquire arch_try_cmpxchg64_acquire
#elif defined(arch_try_cmpxchg64_relaxed)
#define raw_try_cmpxchg64_acquire(...) \
__atomic_op_acquire(arch_try_cmpxchg64, __VA_ARGS__)
#elif defined(arch_try_cmpxchg64)
#define raw_try_cmpxchg64_acquire arch_try_cmpxchg64
#else
#define raw_try_cmpxchg64_acquire(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg64_acquire((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg64_release)
#define raw_try_cmpxchg64_release arch_try_cmpxchg64_release
#elif defined(arch_try_cmpxchg64_relaxed)
#define raw_try_cmpxchg64_release(...) \
__atomic_op_release(arch_try_cmpxchg64, __VA_ARGS__)
#elif defined(arch_try_cmpxchg64)
#define raw_try_cmpxchg64_release arch_try_cmpxchg64
#else
#define raw_try_cmpxchg64_release(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg64_release((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg64_relaxed)
#define raw_try_cmpxchg64_relaxed arch_try_cmpxchg64_relaxed
#elif defined(arch_try_cmpxchg64)
#define raw_try_cmpxchg64_relaxed arch_try_cmpxchg64
#else
#define raw_try_cmpxchg64_relaxed(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg64_relaxed((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg64 support Add generic support for try_cmpxchg64{,_acquire,_release,_relaxed} and their falbacks involving cmpxchg64. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Link: https://lkml.kernel.org/r/20220515184205.103089-2-ubizjak@gmail.com
2022-05-15 18:42:03 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg128)
#define raw_try_cmpxchg128 arch_try_cmpxchg128
#elif defined(arch_try_cmpxchg128_relaxed)
#define raw_try_cmpxchg128(...) \
__atomic_op_fence(arch_try_cmpxchg128, __VA_ARGS__)
#else
#define raw_try_cmpxchg128(_ptr, _oldp, _new) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg128((_ptr), ___o, (_new)); \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg128_acquire)
#define raw_try_cmpxchg128_acquire arch_try_cmpxchg128_acquire
#elif defined(arch_try_cmpxchg128_relaxed)
#define raw_try_cmpxchg128_acquire(...) \
__atomic_op_acquire(arch_try_cmpxchg128, __VA_ARGS__)
#elif defined(arch_try_cmpxchg128)
#define raw_try_cmpxchg128_acquire arch_try_cmpxchg128
#else
#define raw_try_cmpxchg128_acquire(_ptr, _oldp, _new) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg128_acquire((_ptr), ___o, (_new)); \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg128_release)
#define raw_try_cmpxchg128_release arch_try_cmpxchg128_release
#elif defined(arch_try_cmpxchg128_relaxed)
#define raw_try_cmpxchg128_release(...) \
__atomic_op_release(arch_try_cmpxchg128, __VA_ARGS__)
#elif defined(arch_try_cmpxchg128)
#define raw_try_cmpxchg128_release arch_try_cmpxchg128
#else
#define raw_try_cmpxchg128_release(_ptr, _oldp, _new) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg128_release((_ptr), ___o, (_new)); \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_try_cmpxchg128_relaxed)
#define raw_try_cmpxchg128_relaxed arch_try_cmpxchg128_relaxed
#elif defined(arch_try_cmpxchg128)
#define raw_try_cmpxchg128_relaxed arch_try_cmpxchg128
#else
#define raw_try_cmpxchg128_relaxed(_ptr, _oldp, _new) \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg128_relaxed((_ptr), ___o, (_new)); \
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#define raw_cmpxchg_local arch_cmpxchg_local
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#ifdef arch_try_cmpxchg_local
#define raw_try_cmpxchg_local arch_try_cmpxchg_local
#else
#define raw_try_cmpxchg_local(_ptr, _oldp, _new) \
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
___r = raw_cmpxchg_local((_ptr), ___o, (_new)); \
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
#endif
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#define raw_cmpxchg64_local arch_cmpxchg64_local
instrumentation: Wire up cmpxchg128() Wire up the cmpxchg128 family in the atomic wrapper scripts. These provide the generic cmpxchg128 family of functions from the arch_ prefixed version, adding explicit instrumentation where needed. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Arnd Bergmann <arnd@arndb.de> Reviewed-by: Mark Rutland <mark.rutland@arm.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Tested-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230531132323.519237070@infradead.org
2023-05-31 13:08:37 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#ifdef arch_try_cmpxchg64_local
#define raw_try_cmpxchg64_local arch_try_cmpxchg64_local
#else
#define raw_try_cmpxchg64_local(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg64_local((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
#define raw_cmpxchg128_local arch_cmpxchg128_local
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#ifdef arch_try_cmpxchg128_local
#define raw_try_cmpxchg128_local arch_try_cmpxchg128_local
#else
#define raw_try_cmpxchg128_local(_ptr, _oldp, _new) \
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
___r = raw_cmpxchg128_local((_ptr), ___o, (_new)); \
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#endif
#define raw_sync_cmpxchg arch_sync_cmpxchg
locking/atomic: Add generic try_cmpxchg{,64}_local() support Add generic support for try_cmpxchg{,64}_local() and their falbacks. These provides the generic try_cmpxchg_local family of functions from the arch_ prefixed version, also adding explicit instrumentation. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Ingo Molnar <mingo@kernel.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230405141710.3551-2-ubizjak@gmail.com Cc: Linus Torvalds <torvalds@linux-foundation.org>
2023-04-05 14:17:06 +00:00
locking/atomic: Add generic support for sync_try_cmpxchg() and its fallback Provide the generic sync_try_cmpxchg() function from the raw_ prefixed version, also adding explicit instrumentation. The patch amends existing scripts to generate sync_try_cmpxchg() locking primitive and its raw_sync_try_cmpxchg() fallback, while leaving existing macros from the try_cmpxchg() family unchanged. The target can define its own arch_sync_try_cmpxchg() to override the generic version of raw_sync_try_cmpxchg(). This allows the target to generate more optimal assembly than the generic version. Additionally, the patch renames two scripts to better reflect whet they really do. Signed-off-by: Uros Bizjak <ubizjak@gmail.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Cc: Will Deacon <will@kernel.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: linux-kernel@vger.kernel.org
2023-09-25 14:50:23 +00:00
#ifdef arch_sync_try_cmpxchg
#define raw_sync_try_cmpxchg arch_sync_try_cmpxchg
#else
#define raw_sync_try_cmpxchg(_ptr, _oldp, _new) \
({ \
typeof(*(_ptr)) *___op = (_oldp), ___o = *___op, ___r; \
___r = raw_sync_cmpxchg((_ptr), ___o, (_new)); \
if (unlikely(___r != ___o)) \
*___op = ___r; \
likely(___r == ___o); \
})
#endif
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_read() - atomic load with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically loads the value of @v with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_read() elsewhere.
*
* Return: The value loaded from @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_read(const atomic_t *v)
{
return arch_atomic_read(v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_read_acquire() - atomic load with acquire ordering
* @v: pointer to atomic_t
*
* Atomically loads the value of @v with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_read_acquire() elsewhere.
*
* Return: The value loaded from @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_read_acquire(const atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_read_acquire)
return arch_atomic_read_acquire(v);
#else
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
int ret;
if (__native_word(atomic_t)) {
ret = smp_load_acquire(&(v)->counter);
} else {
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = raw_atomic_read(v);
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
__atomic_acquire_fence();
}
return ret;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_set() - atomic set with relaxed ordering
* @v: pointer to atomic_t
* @i: int value to assign
*
* Atomically sets @v to @i with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_set() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_set(atomic_t *v, int i)
{
arch_atomic_set(v, i);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_set_release() - atomic set with release ordering
* @v: pointer to atomic_t
* @i: int value to assign
*
* Atomically sets @v to @i with release ordering.
*
* Safe to use in noinstr code; prefer atomic_set_release() elsewhere.
