Replace a bunch of 'p->state == TASK_RUNNING' with a new helper:
task_is_running(p).
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Davidlohr Bueso <dave@stgolabs.net>
Acked-by: Geert Uytterhoeven <geert@linux-m68k.org>
Acked-by: Will Deacon <will@kernel.org>
Link: https://lore.kernel.org/r/20210611082838.222401495@infradead.org
When a thread detects that a memory location was modified without its
watchpoint being hit, the report notes that a change was detected, but
does not provide concrete values for the change. Knowing the concrete
values can be very helpful in tracking down any racy writers (e.g. as
specific values may only be written in some portions of code, or under
certain conditions).
When we detect a modification, let's report the concrete old/new values,
along with the access's mask of relevant bits (and which relevant bits
were modified). This can make it easier to identify potential racy
writers. As the snapshots are at most 8 bytes, we can only report values
for acceses up to this size, but this appears to cater for the common
case.
When we detect a race via a watchpoint, we may or may not have concrete
values for the modification. To be helpful, let's attempt to log them
when we do as they can be ignored where irrelevant.
The resulting reports appears as follows, with values zero-padded to the
access width:
| ==================================================================
| BUG: KCSAN: data-race in el0_svc_common+0x34/0x25c arch/arm64/kernel/syscall.c:96
|
| race at unknown origin, with read to 0xffff00007ae6aa00 of 8 bytes by task 223 on cpu 1:
| el0_svc_common+0x34/0x25c arch/arm64/kernel/syscall.c:96
| do_el0_svc+0x48/0xec arch/arm64/kernel/syscall.c:178
| el0_svc arch/arm64/kernel/entry-common.c:226 [inline]
| el0_sync_handler+0x1a4/0x390 arch/arm64/kernel/entry-common.c:236
| el0_sync+0x140/0x180 arch/arm64/kernel/entry.S:674
|
| value changed: 0x0000000000000000 -> 0x0000000000000002
|
| Reported by Kernel Concurrency Sanitizer on:
| CPU: 1 PID: 223 Comm: syz-executor.1 Not tainted 5.8.0-rc3-00094-ga73f923ecc8e-dirty #3
| Hardware name: linux,dummy-virt (DT)
| ==================================================================
If an access mask is set, it is shown underneath the "value changed"
line as "bits changed: 0x<bits changed> with mask 0x<non-zero mask>".
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
[ elver@google.com: align "value changed" and "bits changed" lines,
which required massaging the message; do not print bits+mask if no
mask set. ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Now that the reporting code has been refactored, it's clear by
construction that print_report() can only be passed
KCSAN_REPORT_RACE_SIGNAL or KCSAN_REPORT_RACE_UNKNOWN_ORIGIN, and these
can also be distinguished by the presence of `other_info`.
Let's simplify things and remove the report type enum, and instead let's
check `other_info` to distinguish these cases. This allows us to remove
code for cases which are impossible and generally makes the code simpler.
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
[ elver@google.com: add updated comments to kcsan_report_*() functions ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Now that we have separate kcsan_report_*() functions, we can factor the
distinct logic for each of the report cases out of kcsan_report(). While
this means each case has to handle mutual exclusion independently, this
minimizes the conditionality of code and makes it easier to read, and
will permit passing distinct bits of information to print_report() in
future.
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
[ elver@google.com: retain comment about lockdep_off() ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In subsequent patches we'll want to split kcsan_report() into distinct
handlers for each report type. The largest bit of common work is
initializing the `access_info`, so let's factor this out into a helper,
and have the kcsan_report_*() functions pass the `aaccess_info` as a
parameter to kcsan_report().
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
So that we can add more callers of print_report(), lets fold the panic()
call into print_report() so the caller doesn't have to handle this
explicitly.
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The `watchpoint_idx` argument to kcsan_report() isn't meaningful for
races which were not detected by a watchpoint, and it would be clearer
if callers passed the other_info directly so that a NULL value can be
passed in this case.
Given that callers manipulate their watchpoints before passing the index
into kcsan_report_*(), and given we index the `other_infos` array using
this before we sanity-check it, the subsequent sanity check isn't all
that useful.
Let's remove the `watchpoint_idx` sanity check, and move the job of
finding the `other_info` out of kcsan_report().
Other than the removal of the check, there should be no functional
change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Currently kcsan_report() is used to handle three distinct cases:
* The caller hit a watchpoint when attempting an access. Some
information regarding the caller and access are recorded, but no
output is produced.
