linux-stable/rust/kernel/init.rs

1345 lines
45 KiB
Rust
Raw Normal View History

rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
// SPDX-License-Identifier: Apache-2.0 OR MIT
//! API to safely and fallibly initialize pinned `struct`s using in-place constructors.
//!
//! It also allows in-place initialization of big `struct`s that would otherwise produce a stack
//! overflow.
//!
//! Most `struct`s from the [`sync`] module need to be pinned, because they contain self-referential
//! `struct`s from C. [Pinning][pinning] is Rust's way of ensuring data does not move.
//!
//! # Overview
//!
//! To initialize a `struct` with an in-place constructor you will need two things:
//! - an in-place constructor,
//! - a memory location that can hold your `struct` (this can be the [stack], an [`Arc<T>`],
//! [`UniqueArc<T>`], [`Box<T>`] or any other smart pointer that implements [`InPlaceInit`]).
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
//!
//! To get an in-place constructor there are generally three options:
//! - directly creating an in-place constructor using the [`pin_init!`] macro,
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
//! - a custom function/macro returning an in-place constructor provided by someone else,
//! - using the unsafe function [`pin_init_from_closure()`] to manually create an initializer.
//!
//! Aside from pinned initialization, this API also supports in-place construction without pinning,
//! the macros/types/functions are generally named like the pinned variants without the `pin`
//! prefix.
//!
//! # Examples
//!
//! ## Using the [`pin_init!`] macro
//!
//! If you want to use [`PinInit`], then you will have to annotate your `struct` with
//! `#[`[`pin_data`]`]`. It is a macro that uses `#[pin]` as a marker for
//! [structurally pinned fields]. After doing this, you can then create an in-place constructor via
//! [`pin_init!`]. The syntax is almost the same as normal `struct` initializers. The difference is
//! that you need to write `<-` instead of `:` for fields that you want to initialize in-place.
//!
//! ```rust
//! # #![allow(clippy::disallowed_names)]
//! use kernel::sync::{new_mutex, Mutex};
//! # use core::pin::Pin;
//! #[pin_data]
//! struct Foo {
//! #[pin]
//! a: Mutex<usize>,
//! b: u32,
//! }
//!
//! let foo = pin_init!(Foo {
//! a <- new_mutex!(42, "Foo::a"),
//! b: 24,
//! });
//! ```
//!
//! `foo` now is of the type [`impl PinInit<Foo>`]. We can now use any smart pointer that we like
//! (or just the stack) to actually initialize a `Foo`:
//!
//! ```rust
//! # #![allow(clippy::disallowed_names)]
//! # use kernel::sync::{new_mutex, Mutex};
//! # use core::pin::Pin;
//! # #[pin_data]
//! # struct Foo {
//! # #[pin]
//! # a: Mutex<usize>,
//! # b: u32,
//! # }
//! # let foo = pin_init!(Foo {
//! # a <- new_mutex!(42, "Foo::a"),
//! # b: 24,
//! # });
//! let foo: Result<Pin<Box<Foo>>> = Box::pin_init(foo);
//! ```
//!
//! For more information see the [`pin_init!`] macro.
//!
//! ## Using a custom function/macro that returns an initializer
//!
//! Many types from the kernel supply a function/macro that returns an initializer, because the
//! above method only works for types where you can access the fields.
//!
//! ```rust
//! # use kernel::sync::{new_mutex, Arc, Mutex};
//! let mtx: Result<Arc<Mutex<usize>>> = Arc::pin_init(new_mutex!(42, "example::mtx"));
//! ```
//!
//! To declare an init macro/function you just return an [`impl PinInit<T, E>`]:
//!
//! ```rust
//! # #![allow(clippy::disallowed_names)]
//! # use kernel::{sync::Mutex, prelude::*, new_mutex, init::PinInit, try_pin_init};
//! #[pin_data]
//! struct DriverData {
//! #[pin]
//! status: Mutex<i32>,
//! buffer: Box<[u8; 1_000_000]>,
//! }
//!
//! impl DriverData {
//! fn new() -> impl PinInit<Self, Error> {
//! try_pin_init!(Self {
//! status <- new_mutex!(0, "DriverData::status"),
//! buffer: Box::init(kernel::init::zeroed())?,
//! })
//! }
//! }
//! ```
//!