*
* Return: Nothing.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline void
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_set_release(atomic_t *v, int i)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_set_release)
arch_atomic_set_release(v, i);
#else
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
if (__native_word(atomic_t)) {
smp_store_release(&(v)->counter, i);
} else {
__atomic_release_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_set(v, i);
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add() - atomic add with relaxed ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_add() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_add(int i, atomic_t *v)
{
arch_atomic_add(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_return() - atomic add with full ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_add_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline int
raw_atomic_add_return(int i, atomic_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_return)
return arch_atomic_add_return(i, v);
#elif defined(arch_atomic_add_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret;
__atomic_pre_full_fence();
ret = arch_atomic_add_return_relaxed(i, v);
__atomic_post_full_fence();
return ret;
#else
#error "Unable to define raw_atomic_add_return"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_return_acquire() - atomic add with acquire ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_add_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_return_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_return_acquire)
return arch_atomic_add_return_acquire(i, v);
#elif defined(arch_atomic_add_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_add_return_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_add_return_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_return_release() - atomic add with release ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_add_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_return_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_return_release)
return arch_atomic_add_return_release(i, v);
#elif defined(arch_atomic_add_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_add_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_add_return_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_return_relaxed() - atomic add with relaxed ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_add_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_add_return_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_add_return_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_add_return_relaxed"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_add() - atomic add with full ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_add() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_add(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_add)
return arch_atomic_fetch_add(i, v);
#elif defined(arch_atomic_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_add_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_add"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_add_acquire() - atomic add with acquire ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_add_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_add_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_add_acquire)
return arch_atomic_fetch_add_acquire(i, v);
#elif defined(arch_atomic_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_fetch_add_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_add_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_add_release() - atomic add with release ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_add_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_add_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_add_release)
return arch_atomic_fetch_add_release(i, v);
#elif defined(arch_atomic_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_fetch_add_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_add_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_add_relaxed() - atomic add with relaxed ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_add_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_fetch_add_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_fetch_add_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_add_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_add_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub() - atomic subtract with relaxed ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_sub() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_sub(int i, atomic_t *v)
{
arch_atomic_sub(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub_return() - atomic subtract with full ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_sub_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_sub_return(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_sub_return)
return arch_atomic_sub_return(i, v);
#elif defined(arch_atomic_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_sub_return_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_sub_return"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub_return_acquire() - atomic subtract with acquire ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_sub_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_sub_return_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_sub_return_acquire)
return arch_atomic_sub_return_acquire(i, v);
#elif defined(arch_atomic_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_sub_return_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_sub_return_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub_return_release() - atomic subtract with release ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_sub_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_sub_return_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_sub_return_release)
return arch_atomic_sub_return_release(i, v);
#elif defined(arch_atomic_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_sub_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_sub_return_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub_return_relaxed() - atomic subtract with relaxed ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_sub_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_sub_return_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_sub_return_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_sub_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_sub_return_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_sub() - atomic subtract with full ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_sub() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_sub(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_sub)
return arch_atomic_fetch_sub(i, v);
#elif defined(arch_atomic_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_sub_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_sub"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_sub_acquire() - atomic subtract with acquire ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_sub_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_sub_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_sub_acquire)
return arch_atomic_fetch_sub_acquire(i, v);
#elif defined(arch_atomic_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_fetch_sub_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_sub_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_sub_release() - atomic subtract with release ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_sub_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_sub_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_sub_release)
return arch_atomic_fetch_sub_release(i, v);
#elif defined(arch_atomic_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_fetch_sub_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_sub_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_sub_relaxed() - atomic subtract with relaxed ordering
* @i: int value to subtract
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_sub_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_fetch_sub_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_fetch_sub_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_sub_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_sub_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc() - atomic increment with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_inc() elsewhere.
*
* Return: Nothing.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline void
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc)
arch_atomic_inc(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_return() - atomic increment with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_inc_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline int
raw_atomic_inc_return(atomic_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_return)
return arch_atomic_inc_return(v);
#elif defined(arch_atomic_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret;
__atomic_pre_full_fence();
ret = arch_atomic_inc_return_relaxed(v);
__atomic_post_full_fence();
return ret;
#else
return raw_atomic_add_return(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_return_acquire() - atomic increment with acquire ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_inc_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_return_acquire(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_return_acquire)
return arch_atomic_inc_return_acquire(v);
#elif defined(arch_atomic_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_inc_return_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_inc_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_inc_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_add_return_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_return_release() - atomic increment with release ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_inc_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_return_release(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_return_release)
return arch_atomic_inc_return_release(v);
#elif defined(arch_atomic_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_inc_return_relaxed(v);
#elif defined(arch_atomic_inc_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_inc_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_add_return_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_return_relaxed() - atomic increment with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_inc_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_return_relaxed(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_return_relaxed)
return arch_atomic_inc_return_relaxed(v);
#elif defined(arch_atomic_inc_return)
return arch_atomic_inc_return(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_add_return_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_inc() - atomic increment with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_inc() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_inc(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_inc)
return arch_atomic_fetch_inc(v);
#elif defined(arch_atomic_fetch_inc_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_inc_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_add(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_inc_acquire() - atomic increment with acquire ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_inc_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_inc_acquire(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_inc_acquire)
return arch_atomic_fetch_inc_acquire(v);
#elif defined(arch_atomic_fetch_inc_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_fetch_inc_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_fetch_inc)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_inc(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_add_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_inc_release() - atomic increment with release ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_inc_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline int
raw_atomic_fetch_inc_release(atomic_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_inc_release)
return arch_atomic_fetch_inc_release(v);
#elif defined(arch_atomic_fetch_inc_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_fetch_inc_relaxed(v);
#elif defined(arch_atomic_fetch_inc)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_inc(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_add_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_inc_relaxed() - atomic increment with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_inc_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_inc_relaxed(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_inc_relaxed)
return arch_atomic_fetch_inc_relaxed(v);
#elif defined(arch_atomic_fetch_inc)
return arch_atomic_fetch_inc(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_fetch_add_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec() - atomic decrement with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_dec() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline void
raw_atomic_dec(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec)
arch_atomic_dec(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_sub(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_return() - atomic decrement with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_dec_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_return(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_return)
return arch_atomic_dec_return(v);
#elif defined(arch_atomic_dec_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_dec_return_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_sub_return(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_return_acquire() - atomic decrement with acquire ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_dec_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_return_acquire(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_return_acquire)
return arch_atomic_dec_return_acquire(v);
#elif defined(arch_atomic_dec_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_dec_return_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_dec_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_dec_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_sub_return_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_return_release() - atomic decrement with release ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_dec_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_return_release(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_return_release)
return arch_atomic_dec_return_release(v);
#elif defined(arch_atomic_dec_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_dec_return_relaxed(v);
#elif defined(arch_atomic_dec_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_dec_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_sub_return_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_return_relaxed() - atomic decrement with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_dec_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_return_relaxed(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_return_relaxed)
return arch_atomic_dec_return_relaxed(v);
#elif defined(arch_atomic_dec_return)
return arch_atomic_dec_return(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_sub_return_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_dec() - atomic decrement with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_dec() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_dec(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_dec)
return arch_atomic_fetch_dec(v);
#elif defined(arch_atomic_fetch_dec_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_dec_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_sub(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_dec_acquire() - atomic decrement with acquire ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_dec_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_dec_acquire(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_dec_acquire)
return arch_atomic_fetch_dec_acquire(v);
#elif defined(arch_atomic_fetch_dec_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_fetch_dec_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_fetch_dec)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_dec(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_sub_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_dec_release() - atomic decrement with release ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_dec_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_dec_release(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_dec_release)
return arch_atomic_fetch_dec_release(v);
#elif defined(arch_atomic_fetch_dec_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_fetch_dec_relaxed(v);
#elif defined(arch_atomic_fetch_dec)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_dec(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_sub_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_dec_relaxed() - atomic decrement with relaxed ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_dec_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_dec_relaxed(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_dec_relaxed)
return arch_atomic_fetch_dec_relaxed(v);
#elif defined(arch_atomic_fetch_dec)