* A caller which previously setup a watchpoint detected that the
watchpoint has been hit, and possibly detected a change to the
location in memory being watched. This may result in output reporting
the interaction between this caller and the caller which hit the
watchpoint.
* A caller detected a change to a modification to a memory location
which wasn't detected by a watchpoint, for which there is no
information on the other thread. This may result in output reporting
the unexpected change.
... depending on the specific case the caller has distinct pieces of
information available, but the prototype of kcsan_report() has to handle
all three cases. This means that in some cases we pass redundant
information, and in others we don't pass all the information we could
pass. This also means that the report code has to demux these three
cases.
So that we can pass some additional information while also simplifying
the callers and report code, add separate kcsan_report_*() functions for
the distinct cases, updating callers accordingly. As the watchpoint_idx
is unused in the case of kcsan_report_unknown_origin(), this passes a
dummy value into kcsan_report(). Subsequent patches will refactor the
report code to avoid this.
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
[ elver@google.com: try to make kcsan_report_*() names more descriptive ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In kcsan_setup_watchpoint() we store snapshots of a watched value into a
union of u8/u16/u32/u64 sized fields, modify this in place using a
consistent field, then later check for any changes via the u64 field.
We can achieve the safe effect more simply by always treating the field
as a u64, as smaller values will be zero-extended. As the values are
zero-extended, we don't need to truncate the access_mask when we apply
it, and can always apply the full 64-bit access_mask to the 64-bit
value.
Finally, we can store the two snapshots and calculated difference
separately, which makes the code a little easier to read, and will
permit reporting the old/new values in subsequent patches.
There should be no functional change as a result of this patch.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
clang with CONFIG_LTO_CLANG points out that an initcall function should
return an 'int' due to the changes made to the initcall macros in commit
3578ad11f3 ("init: lto: fix PREL32 relocations"):
kernel/kcsan/debugfs.c:274:15: error: returning 'void' from a function with incompatible result type 'int'
late_initcall(kcsan_debugfs_init);
~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~
include/linux/init.h:292:46: note: expanded from macro 'late_initcall'
#define late_initcall(fn) __define_initcall(fn, 7)
Fixes: e36299efe7 ("kcsan, debugfs: Move debugfs file creation out of early init")
Cc: stable <stable@vger.kernel.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Marco Elver <elver@google.com>
Reviewed-by: Nathan Chancellor <nathan@kernel.org>
Reviewed-by: Miguel Ojeda <ojeda@kernel.org>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Printing a 'long' variable using the '%d' format string is wrong
and causes a warning from gcc:
kernel/kcsan/kcsan_test.c: In function 'nthreads_gen_params':
include/linux/kern_levels.h:5:25: error: format '%d' expects argument of type 'int', but argument 3 has type 'long int' [-Werror=format=]
Use the appropriate format modifier.
Fixes: f6a1491403 ("kcsan: Switch to KUNIT_CASE_PARAM for parameterized tests")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Marco Elver <elver@google.com>
Acked-by: Paul E. McKenney <paulmck@kernel.org>
Link: https://lkml.kernel.org/r/20210421135059.3371701-1-arnd@kernel.org
Adds missing license and/or copyright headers for KCSAN source files.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Since KUnit now support parameterized tests via KUNIT_CASE_PARAM, update
KCSAN's test to switch to it for parameterized tests. This simplifies
parameterized tests and gets rid of the "parameters in case name"
workaround (hack).
At the same time, we can increase the maximum number of threads used,
because on systems with too few CPUs, KUnit allows us to now stop at the
maximum useful threads and not unnecessarily execute redundant test
cases with (the same) limited threads as had been the case before.
Reviewed-by: David Gow <davidgow@google.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Per recently added KUnit style recommendations at
Documentation/dev-tools/kunit/style.rst, make the following changes to
the KCSAN test:
1. Rename 'kcsan-test.c' to 'kcsan_test.c'.
2. Rename suite name 'kcsan-test' to 'kcsan'.
3. Rename CONFIG_KCSAN_TEST to CONFIG_KCSAN_KUNIT_TEST and
default to KUNIT_ALL_TESTS.
Reviewed-by: David Gow <davidgow@google.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Commit 56348560d4 ("debugfs: do not attempt to create a new file
before the filesystem is initalized") forbids creating new debugfs files
until debugfs is fully initialized. This means that KCSAN's debugfs
file creation, which happened at the end of __init(), no longer works.
And was apparently never supposed to work!
However, there is no reason to create KCSAN's debugfs file so early.
This commit therefore moves its creation to a late_initcall() callback.