//! ## Manual creation of an initializer
//!
//! Often when working with primitives the previous approaches are not sufficient. That is where
//! [`pin_init_from_closure()`] comes in. This `unsafe` function allows you to create a
//! [`impl PinInit<T, E>`] directly from a closure. Of course you have to ensure that the closure
//! actually does the initialization in the correct way. Here are the things to look out for
//! (we are calling the parameter to the closure `slot`):
//! - when the closure returns `Ok(())`, then it has completed the initialization successfully, so
//! `slot` now contains a valid bit pattern for the type `T`,
//! - when the closure returns `Err(e)`, then the caller may deallocate the memory at `slot`, so
//! you need to take care to clean up anything if your initialization fails mid-way,
//! - you may assume that `slot` will stay pinned even after the closure returns until `drop` of
//! `slot` gets called.
//!
//! ```rust
//! # #![allow(unreachable_pub, clippy::disallowed_names)]
//! use kernel::{prelude::*, init, types::Opaque};
//! use core::{ptr::addr_of_mut, marker::PhantomPinned, pin::Pin};
//! # mod bindings {
//! # #![allow(non_camel_case_types)]
//! # pub struct foo;
//! # pub unsafe fn init_foo(_ptr: *mut foo) {}
//! # pub unsafe fn destroy_foo(_ptr: *mut foo) {}
//! # pub unsafe fn enable_foo(_ptr: *mut foo, _flags: u32) -> i32 { 0 }
//! # }
//! # // `Error::from_errno` is `pub(crate)` in the `kernel` crate, thus provide a workaround.
//! # trait FromErrno {
//! # fn from_errno(errno: core::ffi::c_int) -> Error {
//! # // Dummy error that can be constructed outside the `kernel` crate.
//! # Error::from(core::fmt::Error)
//! # }
//! # }
//! # impl FromErrno for Error {}
//! /// # Invariants
//! ///
//! /// `foo` is always initialized
//! #[pin_data(PinnedDrop)]
//! pub struct RawFoo {
//! #[pin]
//! foo: Opaque<bindings::foo>,
//! #[pin]
//! _p: PhantomPinned,
//! }
//!
//! impl RawFoo {
//! pub fn new(flags: u32) -> impl PinInit<Self, Error> {
//! // SAFETY:
//! // - when the closure returns `Ok(())`, then it has successfully initialized and
//! // enabled `foo`,
//! // - when it returns `Err(e)`, then it has cleaned up before
//! unsafe {
//! init::pin_init_from_closure(move |slot: *mut Self| {
//! // `slot` contains uninit memory, avoid creating a reference.
//! let foo = addr_of_mut!((*slot).foo);
//!
//! // Initialize the `foo`
//! bindings::init_foo(Opaque::raw_get(foo));
//!
//! // Try to enable it.
//! let err = bindings::enable_foo(Opaque::raw_get(foo), flags);
//! if err != 0 {
//! // Enabling has failed, first clean up the foo and then return the error.
//! bindings::destroy_foo(Opaque::raw_get(foo));
//! return Err(Error::from_errno(err));
//! }
//!
//! // All fields of `RawFoo` have been initialized, since `_p` is a ZST.
//! Ok(())
//! })
//! }
//! }
//! }
//!
//! #[pinned_drop]
//! impl PinnedDrop for RawFoo {
//! fn drop(self: Pin<&mut Self>) {
//! // SAFETY: Since `foo` is initialized, destroying is safe.
//! unsafe { bindings::destroy_foo(self.foo.get()) };
//! }
//! }
//! ```
//!
//! For the special case where initializing a field is a single FFI-function call that cannot fail,
//! there exist the helper function [`Opaque::ffi_init`]. This function initialize a single
//! [`Opaque`] field by just delegating to the supplied closure. You can use these in combination
//! with [`pin_init!`].
//!
//! For more information on how to use [`pin_init_from_closure()`], take a look at the uses inside
//! the `kernel` crate. The [`sync`] module is a good starting point.
//!