return arch_atomic_fetch_dec(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_fetch_sub_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_and() - atomic bitwise AND with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_and() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_and(int i, atomic_t *v)
{
arch_atomic_and(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_and() - atomic bitwise AND with full ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_and() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_and(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_and)
return arch_atomic_fetch_and(i, v);
#elif defined(arch_atomic_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_and_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_and"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_and_acquire() - atomic bitwise AND with acquire ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_and_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_and_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_and_acquire)
return arch_atomic_fetch_and_acquire(i, v);
#elif defined(arch_atomic_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_fetch_and_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_and_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_and_release() - atomic bitwise AND with release ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_and_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_and_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_and_release)
return arch_atomic_fetch_and_release(i, v);
#elif defined(arch_atomic_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_fetch_and_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_and_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_and_relaxed() - atomic bitwise AND with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_and_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_fetch_and_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_fetch_and_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_and_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_and_relaxed"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_andnot() - atomic bitwise AND NOT with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & ~@i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_andnot() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline void
raw_atomic_andnot(int i, atomic_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_andnot)
arch_atomic_andnot(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_and(~i, v);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_andnot() - atomic bitwise AND NOT with full ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & ~@i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_andnot() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_andnot(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_andnot)
return arch_atomic_fetch_andnot(i, v);
#elif defined(arch_atomic_fetch_andnot_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_andnot_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_and(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_andnot_acquire() - atomic bitwise AND NOT with acquire ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & ~@i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_andnot_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline int
raw_atomic_fetch_andnot_acquire(int i, atomic_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_andnot_acquire)
return arch_atomic_fetch_andnot_acquire(i, v);
#elif defined(arch_atomic_fetch_andnot_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_fetch_andnot_relaxed(i, v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_fetch_andnot)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_andnot(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_and_acquire(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_andnot_release() - atomic bitwise AND NOT with release ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & ~@i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_andnot_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_andnot_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_andnot_release)
return arch_atomic_fetch_andnot_release(i, v);
#elif defined(arch_atomic_fetch_andnot_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_fetch_andnot_relaxed(i, v);
#elif defined(arch_atomic_fetch_andnot)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_andnot(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_fetch_and_release(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_andnot_relaxed() - atomic bitwise AND NOT with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v & ~@i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_andnot_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_andnot_relaxed(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_andnot_relaxed)
return arch_atomic_fetch_andnot_relaxed(i, v);
#elif defined(arch_atomic_fetch_andnot)
return arch_atomic_fetch_andnot(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_fetch_and_relaxed(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_or() - atomic bitwise OR with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v | @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_or() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_or(int i, atomic_t *v)
{
arch_atomic_or(i, v);
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_or() - atomic bitwise OR with full ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v | @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_or() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_or(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_or)
return arch_atomic_fetch_or(i, v);
#elif defined(arch_atomic_fetch_or_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret;
__atomic_pre_full_fence();
ret = arch_atomic_fetch_or_relaxed(i, v);
__atomic_post_full_fence();
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_or"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_or_acquire() - atomic bitwise OR with acquire ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v | @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_or_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_or_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_or_acquire)
return arch_atomic_fetch_or_acquire(i, v);
#elif defined(arch_atomic_fetch_or_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_fetch_or_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_or_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_or_release() - atomic bitwise OR with release ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v | @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_or_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_or_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_or_release)
return arch_atomic_fetch_or_release(i, v);
#elif defined(arch_atomic_fetch_or_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_fetch_or_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_or_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_or_relaxed() - atomic bitwise OR with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v | @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_or_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_fetch_or_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_fetch_or_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_or_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_or_relaxed"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_xor() - atomic bitwise XOR with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v ^ @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_xor() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic_xor(int i, atomic_t *v)
{
arch_atomic_xor(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_xor() - atomic bitwise XOR with full ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v ^ @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_xor() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_xor(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_xor)
return arch_atomic_fetch_xor(i, v);
#elif defined(arch_atomic_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_fetch_xor_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_xor"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_xor_acquire() - atomic bitwise XOR with acquire ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v ^ @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_xor_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_xor_acquire(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_xor_acquire)
return arch_atomic_fetch_xor_acquire(i, v);
#elif defined(arch_atomic_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret = arch_atomic_fetch_xor_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_xor_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_xor_release() - atomic bitwise XOR with release ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v ^ @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_xor_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_xor_release(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_xor_release)
return arch_atomic_fetch_xor_release(i, v);
#elif defined(arch_atomic_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic_fetch_xor_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_xor_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_xor_relaxed() - atomic bitwise XOR with relaxed ordering
* @i: int value
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v ^ @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_fetch_xor_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline int
raw_atomic_fetch_xor_relaxed(int i, atomic_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic_fetch_xor_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_xor_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic_fetch_xor_relaxed"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_xchg() - atomic exchange with full ordering
* @v: pointer to atomic_t
* @new: int value to assign
*
* Atomically updates @v to @new with full ordering.
*
* Safe to use in noinstr code; prefer atomic_xchg() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic_xchg(atomic_t *v, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_xchg)
return arch_atomic_xchg(v, new);
#elif defined(arch_atomic_xchg_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
ret = arch_atomic_xchg_relaxed(v, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg(&v->counter, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_xchg_acquire() - atomic exchange with acquire ordering
* @v: pointer to atomic_t
* @new: int value to assign
*
* Atomically updates @v to @new with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_xchg_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline int
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic_xchg_acquire(atomic_t *v, int new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_xchg_acquire)
return arch_atomic_xchg_acquire(v, new);
#elif defined(arch_atomic_xchg_relaxed)
int ret = arch_atomic_xchg_relaxed(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_xchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_xchg(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg_acquire(&v->counter, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_xchg_release() - atomic exchange with release ordering
* @v: pointer to atomic_t
* @new: int value to assign
*
* Atomically updates @v to @new with release ordering.
*
* Safe to use in noinstr code; prefer atomic_xchg_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline int
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic_xchg_release(atomic_t *v, int new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_xchg_release)
return arch_atomic_xchg_release(v, new);
#elif defined(arch_atomic_xchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_xchg_relaxed(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic_xchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_xchg(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg_release(&v->counter, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_xchg_relaxed() - atomic exchange with relaxed ordering
* @v: pointer to atomic_t
* @new: int value to assign
*
* Atomically updates @v to @new with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_xchg_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_xchg_relaxed(atomic_t *v, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_xchg_relaxed)
return arch_atomic_xchg_relaxed(v, new);
#elif defined(arch_atomic_xchg)
return arch_atomic_xchg(v, new);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_xchg_relaxed(&v->counter, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_cmpxchg() - atomic compare and exchange with full ordering
* @v: pointer to atomic_t
* @old: int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_cmpxchg() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_cmpxchg(atomic_t *v, int old, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_cmpxchg)
return arch_atomic_cmpxchg(v, old, new);
#elif defined(arch_atomic_cmpxchg_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_cmpxchg_relaxed(v, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg(&v->counter, old, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_cmpxchg_acquire() - atomic compare and exchange with acquire ordering
* @v: pointer to atomic_t
* @old: int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with acquire ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_cmpxchg_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_cmpxchg_acquire(atomic_t *v, int old, int new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_cmpxchg_acquire)
return arch_atomic_cmpxchg_acquire(v, old, new);
#elif defined(arch_atomic_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int ret = arch_atomic_cmpxchg_relaxed(v, old, new);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg_acquire(&v->counter, old, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_cmpxchg_release() - atomic compare and exchange with release ordering
* @v: pointer to atomic_t
* @old: int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with release ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_cmpxchg_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_cmpxchg_release(atomic_t *v, int old, int new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_cmpxchg_release)
return arch_atomic_cmpxchg_release(v, old, new);
#elif defined(arch_atomic_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg_release(&v->counter, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_cmpxchg_relaxed() - atomic compare and exchange with relaxed ordering
* @v: pointer to atomic_t
* @old: int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with relaxed ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_cmpxchg_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_cmpxchg_relaxed(atomic_t *v, int old, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_cmpxchg_relaxed)
return arch_atomic_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic_cmpxchg)
return arch_atomic_cmpxchg(v, old, new);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_cmpxchg_relaxed(&v->counter, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_try_cmpxchg() - atomic compare and exchange with full ordering
* @v: pointer to atomic_t
* @old: pointer to int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_try_cmpxchg() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline bool
raw_atomic_try_cmpxchg(atomic_t *v, int *old, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_try_cmpxchg)
return arch_atomic_try_cmpxchg(v, old, new);
#elif defined(arch_atomic_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic_try_cmpxchg_relaxed(v, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic_cmpxchg(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_try_cmpxchg_acquire() - atomic compare and exchange with acquire ordering
* @v: pointer to atomic_t
* @old: pointer to int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with acquire ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_try_cmpxchg_acquire() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline bool
raw_atomic_try_cmpxchg_acquire(atomic_t *v, int *old, int new)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_try_cmpxchg_acquire)
return arch_atomic_try_cmpxchg_acquire(v, old, new);
#elif defined(arch_atomic_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret = arch_atomic_try_cmpxchg_relaxed(v, old, new);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_try_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_try_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic_cmpxchg_acquire(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_try_cmpxchg_release() - atomic compare and exchange with release ordering
* @v: pointer to atomic_t
* @old: pointer to int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with release ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_try_cmpxchg_release() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_try_cmpxchg_release(atomic_t *v, int *old, int new)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_try_cmpxchg_release)
return arch_atomic_try_cmpxchg_release(v, old, new);
#elif defined(arch_atomic_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_try_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic_try_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_try_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic_cmpxchg_release(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_try_cmpxchg_relaxed() - atomic compare and exchange with relaxed ordering