Cc: "Rafael J. Wysocki" <rafael@kernel.org>
Cc: stable <stable@vger.kernel.org>
Fixes: 56348560d4 ("debugfs: do not attempt to create a new file before the filesystem is initalized")
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Rewrite kcsan_prandom_u32_max() to not depend on code that might be
instrumented, removing any dependency on lib/random32.c. The rewrite
implements a simple linear congruential generator, that is sufficient
for our purposes (for udelay() and skip_watch counter randomness).
The initial motivation for this was to allow enabling KCSAN for
kernel/sched (remove KCSAN_SANITIZE := n from kernel/sched/Makefile),
with CONFIG_DEBUG_PREEMPT=y. Without this change, we could observe
recursion:
check_access() [via instrumentation]
kcsan_setup_watchpoint()
reset_kcsan_skip()
kcsan_prandom_u32_max()
get_cpu_var()
preempt_disable()
preempt_count_add() [in kernel/sched/core.c]
check_access() [via instrumentation]
Note, while this currently does not affect an unmodified kernel, it'd be
good to keep a KCSAN kernel working when KCSAN_SANITIZE := n is removed
from kernel/sched/Makefile to permit testing scheduler code with KCSAN
if desired.
Fixes: cd290ec246 ("kcsan: Use tracing-safe version of prandom")
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The watchpoint encoding masks for size and address were off-by-one bit
each, with the size mask using 1 unnecessary bit and the address mask
missing 1 bit. However, due to the way the size is shifted into the
encoded watchpoint, we were effectively wasting and never using the
extra bit.
For example, on x86 with PAGE_SIZE==4K, we have 1 bit for the is-write
bit, 14 bits for the size bits, and then 49 bits left for the address.
Prior to this fix we would end up with this usage:
[ write<1> | size<14> | wasted<1> | address<48> ]
Fix it by subtracting 1 bit from the GENMASK() end and start ranges of
size and address respectively. The added static_assert()s verify that
the masks are as expected. With the fixed version, we get the expected
usage:
[ write<1> | size<14> | address<49> ]
Functionally no change is expected, since that extra address bit is
insignificant for enabled architectures.
Acked-by: Boqun Feng <boqun.feng@gmail.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Avoid setting up watchpoints on NULL pointers, as otherwise we would
crash inside the KCSAN runtime (when checking for value changes) instead
of the instrumented code.
Because that may be confusing, skip any address less than PAGE_SIZE.
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In preparation of supporting only addresses not within the NULL page,
change the selftest to never use addresses that are less than PAGE_SIZE.
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Fix multiple occurrences of duplicated words in kernel/.
Fix one typo/spello on the same line as a duplicate word. Change one
instance of "the the" to "that the". Otherwise just drop one of the
repeated words.
Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Link: https://lkml.kernel.org/r/98202fa6-8919-ef63-9efe-c0fad5ca7af1@infradead.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In the core runtime, we must minimize any calls to external library
functions to avoid any kind of recursion. This can happen even though
instrumentation is disabled for called functions, but tracing is
enabled.
Most recently, prandom_u32() added a tracepoint, which can cause
problems for KCSAN even if the rcuidle variant is used. For example:
kcsan -> prandom_u32() -> trace_prandom_u32_rcuidle ->
srcu_read_lock_notrace -> __srcu_read_lock -> kcsan ...
While we could disable KCSAN in kcsan_setup_watchpoint(), this does not
solve other unexpected behaviour we may get due recursing into functions
that may not be tolerant to such recursion:
__srcu_read_lock -> kcsan -> ... -> __srcu_read_lock
Therefore, switch to using prandom_u32_state(), which is uninstrumented,
and does not have a tracepoint.
Link: https://lkml.kernel.org/r/20200821063043.1949509-1-elver@google.com
Link: https://lkml.kernel.org/r/20200820172046.GA177701@elver.google.com
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Remove kcsan_counter_inc/dec() functions, as they perform no other
logic, and are no longer needed.
This avoids several calls in kcsan_setup_watchpoint() and
kcsan_found_watchpoint(), as well as lets the compiler warn us about
potential out-of-bounds accesses as the array's size is known at all
usage sites at compile-time.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Use the same pr_fmt throughout for consistency. [ The only exception is
report.c, where the format must be kept precisely as-is. ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Show a message in the kernel log if KCSAN was enabled early.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Remove the debugfs test command, as it is no longer needed now that we
have the KUnit+Torture based kcsan-test module. This is to avoid
confusion around how KCSAN should be tested, as only the kcsan-test
module is maintained.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Simplify checking prefixes and length calculation of constant strings.