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
//! [`sync`]: kernel::sync
//! [pinning]: https://doc.rust-lang.org/std/pin/index.html
//! [structurally pinned fields]:
//! https://doc.rust-lang.org/std/pin/index.html#pinning-is-structural-for-field
//! [stack]: crate::stack_pin_init
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
//! [`Arc<T>`]: crate::sync::Arc
//! [`impl PinInit<Foo>`]: PinInit
//! [`impl PinInit<T, E>`]: PinInit
//! [`impl Init<T, E>`]: Init
//! [`Opaque`]: kernel::types::Opaque
//! [`Opaque::ffi_init`]: kernel::types::Opaque::ffi_init
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
//! [`pin_data`]: ::macros::pin_data
rust: upgrade to Rust 1.68.2 This is the first upgrade to the Rust toolchain since the initial Rust merge, from 1.62.0 to 1.68.2 (i.e. the latest). # Context The kernel currently supports only a single Rust version [1] (rather than a minimum) given our usage of some "unstable" Rust features [2] which do not promise backwards compatibility. The goal is to reach a point where we can declare a minimum version for the toolchain. For instance, by waiting for some of the features to be stabilized. Therefore, the first minimum Rust version that the kernel will support is "in the future". # Upgrade policy Given we will eventually need to reach that minimum version, it would be ideal to upgrade the compiler from time to time to be as close as possible to that goal and find any issues sooner. In the extreme, we could upgrade as soon as a new Rust release is out. Of course, upgrading so often is in stark contrast to what one normally would need for GCC and LLVM, especially given the release schedule: 6 weeks for Rust vs. half a year for LLVM and a year for GCC. Having said that, there is no particular advantage to updating slowly either: kernel developers in "stable" distributions are unlikely to be able to use their distribution-provided Rust toolchain for the kernel anyway [3]. Instead, by routinely upgrading to the latest instead, kernel developers using Linux distributions that track the latest Rust release may be able to use those rather than Rust-provided ones, especially if their package manager allows to pin / hold back / downgrade the version for some days during windows where the version may not match. For instance, Arch, Fedora, Gentoo and openSUSE all provide and track the latest version of Rust as they get released every 6 weeks. Then, when the minimum version is reached, we will stop upgrading and decide how wide the window of support will be. For instance, a year of Rust versions. We will probably want to start small, and then widen it over time, just like the kernel did originally for LLVM, see commit 3519c4d6e08e ("Documentation: add minimum clang/llvm version"). # Unstable features stabilized This upgrade allows us to remove the following unstable features since they were stabilized: - `feature(explicit_generic_args_with_impl_trait)` (1.63). - `feature(core_ffi_c)` (1.64). - `feature(generic_associated_types)` (1.65). - `feature(const_ptr_offset_from)` (1.65, *). - `feature(bench_black_box)` (1.66, *). - `feature(pin_macro)` (1.68). The ones marked with `*` apply only to our old `rust` branch, not mainline yet, i.e. only for code that we may potentially upstream. With this patch applied, the only unstable feature allowed to be used outside the `kernel` crate is `new_uninit`, though other code to be upstreamed may increase the list. Please see [2] for details. # Other required changes Since 1.63, `rustdoc` triggers the `broken_intra_doc_links` lint for links pointing to exported (`#[macro_export]`) `macro_rules`. An issue was opened upstream [4], but it turns out it is intended behavior. For the moment, just add an explicit reference for each link. Later we can revisit this if `rustdoc` removes the compatibility measure. Nevertheless, this was helpful to discover a link that was pointing to the wrong place unintentionally. Since that one was actually wrong, it is fixed in a previous commit independently. Another change was the addition of `cfg(no_rc)` and `cfg(no_sync)` in upstream [5], thus remove our original changes for that. Similarly, upstream now tests that it compiles successfully