* @v: pointer to atomic_t
* @old: pointer to int value to compare with
* @new: int value to assign
*
* If (@v == @old), atomically updates @v to @new with relaxed ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_try_cmpxchg_relaxed() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_try_cmpxchg_relaxed(atomic_t *v, int *old, int new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_try_cmpxchg_relaxed)
return arch_atomic_try_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic_try_cmpxchg)
return arch_atomic_try_cmpxchg(v, old, new);
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
int r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic_cmpxchg_relaxed(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_sub_and_test() - atomic subtract and test if zero with full ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_sub_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_sub_and_test(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_sub_and_test)
return arch_atomic_sub_and_test(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_sub_return(i, v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_and_test() - atomic decrement and test if zero with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_dec_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_and_test(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_and_test)
return arch_atomic_dec_and_test(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_dec_return(v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_and_test() - atomic increment and test if zero with full ordering
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_inc_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_and_test(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_and_test)
return arch_atomic_inc_and_test(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_inc_return(v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_negative() - atomic add and test if negative with full ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic_add_negative() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_negative(int i, atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_negative)
return arch_atomic_add_negative(i, v);
#elif defined(arch_atomic_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret;
__atomic_pre_full_fence();
ret = arch_atomic_add_negative_relaxed(i, v);
__atomic_post_full_fence();
return ret;
#else
return raw_atomic_add_return(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_negative_acquire() - atomic add and test if negative with acquire ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic_add_negative_acquire() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_negative_acquire(int i, atomic_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_negative_acquire)
return arch_atomic_add_negative_acquire(i, v);
#elif defined(arch_atomic_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret = arch_atomic_add_negative_relaxed(i, v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic_add_negative)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_negative(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_add_return_acquire(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_negative_release() - atomic add and test if negative with release ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic_add_negative_release() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_negative_release(int i, atomic_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_negative_release)
return arch_atomic_add_negative_release(i, v);
#elif defined(arch_atomic_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic_add_negative_relaxed(i, v);
#elif defined(arch_atomic_add_negative)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic_add_negative(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic_add_return_release(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_negative_relaxed() - atomic add and test if negative with relaxed ordering
* @i: int value to add
* @v: pointer to atomic_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic_add_negative_relaxed() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_negative_relaxed(int i, atomic_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_negative_relaxed)
return arch_atomic_add_negative_relaxed(i, v);
#elif defined(arch_atomic_add_negative)
return arch_atomic_add_negative(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_add_return_relaxed(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_fetch_add_unless() - atomic add unless value with full ordering
* @v: pointer to atomic_t
* @a: int value to add
* @u: int value to compare with
*
* If (@v != @u), atomically updates @v to (@v + @a) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_fetch_add_unless() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_fetch_add_unless(atomic_t *v, int a, int u)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_fetch_add_unless)
return arch_atomic_fetch_add_unless(v, a, u);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int c = raw_atomic_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c == u))
break;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic_try_cmpxchg(v, &c, c + a));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return c;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_add_unless() - atomic add unless value with full ordering
* @v: pointer to atomic_t
* @a: int value to add
* @u: int value to compare with
*
* If (@v != @u), atomically updates @v to (@v + @a) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_add_unless() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_add_unless(atomic_t *v, int a, int u)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_add_unless)
return arch_atomic_add_unless(v, a, u);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_fetch_add_unless(v, a, u) != u;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_not_zero() - atomic increment unless zero with full ordering
* @v: pointer to atomic_t
*
* If (@v != 0), atomically updates @v to (@v + 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_inc_not_zero() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_not_zero(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_not_zero)
return arch_atomic_inc_not_zero(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic_add_unless(v, 1, 0);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_inc_unless_negative() - atomic increment unless negative with full ordering
* @v: pointer to atomic_t
*
* If (@v >= 0), atomically updates @v to (@v + 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_inc_unless_negative() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_inc_unless_negative(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_inc_unless_negative)
return arch_atomic_inc_unless_negative(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int c = raw_atomic_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c < 0))
return false;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic_try_cmpxchg(v, &c, c + 1));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return true;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_unless_positive() - atomic decrement unless positive with full ordering
* @v: pointer to atomic_t
*
* If (@v <= 0), atomically updates @v to (@v - 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_dec_unless_positive() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_unless_positive(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_unless_positive)
return arch_atomic_dec_unless_positive(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int c = raw_atomic_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c > 0))
return false;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic_try_cmpxchg(v, &c, c - 1));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return true;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic_dec_if_positive() - atomic decrement if positive with full ordering
* @v: pointer to atomic_t
*
* If (@v > 0), atomically updates @v to (@v - 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic_dec_if_positive() elsewhere.
*
locking/atomic: scripts: fix ${atomic}_dec_if_positive() kerneldoc The ${atomic}_dec_if_positive() ops are unlike all the other conditional atomic ops. Rather than returning a boolean success value, these return the value that the atomic variable would be updated to, even when no update is performed. We missed this when adding kerneldoc comments, and the documentation for ${atomic}_dec_if_positive() erroneously states: | Return: @true if @v was updated, @false otherwise. Ideally we'd clean this up by aligning ${atomic}_dec_if_positive() with the usual atomic op conventions: with ${atomic}_fetch_dec_if_positive() for those who care about the value of the varaible, and ${atomic}_dec_if_positive() returning a boolean success value. In the mean time, align the documentation with the current reality. Fixes: ad8110706f381170 ("locking/atomic: scripts: generate kerneldoc comments") Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Link: https://lore.kernel.org/r/20230615132734.1119765-1-mark.rutland@arm.com
2023-06-15 13:27:34 +00:00
* Return: The old value of (@v - 1), regardless of whether @v was updated.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline int
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic_dec_if_positive(atomic_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic_dec_if_positive)
return arch_atomic_dec_if_positive(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
int dec, c = raw_atomic_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
dec = c - 1;
if (unlikely(dec < 0))
break;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic_try_cmpxchg(v, &c, dec));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return dec;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#ifdef CONFIG_GENERIC_ATOMIC64
#include <asm-generic/atomic64.h>
#endif
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_read() - atomic load with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically loads the value of @v with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_read() elsewhere.
*
* Return: The value loaded from @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_read(const atomic64_t *v)
{
return arch_atomic64_read(v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_read_acquire() - atomic load with acquire ordering
* @v: pointer to atomic64_t
*
* Atomically loads the value of @v with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_read_acquire() elsewhere.
*
* Return: The value loaded from @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_read_acquire(const atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_read_acquire)
return arch_atomic64_read_acquire(v);
#else
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
s64 ret;
if (__native_word(atomic64_t)) {
ret = smp_load_acquire(&(v)->counter);
} else {
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = raw_atomic64_read(v);
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
__atomic_acquire_fence();
}
return ret;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_set() - atomic set with relaxed ordering
* @v: pointer to atomic64_t
* @i: s64 value to assign
*
* Atomically sets @v to @i with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_set() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_set(atomic64_t *v, s64 i)
{
arch_atomic64_set(v, i);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_set_release() - atomic set with release ordering
* @v: pointer to atomic64_t
* @i: s64 value to assign
*
* Atomically sets @v to @i with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_set_release() elsewhere.
*
* Return: Nothing.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline void
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_set_release(atomic64_t *v, s64 i)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_set_release)
arch_atomic64_set_release(v, i);
#else
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
if (__native_word(atomic64_t)) {
smp_store_release(&(v)->counter, i);
} else {
__atomic_release_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_set(v, i);
atomics: Fix atomic64_{read_acquire,set_release} fallbacks Arnd reports that on 32-bit architectures, the fallbacks for atomic64_read_acquire() and atomic64_set_release() are broken as they use smp_load_acquire() and smp_store_release() respectively, which do not work on types larger than the native word size. Since those contain compiletime_assert_atomic_type(), any attempt to use those fallbacks will result in a build-time error. e.g. with the following added to arch/arm/kernel/setup.c: | void test_atomic64(atomic64_t *v) | { | atomic64_set_release(v, 5); | atomic64_read_acquire(v); | } The compiler will complain as follows: | In file included from <command-line>: | In function 'arch_atomic64_set_release', | inlined from 'test_atomic64' at ./include/linux/atomic/atomic-instrumented.h:669:2: | ././include/linux/compiler_types.h:346:38: error: call to '__compiletime_assert_9' declared with attribute error: Need native word sized stores/loads for atomicity. | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^ | ././include/linux/compiler_types.h:327:4: note: in definition of macro '__compiletime_assert' | 327 | prefix ## suffix(); \ | | ^~~~~~ | ././include/linux/compiler_types.h:346:2: note: in expansion of macro '_compiletime_assert' | 346 | _compiletime_assert(condition, msg, __compiletime_assert_, __COUNTER__) | | ^~~~~~~~~~~~~~~~~~~ | ././include/linux/compiler_types.h:349:2: note: in expansion of macro 'compiletime_assert' | 349 | compiletime_assert(__native_word(t), \ | | ^~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:133:2: note: in expansion of macro 'compiletime_assert_atomic_type' | 133 | compiletime_assert_atomic_type(*p); \ | | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ | ./include/asm-generic/barrier.h:164:55: note: in expansion of macro '__smp_store_release' | 164 | #define smp_store_release(p, v) do { kcsan_release(); __smp_store_release(p, v); } while (0) | | ^~~~~~~~~~~~~~~~~~~ | ./include/linux/atomic/atomic-arch-fallback.h:1270:2: note: in expansion of macro 'smp_store_release' | 1270 | smp_store_release(&(v)->counter, i); | | ^~~~~~~~~~~~~~~~~ | make[2]: *** [scripts/Makefile.build:288: arch/arm/kernel/setup.o] Error 1 | make[1]: *** [scripts/Makefile.build:550: arch/arm/kernel] Error 2 | make: *** [Makefile:1831: arch/arm] Error 2 Fix this by only using smp_load_acquire() and smp_store_release() for native atomic types, and otherwise falling back to the regular barriers necessary for acquire/release semantics, as we do in the more generic acquire and release fallbacks. Since the fallback templates are used to generate the atomic64_*() and atomic_*() operations, the __native_word() check is added to both. For the atomic_*() operations, which are always 32-bit, the __native_word() check is redundant but not harmful, as it is always true. For the example above this works as expected on 32-bit, e.g. for arm multi_v7_defconfig: | <test_atomic64>: | push {r4, r5} | dmb ish | pldw [r0] | mov r2, #5 | mov r3, #0 | ldrexd r4, [r0] | strexd r4, r2, [r0] | teq r4, #0 | bne 484 <test_atomic64+0x14> | ldrexd r2, [r0] | dmb ish | pop {r4, r5} | bx lr ... and also on 64-bit, e.g. for arm64 defconfig: | <test_atomic64>: | bti c | paciasp | mov x1, #0x5 | stlr x1, [x0] | ldar x0, [x0] | autiasp | ret Reported-by: Arnd Bergmann <arnd@arndb.de> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Ard Biesheuvel <ardb@kernel.org> Reviewed-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20220207101943.439825-1-mark.rutland@arm.com
2022-02-07 10:19:43 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add() - atomic add with relaxed ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_add() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_add(s64 i, atomic64_t *v)
{
arch_atomic64_add(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_return() - atomic add with full ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline s64
raw_atomic64_add_return(s64 i, atomic64_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_return)
return arch_atomic64_add_return(i, v);
#elif defined(arch_atomic64_add_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret;
__atomic_pre_full_fence();
ret = arch_atomic64_add_return_relaxed(i, v);
__atomic_post_full_fence();
return ret;
#else
#error "Unable to define raw_atomic64_add_return"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_return_acquire() - atomic add with acquire ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_return_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_return_acquire)
return arch_atomic64_add_return_acquire(i, v);
#elif defined(arch_atomic64_add_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_add_return_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_add_return_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_return_release() - atomic add with release ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_return_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_return_release)
return arch_atomic64_add_return_release(i, v);