For the former, the kernel provides str_has_prefix(), and the latter we
should just use strlen("..") because GCC and Clang have optimizations
that optimize these into constants.
No functional change intended.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Simplify counter ID to name mapping by using an array with designated
inits. This way, we can turn a run-time BUG() into a compile-time static
assertion failure if a counter name is missing.
No functional change intended.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Changes kcsan-test module to support checking reports that include
compound instrumentation. Since we should not fail the test if this
support is unavailable, we have to add a config variable that the test
can use to decide what to check for.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Add missing CONFIG_KCSAN_IGNORE_ATOMICS checks for the builtin atomics
instrumentation.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
For compound instrumentation and assert accesses, skew the watchpoint
delay to be longer if randomized. This is useful to improve race
detection for such accesses.
For compound accesses we should increase the delay as we've aggregated
both read and write instrumentation. By giving up 1 call into the
runtime, we're less likely to set up a watchpoint and thus less likely
to detect a race. We can balance this by increasing the watchpoint
delay.
For assert accesses, we know these are of increased interest, and we
wish to increase our chances of detecting races for such checks.
Note that, kcsan_udelay_{task,interrupt} define the upper bound delays.
When randomized, delays are uniformly distributed between [0, delay].
Skewing the delay does not break this promise as long as the defined
upper bounds are still adhered to. The current skew results in delays
uniformly distributed between [delay/2, delay].
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Add support for compounded read-write instrumentation if supported by
the compiler. Adds the necessary instrumentation functions, and a new
type which is used to generate a more descriptive report.
Furthermore, such compounded memory access instrumentation is excluded
from the "assume aligned writes up to word size are atomic" rule,
because we cannot assume that the compiler emits code that is atomic for
compound ops.
LLVM/Clang added support for the feature in:
785d41a261
The new instrumentation is emitted for sets of memory accesses in the
same basic block to the same address with at least one read appearing
before a write. These typically result from compound operations such as
++, --, +=, -=, |=, &=, etc. but also equivalent forms such as "var =
var + 1". Where the compiler determines that it is equivalent to emit a
call to a single __tsan_read_write instead of separate __tsan_read and
__tsan_write, we can then benefit from improved performance and better
reporting for such access patterns.
The new reports now show that the ops are both reads and writes, for
example:
read-write to 0xffffffff90548a38 of 8 bytes by task 143 on cpu 3:
test_kernel_rmw_array+0x45/0xa0
access_thread+0x71/0xb0
kthread+0x21e/0x240
ret_from_fork+0x22/0x30
read-write to 0xffffffff90548a38 of 8 bytes by task 144 on cpu 2:
test_kernel_rmw_array+0x45/0xa0
access_thread+0x71/0xb0
kthread+0x21e/0x240
ret_from_fork+0x22/0x30
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Adds test case to kcsan-test module, to test atomic builtin
instrumentation works.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Some architectures (currently e.g. s390 partially) implement atomics
using the compiler's atomic builtins (__atomic_*, __sync_*). To support
enabling KCSAN on such architectures in future, or support experimental
use of these builtins, implement support for them.
We should also avoid breaking KCSAN kernels due to use (accidental or
otherwise) of atomic builtins in drivers, as has happened in the past:
https://lkml.kernel.org/r/5231d2c0-41d9-6721-e15f-a7eedf3ce69e@infradead.org
The instrumentation is subtly different from regular reads/writes: TSAN
instrumentation replaces the use of atomic builtins with a call into the
runtime, and the runtime's job is to also execute the desired atomic
operation. We rely on the __atomic_* compiler builtins, available with
all KCSAN-supported compilers, to implement each TSAN atomic
instrumentation function.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Pull v5.9 KCSAN bits from Paul E. McKenney.
Perhaps the most important change is that GCC 11 now has all fixes in place
to support KCSAN, so GCC support can be enabled again.
Signed-off-by: Ingo Molnar <mingo@kernel.org>
To improve the general usefulness of the IRQ state trace events with
KCSAN enabled, save and restore the trace information when entering and
exiting the KCSAN runtime as well as when generating a KCSAN report.
Without this, reporting the IRQ trace events (whether via a KCSAN report
or outside of KCSAN via a lockdep report) is rather useless due to
continuously being touched by KCSAN. This is because if KCSAN is
enabled, every instrumented memory access causes changes to IRQ trace
events (either by KCSAN disabling/enabling interrupts or taking
report_lock when generating a report).