with `#[cfg(not(no_global_oom_handling))]` [6], which allow us to get rid of some changes, such as an `#[allow(dead_code)]`. In addition, remove another `#[allow(dead_code)]` due to new uses within the standard library. Finally, add `try_extend_trusted` and move the code in `spec_extend.rs` since upstream moved it for the infallible version. # `alloc` upgrade and reviewing There are a large amount of changes, but the vast majority of them are due to our `alloc` fork being upgraded at once. There are two kinds of changes to be aware of: the ones coming from upstream, which we should follow as closely as possible, and the updates needed in our added fallible APIs to keep them matching the newer infallible APIs coming from upstream. Instead of taking a look at the diff of this patch, an alternative approach is reviewing a diff of the changes between upstream `alloc` and the kernel's. This allows to easily inspect the kernel additions only, especially to check if the fallible methods we already have still match the infallible ones in the new version coming from upstream. Another approach is reviewing the changes introduced in the additions in the kernel fork between the two versions. This is useful to spot potentially unintended changes to our additions. To apply these approaches, one may follow steps similar to the following to generate a pair of patches that show the differences between upstream Rust and the kernel (for the subset of `alloc` we use) before and after applying this patch: # Get the difference with respect to the old version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > old.patch git -C linux restore rust/alloc # Apply this patch. git -C linux am rust-upgrade.patch # Get the difference with respect to the new version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > new.patch git -C linux restore rust/alloc Now one may check the `new.patch` to take a look at the additions (first approach) or at the difference between those two patches (second approach). For the latter, a side-by-side tool is recommended. Link: https://rust-for-linux.com/rust-version-policy [1] Link: https://github.com/Rust-for-Linux/linux/issues/2 [2] Link: https://lore.kernel.org/rust-for-linux/CANiq72mT3bVDKdHgaea-6WiZazd8Mvurqmqegbe5JZxVyLR8Yg@mail.gmail.com/ [3] Link: https://github.com/rust-lang/rust/issues/106142 [4] Link: https://github.com/rust-lang/rust/pull/89891 [5] Link: https://github.com/rust-lang/rust/pull/98652 [6] Reviewed-by: Björn Roy Baron <bjorn3_gh@protonmail.com> Reviewed-by: Gary Guo <gary@garyguo.net> Reviewed-By: Martin Rodriguez Reboredo <yakoyoku@gmail.com> Tested-by: Ariel Miculas <amiculas@cisco.com> Tested-by: David Gow <davidgow@google.com> Tested-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20230418214347.324156-4-ojeda@kernel.org [ Removed `feature(core_ffi_c)` from `uapi` ] Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-18 21:43:47 +00:00
//! [`pin_init!`]: crate::pin_init!
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
use crate::{
error::{self, Error},
sync::UniqueArc,
types::{Opaque, ScopeGuard},
};
use alloc::boxed::Box;
use core::{
alloc::AllocError,
cell::UnsafeCell,
convert::Infallible,
marker::PhantomData,
mem::MaybeUninit,
num::*,
pin::Pin,
ptr::{self, NonNull},
};
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
#[doc(hidden)]
pub mod __internal;
#[doc(hidden)]
pub mod macros;
/// Initialize and pin a type directly on the stack.
///
/// # Examples
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, macros::pin_data, pin_init, stack_pin_init, init::*, sync::Mutex, new_mutex};
/// # use core::pin::Pin;
/// #[pin_data]
/// struct Foo {
/// #[pin]
/// a: Mutex<usize>,
/// b: Bar,
/// }
///
/// #[pin_data]
/// struct Bar {
/// x: u32,
/// }
///
/// stack_pin_init!(let foo = pin_init!(Foo {
/// a <- new_mutex!(42),
/// b: Bar {
/// x: 64,
/// },
/// }));
/// let foo: Pin<&mut Foo> = foo;
/// pr_info!("a: {}", &*foo.a.lock());
/// ```
///
/// # Syntax
///
/// A normal `let` binding with optional type annotation. The expression is expected to implement
/// [`PinInit`]/[`Init`] with the error type [`Infallible`]. If you want to use a different error
/// type, then use [`stack_try_pin_init!`].