#elif defined(arch_atomic64_add_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_add_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_add_return_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_return_relaxed() - atomic add with relaxed ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_add_return_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_add_return_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_add_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_add_return_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_add() - atomic add with full ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_add() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_add(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_add)
return arch_atomic64_fetch_add(i, v);
#elif defined(arch_atomic64_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_add_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_add"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_add_acquire() - atomic add with acquire ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_add_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_add_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_add_acquire)
return arch_atomic64_fetch_add_acquire(i, v);
#elif defined(arch_atomic64_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_fetch_add_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_add_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_add_release() - atomic add with release ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_add_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_add_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_add_release)
return arch_atomic64_fetch_add_release(i, v);
#elif defined(arch_atomic64_fetch_add_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_add_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_add_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_add_relaxed() - atomic add with relaxed ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_add_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_fetch_add_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_fetch_add_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_add_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_add)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_add(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_add_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub() - atomic subtract with relaxed ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_sub(s64 i, atomic64_t *v)
{
arch_atomic64_sub(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub_return() - atomic subtract with full ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_sub_return(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_sub_return)
return arch_atomic64_sub_return(i, v);
#elif defined(arch_atomic64_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_sub_return_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_sub_return"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub_return_acquire() - atomic subtract with acquire ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_sub_return_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_sub_return_acquire)
return arch_atomic64_sub_return_acquire(i, v);
#elif defined(arch_atomic64_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_sub_return_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_sub_return_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub_return_release() - atomic subtract with release ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_sub_return_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_sub_return_release)
return arch_atomic64_sub_return_release(i, v);
#elif defined(arch_atomic64_sub_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_sub_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_sub_return_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub_return_relaxed() - atomic subtract with relaxed ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_sub_return_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_sub_return_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_sub_return_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_sub_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_sub_return(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_sub_return_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_sub() - atomic subtract with full ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_sub() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_sub(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_sub)
return arch_atomic64_fetch_sub(i, v);
#elif defined(arch_atomic64_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_sub_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_sub"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_sub_acquire() - atomic subtract with acquire ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_sub_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_sub_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_sub_acquire)
return arch_atomic64_fetch_sub_acquire(i, v);
#elif defined(arch_atomic64_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_fetch_sub_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_sub_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_sub_release() - atomic subtract with release ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_sub_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_sub_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_sub_release)
return arch_atomic64_fetch_sub_release(i, v);
#elif defined(arch_atomic64_fetch_sub_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_sub_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_sub_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_sub_relaxed() - atomic subtract with relaxed ordering
* @i: s64 value to subtract
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_sub_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_fetch_sub_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_fetch_sub_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_sub_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_sub)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_sub(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_sub_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc() - atomic increment with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc() elsewhere.
*
* Return: Nothing.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline void
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc)
arch_atomic64_inc(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_return() - atomic increment with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_return(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_return)
return arch_atomic64_inc_return(v);
#elif defined(arch_atomic64_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret;
__atomic_pre_full_fence();
ret = arch_atomic64_inc_return_relaxed(v);
__atomic_post_full_fence();
return ret;
#else
return raw_atomic64_add_return(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_return_acquire() - atomic increment with acquire ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_return_acquire(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_return_acquire)
return arch_atomic64_inc_return_acquire(v);
#elif defined(arch_atomic64_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_inc_return_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_inc_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_inc_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_add_return_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_return_release() - atomic increment with release ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_return_release(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_return_release)
return arch_atomic64_inc_return_release(v);
#elif defined(arch_atomic64_inc_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_inc_return_relaxed(v);
#elif defined(arch_atomic64_inc_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_inc_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_add_return_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_return_relaxed() - atomic increment with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_return_relaxed(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_return_relaxed)
return arch_atomic64_inc_return_relaxed(v);
#elif defined(arch_atomic64_inc_return)
return arch_atomic64_inc_return(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_add_return_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_inc() - atomic increment with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_inc() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_inc(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_inc)
return arch_atomic64_fetch_inc(v);
#elif defined(arch_atomic64_fetch_inc_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_inc_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_add(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_inc_acquire() - atomic increment with acquire ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_inc_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline s64
raw_atomic64_fetch_inc_acquire(atomic64_t *v)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_inc_acquire)
return arch_atomic64_fetch_inc_acquire(v);
#elif defined(arch_atomic64_fetch_inc_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_fetch_inc_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_fetch_inc)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_inc(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_add_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_inc_release() - atomic increment with release ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_inc_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_inc_release(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_inc_release)
return arch_atomic64_fetch_inc_release(v);
#elif defined(arch_atomic64_fetch_inc_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_inc_relaxed(v);
#elif defined(arch_atomic64_fetch_inc)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_inc(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_add_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_inc_relaxed() - atomic increment with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_inc_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_inc_relaxed(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_inc_relaxed)
return arch_atomic64_fetch_inc_relaxed(v);
#elif defined(arch_atomic64_fetch_inc)
return arch_atomic64_fetch_inc(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_fetch_add_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec() - atomic decrement with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline void
raw_atomic64_dec(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec)
arch_atomic64_dec(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_sub(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_return() - atomic decrement with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec_return() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_return(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_return)
return arch_atomic64_dec_return(v);
#elif defined(arch_atomic64_dec_return_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_dec_return_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_sub_return(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_return_acquire() - atomic decrement with acquire ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec_return_acquire() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_return_acquire(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_return_acquire)
return arch_atomic64_dec_return_acquire(v);
#elif defined(arch_atomic64_dec_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_dec_return_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_dec_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_dec_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_sub_return_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_return_release() - atomic decrement with release ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec_return_release() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_return_release(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_return_release)
return arch_atomic64_dec_return_release(v);
#elif defined(arch_atomic64_dec_return_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_dec_return_relaxed(v);
#elif defined(arch_atomic64_dec_return)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_dec_return(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_sub_return_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_return_relaxed() - atomic decrement with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec_return_relaxed() elsewhere.
*
* Return: The updated value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_return_relaxed(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_return_relaxed)
return arch_atomic64_dec_return_relaxed(v);
#elif defined(arch_atomic64_dec_return)
return arch_atomic64_dec_return(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_sub_return_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_dec() - atomic decrement with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_dec() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_dec(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_dec)
return arch_atomic64_fetch_dec(v);
#elif defined(arch_atomic64_fetch_dec_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_dec_relaxed(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_sub(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_dec_acquire() - atomic decrement with acquire ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_dec_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_dec_acquire(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_dec_acquire)
return arch_atomic64_fetch_dec_acquire(v);
#elif defined(arch_atomic64_fetch_dec_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_fetch_dec_relaxed(v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_fetch_dec)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_dec(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_sub_acquire(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_dec_release() - atomic decrement with release ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_dec_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_dec_release(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_dec_release)
return arch_atomic64_fetch_dec_release(v);
#elif defined(arch_atomic64_fetch_dec_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_dec_relaxed(v);
#elif defined(arch_atomic64_fetch_dec)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_dec(v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_sub_release(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_dec_relaxed() - atomic decrement with relaxed ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_dec_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_dec_relaxed(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_dec_relaxed)
return arch_atomic64_fetch_dec_relaxed(v);
#elif defined(arch_atomic64_fetch_dec)
return arch_atomic64_fetch_dec(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_fetch_sub_relaxed(1, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_and() - atomic bitwise AND with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_and() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_and(s64 i, atomic64_t *v)
{
arch_atomic64_and(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_and() - atomic bitwise AND with full ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_and() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_and(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_and)
return arch_atomic64_fetch_and(i, v);
#elif defined(arch_atomic64_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_and_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_and"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_and_acquire() - atomic bitwise AND with acquire ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_and_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_and_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_and_acquire)
return arch_atomic64_fetch_and_acquire(i, v);
#elif defined(arch_atomic64_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_fetch_and_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_and_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_and_release() - atomic bitwise AND with release ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_and_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_and_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_and_release)
return arch_atomic64_fetch_and_release(i, v);
#elif defined(arch_atomic64_fetch_and_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_and_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_and_release"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_and_relaxed() - atomic bitwise AND with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_and_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_fetch_and_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_fetch_and_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_and_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_and)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_and(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_and_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_andnot() - atomic bitwise AND NOT with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & ~@i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_andnot() elsewhere.