Before "lockdep: Prepare for NMI IRQ state tracking", KCSAN avoided
touching the IRQ trace events via raw_local_irq_save/restore() and
lockdep_off/on().
Fixes: 248591f5d2 ("kcsan: Make KCSAN compatible with new IRQ state tracking")
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Link: https://lore.kernel.org/r/20200729110916.3920464-2-elver@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The new IRQ state tracking code does not honor lockdep_off(), and as
such we should again permit tracing by using non-raw functions in
core.c. Update the lockdep_off() comment in report.c, to reflect the
fact there is still a potential risk of deadlock due to using printk()
from scheduler code.
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Ingo Molnar <mingo@kernel.org>
Link: https://lkml.kernel.org/r/20200624113246.GA170324@elver.google.com
Disable branch tracing in core KCSAN runtime if branches are being
traced (TRACE_BRANCH_PROFILING). This it to avoid its performance
impact, but also avoid recursion in case KCSAN is enabled for the branch
tracing runtime.
The latter had already been a problem for KASAN:
https://lore.kernel.org/lkml/CANpmjNOeXmD5E3O50Z3MjkiuCYaYOPyi+1rq=GZvEKwBvLR0Ug@mail.gmail.com/
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Simplify the set of compiler flags for the runtime by removing cc-option
from -fno-stack-protector, because all supported compilers support it.
This saves us one compiler invocation during build.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Add a test that KCSAN nor the compiler gets confused about accesses to
jiffies on different architectures.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Remove existing special atomic rules from kcsan_is_atomic_special()
because they are no longer needed. Since we rely on the compiler
emitting instrumentation distinguishing volatile accesses, the rules
have become redundant.
Let's keep kcsan_is_atomic_special() around, so that we have an obvious
place to add special rules should the need arise in future.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Rename 'test.c' to 'selftest.c' to better reflect its purpose (Kconfig
variable and code inside already match this). This is to avoid confusion
with the test suite module in 'kcsan-test.c'.
No functional change.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
The functions here should not be forward declared for explicit use
elsewhere in the kernel, as they should only be emitted by the compiler
due to sanitizer instrumentation. Add forward declarations a line above
their definition to shut up warnings in W=1 builds.
Link: https://lkml.kernel.org/r/202006060103.jSCpnV1g%lkp@intel.com
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Instead of __no_kcsan_or_inline, prefer '__no_kcsan inline' in test --
this is in case we decide to remove __no_kcsan_or_inline.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
This adds KCSAN test focusing on behaviour of the integrated runtime.
Tests various race scenarios, and verifies the reports generated to
console. Makes use of KUnit for test organization, and the Torture
framework for test thread control.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
In the kernel, the "volatile" keyword is used in various concurrent
contexts, whether in low-level synchronization primitives or for
legacy reasons. If supported by the compiler, it will be assumed
that aligned volatile accesses up to sizeof(long long) (matching
compiletime_assert_rwonce_type()) are atomic.
Recent versions of Clang [1] (GCC tentative [2]) can instrument
volatile accesses differently. Add the option (required) to enable the
instrumentation, and provide the necessary runtime functions. None of
the updated compilers are widely available yet (Clang 11 will be the
first release to support the feature).
[1] 5a2c31116f
[2] https://gcc.gnu.org/pipermail/gcc-patches/2020-April/544452.html
This change allows removing of any explicit checks in primitives such as
READ_ONCE() and WRITE_ONCE().
[ bp: Massage commit message a bit. ]
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Will Deacon <will@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20200521142047.169334-4-elver@google.com
The __kcsan_{enable,disable}_current() variants only call into KCSAN if
KCSAN is enabled for the current compilation unit. Note: This is
typically not what we want, as we usually want to ensure that even calls
into other functions still have KCSAN disabled.
These variants may safely be used in header files that are shared
between regular kernel code and code that does not link the KCSAN
runtime.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
A spin lock is held in insert_report_filterlist(), so the krealloc()
should use GFP_ATOMIC. This commit therefore makes this change.
Reviewed-by: Marco Elver <elver@google.com>
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Reporting hides KCSAN runtime functions in the stack trace, with
filtering done based on function names. Currently this included all
functions (or modules) that would match "kcsan_". Make the filter aware
of KCSAN tests, which contain "kcsan_test", and are no longer skipped in
the report.
This is in preparation for adding a KCSAN test module.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
Pass string length as returned by scnprintf() to strnstr(), since
strnstr() searches exactly len bytes in haystack, even if it contains a
NUL-terminator before haystack+len.
Signed-off-by: Marco Elver <elver@google.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>