rust: upgrade to Rust 1.68.2 This is the first upgrade to the Rust toolchain since the initial Rust merge, from 1.62.0 to 1.68.2 (i.e. the latest). # Context The kernel currently supports only a single Rust version [1] (rather than a minimum) given our usage of some "unstable" Rust features [2] which do not promise backwards compatibility. The goal is to reach a point where we can declare a minimum version for the toolchain. For instance, by waiting for some of the features to be stabilized. Therefore, the first minimum Rust version that the kernel will support is "in the future". # Upgrade policy Given we will eventually need to reach that minimum version, it would be ideal to upgrade the compiler from time to time to be as close as possible to that goal and find any issues sooner. In the extreme, we could upgrade as soon as a new Rust release is out. Of course, upgrading so often is in stark contrast to what one normally would need for GCC and LLVM, especially given the release schedule: 6 weeks for Rust vs. half a year for LLVM and a year for GCC. Having said that, there is no particular advantage to updating slowly either: kernel developers in "stable" distributions are unlikely to be able to use their distribution-provided Rust toolchain for the kernel anyway [3]. Instead, by routinely upgrading to the latest instead, kernel developers using Linux distributions that track the latest Rust release may be able to use those rather than Rust-provided ones, especially if their package manager allows to pin / hold back / downgrade the version for some days during windows where the version may not match. For instance, Arch, Fedora, Gentoo and openSUSE all provide and track the latest version of Rust as they get released every 6 weeks. Then, when the minimum version is reached, we will stop upgrading and decide how wide the window of support will be. For instance, a year of Rust versions. We will probably want to start small, and then widen it over time, just like the kernel did originally for LLVM, see commit 3519c4d6e08e ("Documentation: add minimum clang/llvm version"). # Unstable features stabilized This upgrade allows us to remove the following unstable features since they were stabilized: - `feature(explicit_generic_args_with_impl_trait)` (1.63). - `feature(core_ffi_c)` (1.64). - `feature(generic_associated_types)` (1.65). - `feature(const_ptr_offset_from)` (1.65, *). - `feature(bench_black_box)` (1.66, *). - `feature(pin_macro)` (1.68). The ones marked with `*` apply only to our old `rust` branch, not mainline yet, i.e. only for code that we may potentially upstream. With this patch applied, the only unstable feature allowed to be used outside the `kernel` crate is `new_uninit`, though other code to be upstreamed may increase the list. Please see [2] for details. # Other required changes Since 1.63, `rustdoc` triggers the `broken_intra_doc_links` lint for links pointing to exported (`#[macro_export]`) `macro_rules`. An issue was opened upstream [4], but it turns out it is intended behavior. For the moment, just add an explicit reference for each link. Later we can revisit this if `rustdoc` removes the compatibility measure. Nevertheless, this was helpful to discover a link that was pointing to the wrong place unintentionally. Since that one was actually wrong, it is fixed in a previous commit independently. Another change was the addition of `cfg(no_rc)` and `cfg(no_sync)` in upstream [5], thus remove our original changes for that. Similarly, upstream now tests that it compiles successfully with `#[cfg(not(no_global_oom_handling))]` [6], which allow us to get rid of some changes, such as an `#[allow(dead_code)]`. In addition, remove another `#[allow(dead_code)]` due to new uses within the standard library. Finally, add `try_extend_trusted` and move the code in `spec_extend.rs` since upstream moved it for the infallible version. # `alloc` upgrade and reviewing There are a large amount of changes, but the vast majority of them are due to our `alloc` fork being upgraded at once. There are two kinds of changes to be aware of: the ones coming from upstream, which we should follow as closely as possible, and the updates needed in our added fallible APIs to keep them matching the newer infallible APIs coming from upstream. Instead of taking a look at the diff of this patch, an alternative approach is reviewing a diff of the changes between upstream `alloc` and the kernel's. This allows to easily inspect the kernel additions only, especially to check if the fallible methods we already have still match the infallible ones in the new version coming from upstream. Another approach is reviewing the changes introduced in the additions in the kernel fork between the two versions. This is useful to spot potentially unintended changes to our additions. To apply these approaches, one may follow steps similar to the following to generate a pair of patches that show the differences between upstream Rust and the kernel (for the subset of `alloc` we use) before and after applying this patch: # Get the difference with respect to the old version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > old.patch git -C linux restore rust/alloc # Apply this patch. git -C linux am rust-upgrade.patch # Get the difference with respect to the new version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > new.patch git -C linux restore rust/alloc Now one may check the `new.patch` to take a look at the additions (first approach) or at the difference between those two patches (second approach). For the latter, a side-by-side tool is recommended. Link: https://rust-for-linux.com/rust-version-policy [1] Link: https://github.com/Rust-for-Linux/linux/issues/2 [2] Link: https://lore.kernel.org/rust-for-linux/CANiq72mT3bVDKdHgaea-6WiZazd8Mvurqmqegbe5JZxVyLR8Yg@mail.gmail.com/ [3] Link: https://github.com/rust-lang/rust/issues/106142 [4] Link: https://github.com/rust-lang/rust/pull/89891 [5] Link: https://github.com/rust-lang/rust/pull/98652 [6] Reviewed-by: Björn Roy Baron <bjorn3_gh@protonmail.com> Reviewed-by: Gary Guo <gary@garyguo.net> Reviewed-By: Martin Rodriguez Reboredo <yakoyoku@gmail.com> Tested-by: Ariel Miculas <amiculas@cisco.com> Tested-by: David Gow <davidgow@google.com> Tested-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20230418214347.324156-4-ojeda@kernel.org [ Removed `feature(core_ffi_c)` from `uapi` ] Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-18 21:43:47 +00:00
///
/// [`stack_try_pin_init!`]: crate::stack_try_pin_init!