*
* Return: Nothing.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline void
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_andnot(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_andnot)
arch_atomic64_andnot(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_and(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_andnot() - atomic bitwise AND NOT with full ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & ~@i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_andnot() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_andnot(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_andnot)
return arch_atomic64_fetch_andnot(i, v);
#elif defined(arch_atomic64_fetch_andnot_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret;
__atomic_pre_full_fence();
ret = arch_atomic64_fetch_andnot_relaxed(i, v);
__atomic_post_full_fence();
return ret;
#else
return raw_atomic64_fetch_and(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_andnot_acquire() - atomic bitwise AND NOT with acquire ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & ~@i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_andnot_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_andnot_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_andnot_acquire)
return arch_atomic64_fetch_andnot_acquire(i, v);
#elif defined(arch_atomic64_fetch_andnot_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_fetch_andnot_relaxed(i, v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_fetch_andnot)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_andnot(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_and_acquire(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_andnot_release() - atomic bitwise AND NOT with release ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & ~@i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_andnot_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_andnot_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_andnot_release)
return arch_atomic64_fetch_andnot_release(i, v);
#elif defined(arch_atomic64_fetch_andnot_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_andnot_relaxed(i, v);
#elif defined(arch_atomic64_fetch_andnot)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_andnot(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_fetch_and_release(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_andnot_relaxed() - atomic bitwise AND NOT with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v & ~@i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_andnot_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_andnot_relaxed(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_andnot_relaxed)
return arch_atomic64_fetch_andnot_relaxed(i, v);
#elif defined(arch_atomic64_fetch_andnot)
return arch_atomic64_fetch_andnot(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_fetch_and_relaxed(~i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_or() - atomic bitwise OR with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v | @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_or() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_or(s64 i, atomic64_t *v)
{
arch_atomic64_or(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_or() - atomic bitwise OR with full ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v | @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_or() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_or(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_or)
return arch_atomic64_fetch_or(i, v);
#elif defined(arch_atomic64_fetch_or_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_or_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_or"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_or_acquire() - atomic bitwise OR with acquire ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v | @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_or_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_or_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_or_acquire)
return arch_atomic64_fetch_or_acquire(i, v);
#elif defined(arch_atomic64_fetch_or_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_fetch_or_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_or_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_or_release() - atomic bitwise OR with release ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v | @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_or_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_or_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_or_release)
return arch_atomic64_fetch_or_release(i, v);
#elif defined(arch_atomic64_fetch_or_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_or_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_or_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_or_relaxed() - atomic bitwise OR with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v | @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_or_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_fetch_or_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_fetch_or_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_or_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_or)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_or(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_or_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_xor() - atomic bitwise XOR with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v ^ @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_xor() elsewhere.
*
* Return: Nothing.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline void
raw_atomic64_xor(s64 i, atomic64_t *v)
{
arch_atomic64_xor(i, v);
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_xor() - atomic bitwise XOR with full ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v ^ @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_xor() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_xor(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_xor)
return arch_atomic64_fetch_xor(i, v);
#elif defined(arch_atomic64_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_fetch_xor_relaxed(i, v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_xor"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_xor_acquire() - atomic bitwise XOR with acquire ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v ^ @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_xor_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_xor_acquire(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_xor_acquire)
return arch_atomic64_fetch_xor_acquire(i, v);
#elif defined(arch_atomic64_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret = arch_atomic64_fetch_xor_relaxed(i, v);
__atomic_acquire_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_xor_acquire"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_xor_release() - atomic bitwise XOR with release ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v ^ @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_xor_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_xor_release(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_xor_release)
return arch_atomic64_fetch_xor_release(i, v);
#elif defined(arch_atomic64_fetch_xor_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_release_fence();
return arch_atomic64_fetch_xor_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_xor_release"
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_xor_relaxed() - atomic bitwise XOR with relaxed ordering
* @i: s64 value
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v ^ @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_fetch_xor_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
static __always_inline s64
raw_atomic64_fetch_xor_relaxed(s64 i, atomic64_t *v)
{
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#if defined(arch_atomic64_fetch_xor_relaxed)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_xor_relaxed(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_fetch_xor)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_fetch_xor(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
#error "Unable to define raw_atomic64_fetch_xor_relaxed"
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_xchg() - atomic exchange with full ordering
* @v: pointer to atomic64_t
* @new: s64 value to assign
*
* Atomically updates @v to @new with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_xchg() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic64_xchg(atomic64_t *v, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_xchg)
return arch_atomic64_xchg(v, new);
#elif defined(arch_atomic64_xchg_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
ret = arch_atomic64_xchg_relaxed(v, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg(&v->counter, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_xchg_acquire() - atomic exchange with acquire ordering
* @v: pointer to atomic64_t
* @new: s64 value to assign
*
* Atomically updates @v to @new with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_xchg_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline s64
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic64_xchg_acquire(atomic64_t *v, s64 new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_xchg_acquire)
return arch_atomic64_xchg_acquire(v, new);
#elif defined(arch_atomic64_xchg_relaxed)
s64 ret = arch_atomic64_xchg_relaxed(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_xchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_xchg(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg_acquire(&v->counter, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_xchg_release() - atomic exchange with release ordering
* @v: pointer to atomic64_t
* @new: s64 value to assign
*
* Atomically updates @v to @new with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_xchg_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline s64
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
raw_atomic64_xchg_release(atomic64_t *v, s64 new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_xchg_release)
return arch_atomic64_xchg_release(v, new);
#elif defined(arch_atomic64_xchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_xchg_relaxed(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#elif defined(arch_atomic64_xchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_xchg(v, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_xchg_release(&v->counter, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_xchg_relaxed() - atomic exchange with relaxed ordering
* @v: pointer to atomic64_t
* @new: s64 value to assign
*
* Atomically updates @v to @new with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_xchg_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_xchg_relaxed(atomic64_t *v, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_xchg_relaxed)
return arch_atomic64_xchg_relaxed(v, new);
#elif defined(arch_atomic64_xchg)
return arch_atomic64_xchg(v, new);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_xchg_relaxed(&v->counter, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_cmpxchg() - atomic compare and exchange with full ordering
* @v: pointer to atomic64_t
* @old: s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_cmpxchg() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_cmpxchg(atomic64_t *v, s64 old, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_cmpxchg)
return arch_atomic64_cmpxchg(v, old, new);
#elif defined(arch_atomic64_cmpxchg_relaxed)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 ret;
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_cmpxchg_relaxed(v, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg(&v->counter, old, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_cmpxchg_acquire() - atomic compare and exchange with acquire ordering
* @v: pointer to atomic64_t
* @old: s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with acquire ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_cmpxchg_acquire() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_cmpxchg_acquire(atomic64_t *v, s64 old, s64 new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_cmpxchg_acquire)
return arch_atomic64_cmpxchg_acquire(v, old, new);
#elif defined(arch_atomic64_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 ret = arch_atomic64_cmpxchg_relaxed(v, old, new);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg_acquire(&v->counter, old, new);
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_cmpxchg_release() - atomic compare and exchange with release ordering
* @v: pointer to atomic64_t
* @old: s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with release ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_cmpxchg_release() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_cmpxchg_release(atomic64_t *v, s64 old, s64 new)
locking/atomic: make atomic*_{cmp,}xchg optional Most architectures define the atomic/atomic64 xchg and cmpxchg operations in terms of arch_xchg and arch_cmpxchg respectfully. Add fallbacks for these cases and remove the trivial cases from arch code. On some architectures the existing definitions are kept as these are used to build other arch_atomic*() operations. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-5-mark.rutland@arm.com
2023-06-05 07:01:01 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_cmpxchg_release)
return arch_atomic64_cmpxchg_release(v, old, new);
#elif defined(arch_atomic64_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic64_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_cmpxchg_release(&v->counter, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_cmpxchg_relaxed() - atomic compare and exchange with relaxed ordering
* @v: pointer to atomic64_t
* @old: s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with relaxed ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_cmpxchg_relaxed() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_cmpxchg_relaxed(atomic64_t *v, s64 old, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_cmpxchg_relaxed)
return arch_atomic64_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic64_cmpxchg)
return arch_atomic64_cmpxchg(v, old, new);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_cmpxchg_relaxed(&v->counter, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_try_cmpxchg() - atomic compare and exchange with full ordering
* @v: pointer to atomic64_t
* @old: pointer to s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_try_cmpxchg() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline bool
raw_atomic64_try_cmpxchg(atomic64_t *v, s64 *old, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_try_cmpxchg)
return arch_atomic64_try_cmpxchg(v, old, new);
#elif defined(arch_atomic64_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_pre_full_fence();
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
ret = arch_atomic64_try_cmpxchg_relaxed(v, old, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
__atomic_post_full_fence();
return ret;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic64_cmpxchg(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_try_cmpxchg_acquire() - atomic compare and exchange with acquire ordering
* @v: pointer to atomic64_t
* @old: pointer to s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with acquire ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_try_cmpxchg_acquire() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline bool
raw_atomic64_try_cmpxchg_acquire(atomic64_t *v, s64 *old, s64 new)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_try_cmpxchg_acquire)
return arch_atomic64_try_cmpxchg_acquire(v, old, new);
#elif defined(arch_atomic64_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret = arch_atomic64_try_cmpxchg_relaxed(v, old, new);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_try_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_try_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic64_cmpxchg_acquire(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_try_cmpxchg_release() - atomic compare and exchange with release ordering
* @v: pointer to atomic64_t
* @old: pointer to s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with release ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_try_cmpxchg_release() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
static __always_inline bool
raw_atomic64_try_cmpxchg_release(atomic64_t *v, s64 *old, s64 new)
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_try_cmpxchg_release)
return arch_atomic64_try_cmpxchg_release(v, old, new);
#elif defined(arch_atomic64_try_cmpxchg_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_try_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic64_try_cmpxchg)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_try_cmpxchg(v, old, new);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic64_cmpxchg_release(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_try_cmpxchg_relaxed() - atomic compare and exchange with relaxed ordering
* @v: pointer to atomic64_t
* @old: pointer to s64 value to compare with
* @new: s64 value to assign
*
* If (@v == @old), atomically updates @v to @new with relaxed ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified, @old is updated to the current value of @v,
* and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_try_cmpxchg_relaxed() elsewhere.
*
* Return: @true if the exchange occured, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_try_cmpxchg_relaxed(atomic64_t *v, s64 *old, s64 new)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_try_cmpxchg_relaxed)
return arch_atomic64_try_cmpxchg_relaxed(v, old, new);
#elif defined(arch_atomic64_try_cmpxchg)
return arch_atomic64_try_cmpxchg(v, old, new);
#else
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
s64 r, o = *old;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
r = raw_atomic64_cmpxchg_relaxed(v, o, new);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
if (unlikely(r != o))
*old = r;
return likely(r == o);
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_sub_and_test() - atomic subtract and test if zero with full ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_sub_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_sub_and_test(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_sub_and_test)
return arch_atomic64_sub_and_test(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_sub_return(i, v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_and_test() - atomic decrement and test if zero with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v - 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_dec_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_and_test(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_and_test)
return arch_atomic64_dec_and_test(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_dec_return(v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_and_test() - atomic increment and test if zero with full ordering
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + 1) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_inc_and_test() elsewhere.
*
* Return: @true if the resulting value of @v is zero, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_and_test(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_and_test)
return arch_atomic64_inc_and_test(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_inc_return(v) == 0;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_negative() - atomic add and test if negative with full ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with full ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_negative() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_negative(s64 i, atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_negative)
return arch_atomic64_add_negative(i, v);
#elif defined(arch_atomic64_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret;
__atomic_pre_full_fence();
ret = arch_atomic64_add_negative_relaxed(i, v);
__atomic_post_full_fence();
return ret;
#else
return raw_atomic64_add_return(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_negative_acquire() - atomic add and test if negative with acquire ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with acquire ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_negative_acquire() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_negative_acquire(s64 i, atomic64_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_negative_acquire)
return arch_atomic64_add_negative_acquire(i, v);
#elif defined(arch_atomic64_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
bool ret = arch_atomic64_add_negative_relaxed(i, v);
__atomic_acquire_fence();
return ret;
#elif defined(arch_atomic64_add_negative)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_negative(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_add_return_acquire(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_negative_release() - atomic add and test if negative with release ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with release ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_negative_release() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_negative_release(s64 i, atomic64_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_negative_release)
return arch_atomic64_add_negative_release(i, v);
#elif defined(arch_atomic64_add_negative_relaxed)
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
__atomic_release_fence();
return arch_atomic64_add_negative_relaxed(i, v);
#elif defined(arch_atomic64_add_negative)
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
return arch_atomic64_add_negative(i, v);
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
#else
return raw_atomic64_add_return_release(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_negative_relaxed() - atomic add and test if negative with relaxed ordering
* @i: s64 value to add
* @v: pointer to atomic64_t
*
* Atomically updates @v to (@v + @i) with relaxed ordering.
*
* Safe to use in noinstr code; prefer atomic64_add_negative_relaxed() elsewhere.
*
* Return: @true if the resulting value of @v is negative, @false otherwise.
*/
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_negative_relaxed(s64 i, atomic64_t *v)
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_negative_relaxed)
return arch_atomic64_add_negative_relaxed(i, v);
#elif defined(arch_atomic64_add_negative)
return arch_atomic64_add_negative(i, v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_add_return_relaxed(i, v) < 0;
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
atomics: Provide atomic_add_negative() variants atomic_add_negative() does not provide the relaxed/acquire/release variants. Provide them in preparation for a new scalable reference count algorithm. Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lore.kernel.org/r/20230323102800.101763813@linutronix.de
2023-03-23 20:55:30 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_fetch_add_unless() - atomic add unless value with full ordering
* @v: pointer to atomic64_t
* @a: s64 value to add
* @u: s64 value to compare with
*
* If (@v != @u), atomically updates @v to (@v + @a) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_fetch_add_unless() elsewhere.
*
* Return: The original value of @v.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_fetch_add_unless(atomic64_t *v, s64 a, s64 u)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_fetch_add_unless)
return arch_atomic64_fetch_add_unless(v, a, u);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 c = raw_atomic64_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c == u))
break;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic64_try_cmpxchg(v, &c, c + a));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return c;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_add_unless() - atomic add unless value with full ordering
* @v: pointer to atomic64_t
* @a: s64 value to add
* @u: s64 value to compare with
*
* If (@v != @u), atomically updates @v to (@v + @a) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_add_unless() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_add_unless(atomic64_t *v, s64 a, s64 u)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_add_unless)
return arch_atomic64_add_unless(v, a, u);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_fetch_add_unless(v, a, u) != u;
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_not_zero() - atomic increment unless zero with full ordering
* @v: pointer to atomic64_t
*
* If (@v != 0), atomically updates @v to (@v + 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_inc_not_zero() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_not_zero(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_not_zero)
return arch_atomic64_inc_not_zero(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
return raw_atomic64_add_unless(v, 1, 0);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_inc_unless_negative() - atomic increment unless negative with full ordering
* @v: pointer to atomic64_t
*
* If (@v >= 0), atomically updates @v to (@v + 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_inc_unless_negative() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_inc_unless_negative(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_inc_unless_negative)
return arch_atomic64_inc_unless_negative(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 c = raw_atomic64_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c < 0))
return false;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic64_try_cmpxchg(v, &c, c + 1));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return true;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_unless_positive() - atomic decrement unless positive with full ordering
* @v: pointer to atomic64_t
*
* If (@v <= 0), atomically updates @v to (@v - 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_dec_unless_positive() elsewhere.
*
* Return: @true if @v was updated, @false otherwise.
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline bool
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_unless_positive(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_unless_positive)
return arch_atomic64_dec_unless_positive(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 c = raw_atomic64_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
if (unlikely(c > 0))
return false;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic64_try_cmpxchg(v, &c, c - 1));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return true;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
/**
* raw_atomic64_dec_if_positive() - atomic decrement if positive with full ordering
* @v: pointer to atomic64_t
*
* If (@v > 0), atomically updates @v to (@v - 1) with full ordering.
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
* Otherwise, @v is not modified and relaxed ordering is provided.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*
* Safe to use in noinstr code; prefer atomic64_dec_if_positive() elsewhere.