#[macro_export]
macro_rules! stack_pin_init {
(let $var:ident $(: $t:ty)? = $val:expr) => {
let val = $val;
let mut $var = ::core::pin::pin!($crate::init::__internal::StackInit$(::<$t>)?::uninit());
let mut $var = match $crate::init::__internal::StackInit::init($var, val) {
Ok(res) => res,
Err(x) => {
let x: ::core::convert::Infallible = x;
match x {}
}
};
};
}
/// Initialize and pin a type directly on the stack.
///
/// # Examples
///
/// ```rust,ignore
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, stack_try_pin_init, init::*, sync::Mutex, new_mutex};
/// # use macros::pin_data;
/// # use core::{alloc::AllocError, pin::Pin};
/// #[pin_data]
/// struct Foo {
/// #[pin]
/// a: Mutex<usize>,
/// b: Box<Bar>,
/// }
///
/// struct Bar {
/// x: u32,
/// }
///
/// stack_try_pin_init!(let foo: Result<Pin<&mut Foo>, AllocError> = pin_init!(Foo {
/// a <- new_mutex!(42),
/// b: Box::try_new(Bar {
/// x: 64,
/// })?,
/// }));
/// let foo = foo.unwrap();
/// pr_info!("a: {}", &*foo.a.lock());
/// ```
///
/// ```rust,ignore
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, stack_try_pin_init, init::*, sync::Mutex, new_mutex};
/// # use macros::pin_data;
/// # use core::{alloc::AllocError, pin::Pin};
/// #[pin_data]
/// struct Foo {
/// #[pin]
/// a: Mutex<usize>,
/// b: Box<Bar>,
/// }
///
/// struct Bar {
/// x: u32,
/// }
///
/// stack_try_pin_init!(let foo: Pin<&mut Foo> =? pin_init!(Foo {
/// a <- new_mutex!(42),
/// b: Box::try_new(Bar {
/// x: 64,
/// })?,
/// }));
/// pr_info!("a: {}", &*foo.a.lock());
/// # Ok::<_, AllocError>(())
/// ```
///
/// # Syntax
///
/// A normal `let` binding with optional type annotation. The expression is expected to implement
/// [`PinInit`]/[`Init`]. This macro assigns a result to the given variable, adding a `?` after the
/// `=` will propagate this error.
#[macro_export]
macro_rules! stack_try_pin_init {
(let $var:ident $(: $t:ty)? = $val:expr) => {
let val = $val;
let mut $var = ::core::pin::pin!($crate::init::__internal::StackInit$(::<$t>)?::uninit());
let mut $var = $crate::init::__internal::StackInit::init($var, val);
};
(let $var:ident $(: $t:ty)? =? $val:expr) => {
let val = $val;
let mut $var = ::core::pin::pin!($crate::init::__internal::StackInit$(::<$t>)?::uninit());
let mut $var = $crate::init::__internal::StackInit::init($var, val)?;
};
}
/// Construct an in-place, pinned initializer for `struct`s.
///
/// This macro defaults the error to [`Infallible`]. If you need [`Error`], then use
/// [`try_pin_init!`].