*
locking/atomic: scripts: fix ${atomic}_dec_if_positive() kerneldoc The ${atomic}_dec_if_positive() ops are unlike all the other conditional atomic ops. Rather than returning a boolean success value, these return the value that the atomic variable would be updated to, even when no update is performed. We missed this when adding kerneldoc comments, and the documentation for ${atomic}_dec_if_positive() erroneously states: | Return: @true if @v was updated, @false otherwise. Ideally we'd clean this up by aligning ${atomic}_dec_if_positive() with the usual atomic op conventions: with ${atomic}_fetch_dec_if_positive() for those who care about the value of the varaible, and ${atomic}_dec_if_positive() returning a boolean success value. In the mean time, align the documentation with the current reality. Fixes: ad8110706f381170 ("locking/atomic: scripts: generate kerneldoc comments") Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Link: https://lore.kernel.org/r/20230615132734.1119765-1-mark.rutland@arm.com
2023-06-15 13:27:34 +00:00
* Return: The old value of (@v - 1), regardless of whether @v was updated.
locking/atomic: scripts: generate kerneldoc comments Currently the atomics are documented in Documentation/atomic_t.txt, and have no kerneldoc comments. There are a sufficient number of gotchas (e.g. semantics, noinstr-safety) that it would be nice to have comments to call these out, and it would be nice to have kerneldoc comments such that these can be collated. While it's possible to derive the semantics from the code, this can be painful given the amount of indirection we currently have (e.g. fallback paths), and it's easy to be mislead by naming, e.g. * The unconditional void-returning ops *only* have relaxed variants without a _relaxed suffix, and can easily be mistaken for being fully ordered. It would be nice to give these a _relaxed() suffix, but this would result in significant churn throughout the kernel. * Our naming of conditional and unconditional+test ops is rather inconsistent, and it can be difficult to derive the name of an operation, or to identify where an op is conditional or unconditional+test. Some ops are clearly conditional: - dec_if_positive - add_unless - dec_unless_positive - inc_unless_negative Some ops are clearly unconditional+test: - sub_and_test - dec_and_test - inc_and_test However, what exactly those test is not obvious. A _test_zero suffix might be clearer. Others could be read ambiguously: - inc_not_zero // conditional - add_negative // unconditional+test It would probably be worth renaming these, e.g. to inc_unless_zero and add_test_negative. As a step towards making this more consistent and easier to understand, this patch adds kerneldoc comments for all generated *atomic*_*() functions. These are generated from templates, with some common text shared, making it easy to extend these in future if necessary. I've tried to make these as consistent and clear as possible, and I've deliberately ensured: * All ops have their ordering explicitly mentioned in the short and long description. * All test ops have "test" in their short description. * All ops are described as an expression using their usual C operator. For example: andnot: "Atomically updates @v to (@v & ~@i)" inc: "Atomically updates @v to (@v + 1)" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All conditional ops have their condition described as an expression using the usual C operators. For example: add_unless: "If (@v != @u), atomically updates @v to (@v + @i)" cmpxchg: "If (@v == @old), atomically updates @v to @new" Which may be clearer to non-naative English speakers, and allows all the operations to be described in the same style. * All bitwise ops (and,andnot,or,xor) explicitly mention that they are bitwise in their short description, so that they are not mistaken for performing their logical equivalents. * The noinstr safety of each op is explicitly described, with a description of whether or not to use the raw_ form of the op. There should be no functional change as a result of this patch. Reported-by: Paul E. McKenney <paulmck@kernel.org> Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-26-mark.rutland@arm.com
2023-06-05 07:01:22 +00:00
*/
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
static __always_inline s64
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
raw_atomic64_dec_if_positive(atomic64_t *v)
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
{
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
#if defined(arch_atomic64_dec_if_positive)
return arch_atomic64_dec_if_positive(v);
#else
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
s64 dec, c = raw_atomic64_read(v);
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
do {
dec = c - 1;
if (unlikely(dec < 0))
break;
locking/atomic: scripts: restructure fallback ifdeffery Currently the various ordering variants of an atomic operation are defined in groups of full/acquire/release/relaxed ordering variants with some shared ifdeffery and several potential definitions of each ordering variant in different branches of the shared ifdeffery. As an ordering variant can have several potential definitions down different branches of the shared ifdeffery, it can be painful for a human to find a relevant definition, and we don't have a good location to place anything common to all definitions of an ordering variant (e.g. kerneldoc). Historically the grouping of full/acquire/release/relaxed ordering variants was necessary as we filled in the missing atomics in the same namespace as the architecture used. It would be easy to accidentally define one ordering fallback in terms of another ordering fallback with redundant barriers, and avoiding that would otherwise require a lot of baroque ifdeffery. With recent changes we no longer need to fill in the missing atomics in the arch_atomic*_<op>() namespace, and only need to fill in the raw_atomic*_<op>() namespace. Due to this, there's no risk of a namespace collision, and we can define each raw_atomic*_<op> ordering variant with its own ifdeffery checking for the arch_atomic*_<op> ordering variants. Restructure the fallbacks in this way, with each ordering variant having its own ifdeffery of the form: | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Note that where there's no relevant arch_atomic*_<op>() ordering variant, we'll define the operation in terms of a distinct raw_atomic*_<otherop>(), as this itself might have been filled in with a fallback. As we now generate the raw_atomic*_<op>() implementations directly, we no longer need the trivial wrappers, so they are removed. This makes the ifdeffery easier to follow, and will allow for further improvements in subsequent patches. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-21-mark.rutland@arm.com
2023-06-05 07:01:17 +00:00
} while (!raw_atomic64_try_cmpxchg(v, &c, dec));
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
return dec;
#endif
locking/atomic: scripts: simplify raw_atomic*() definitions Currently each ordering variant has several potential definitions, with a mixture of preprocessor and C definitions, including several copies of its C prototype, e.g. | #if defined(arch_atomic_fetch_andnot_acquire) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot_acquire | #elif defined(arch_atomic_fetch_andnot_relaxed) | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | } | #elif defined(arch_atomic_fetch_andnot) | #define raw_atomic_fetch_andnot_acquire arch_atomic_fetch_andnot | #else | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | return raw_atomic_fetch_and_acquire(~i, v); | } | #endif Make this a bit simpler by defining the C prototype once, and writing the various potential definitions as plain C code guarded by ifdeffery. For example, the above becomes: | static __always_inline int | raw_atomic_fetch_andnot_acquire(int i, atomic_t *v) | { | #if defined(arch_atomic_fetch_andnot_acquire) | return arch_atomic_fetch_andnot_acquire(i, v); | #elif defined(arch_atomic_fetch_andnot_relaxed) | int ret = arch_atomic_fetch_andnot_relaxed(i, v); | __atomic_acquire_fence(); | return ret; | #elif defined(arch_atomic_fetch_andnot) | return arch_atomic_fetch_andnot(i, v); | #else | return raw_atomic_fetch_and_acquire(~i, v); | #endif | } Which is far easier to read. As we now always have a single copy of the C prototype wrapping all the potential definitions, we now have an obvious single location for kerneldoc comments. At the same time, the fallbacks for raw_atomic*_xhcg() are made to use 'new' rather than 'i' as the name of the new value. This is what the existing fallback template used, and is more consistent with the raw_atomic{_try,}cmpxchg() fallbacks. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Kees Cook <keescook@chromium.org> Link: https://lore.kernel.org/r/20230605070124.3741859-24-mark.rutland@arm.com
2023-06-05 07:01:20 +00:00
}
locking/atomics: Flip fallbacks and instrumentation Currently instrumentation of atomic primitives is done at the architecture level, while composites or fallbacks are provided at the generic level. The result is that there are no uninstrumented variants of the fallbacks. Since there is now need of such variants to isolate text poke from any form of instrumentation invert this ordering. Doing this means moving the instrumentation into the generic code as well as having (for now) two variants of the fallbacks. Notes: - the various *cond_read* primitives are not proper fallbacks and got moved into linux/atomic.c. No arch_ variants are generated because the base primitives smp_cond_load*() are instrumented. - once all architectures are moved over to arch_atomic_ one of the fallback variants can be removed and some 2300 lines reclaimed. - atomic_{read,set}*() are no longer double-instrumented Reported-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Mark Rutland <mark.rutland@arm.com> Link: https://lkml.kernel.org/r/20200505134058.769149955@linutronix.de
2020-01-24 21:13:03 +00:00
#endif /* _LINUX_ATOMIC_FALLBACK_H */
locking/atomic: scripts: Clarify ordering of conditional atomics Conditional atomic operations (e.g. cmpxchg()) only provide ordering when the condition holds; when the condition does not hold, the location is not modified and relaxed ordering is provided. Where ordering is needed for failed conditional atomics, it is necessary to use smp_mb__before_atomic() and/or smp_mb__after_atomic(). This is explained tersely in memory-barriers.txt, and is implied but not explicitly stated in the kerneldoc comments for the conditional operations. The lack of an explicit statement has lead to some off-list queries about the ordering semantics of failing conditional operations, so evidently this is confusing. Update the kerneldoc comments to explicitly describe the lack of ordering for failed conditional atomic operations. For most conditional atomic operations, this is written as: | If (${condition}), atomically updates @v to (${new}) with ${desc_order} ordering. | Otherwise, @v is not modified and relaxed ordering is provided. For the try_cmpxchg() operations, this is written as: | If (${condition}), atomically updates @v to @new with ${desc_order} ordering. | Otherwise, @v is not modified, @old is updated to the current value of @v, | and relaxed ordering is provided. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Ingo Molnar <mingo@kernel.org> Reviewed-by: Paul E. McKenney <paulmck@kernel.org> Reviewed-by: Nhat Pham <nphamcs@gmail.com> Link: https://lore.kernel.org/r/20240209124010.2096198-1-mark.rutland@arm.com
2024-02-09 12:40:10 +00:00
// 14850c0b0db20c62fdc78ccd1d42b98b88d76331