///
/// The syntax is almost identical to that of a normal `struct` initializer:
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, macros::pin_data, init::*};
/// # use core::pin::Pin;
/// #[pin_data]
/// struct Foo {
/// a: usize,
/// b: Bar,
/// }
///
/// #[pin_data]
/// struct Bar {
/// x: u32,
/// }
///
/// # fn demo() -> impl PinInit<Foo> {
/// let a = 42;
///
/// let initializer = pin_init!(Foo {
/// a,
/// b: Bar {
/// x: 64,
/// },
/// });
/// # initializer }
/// # Box::pin_init(demo()).unwrap();
/// ```
///
/// Arbitrary Rust expressions can be used to set the value of a variable.
///
/// The fields are initialized in the order that they appear in the initializer. So it is possible
/// to read already initialized fields using raw pointers.
///
/// IMPORTANT: You are not allowed to create references to fields of the struct inside of the
/// initializer.
///
/// # Init-functions
///
/// When working with this API it is often desired to let others construct your types without
/// giving access to all fields. This is where you would normally write a plain function `new`
/// that would return a new instance of your type. With this API that is also possible.
/// However, there are a few extra things to keep in mind.
///
/// To create an initializer function, simply declare it like this:
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, prelude::*, init::*};
/// # use core::pin::Pin;
/// # #[pin_data]
/// # struct Foo {
/// # a: usize,
/// # b: Bar,
/// # }
/// # #[pin_data]
/// # struct Bar {
/// # x: u32,
/// # }
/// impl Foo {
/// fn new() -> impl PinInit<Self> {
/// pin_init!(Self {
/// a: 42,
/// b: Bar {
/// x: 64,
/// },
/// })
/// }
/// }
/// ```
///
/// Users of `Foo` can now create it like this:
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, macros::pin_data, init::*};
/// # use core::pin::Pin;
/// # #[pin_data]
/// # struct Foo {
/// # a: usize,
/// # b: Bar,
/// # }
/// # #[pin_data]
/// # struct Bar {
/// # x: u32,
/// # }
/// # impl Foo {
/// # fn new() -> impl PinInit<Self> {
/// # pin_init!(Self {
/// # a: 42,
/// # b: Bar {
/// # x: 64,
/// # },
/// # })
/// # }
/// # }
/// let foo = Box::pin_init(Foo::new());
/// ```
///
/// They can also easily embed it into their own `struct`s:
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// # use kernel::{init, pin_init, macros::pin_data, init::*};
/// # use core::pin::Pin;
/// # #[pin_data]
/// # struct Foo {
/// # a: usize,
/// # b: Bar,
/// # }
/// # #[pin_data]
/// # struct Bar {
/// # x: u32,
/// # }
/// # impl Foo {
/// # fn new() -> impl PinInit<Self> {
/// # pin_init!(Self {
/// # a: 42,
/// # b: Bar {
/// # x: 64,
/// # },
/// # })
/// # }
/// # }
/// #[pin_data]
/// struct FooContainer {
/// #[pin]
/// foo1: Foo,
/// #[pin]
/// foo2: Foo,
/// other: u32,
/// }
///
/// impl FooContainer {
/// fn new(other: u32) -> impl PinInit<Self> {
/// pin_init!(Self {
/// foo1 <- Foo::new(),
/// foo2 <- Foo::new(),
/// other,
/// })
/// }
/// }
/// ```
///
/// Here we see that when using `pin_init!` with `PinInit`, one needs to write `<-` instead of `:`.
/// This signifies that the given field is initialized in-place. As with `struct` initializers, just
/// writing the field (in this case `other`) without `:` or `<-` means `other: other,`.
///
/// # Syntax
///
/// As already mentioned in the examples above, inside of `pin_init!` a `struct` initializer with
/// the following modifications is expected:
/// - Fields that you want to initialize in-place have to use `<-` instead of `:`.
/// - In front of the initializer you can write `&this in` to have access to a [`NonNull<Self>`]
/// pointer named `this` inside of the initializer.
/// - Using struct update syntax one can place `..Zeroable::zeroed()` at the very end of the
/// struct, this initializes every field with 0 and then runs all initializers specified in the
/// body. This can only be done if [`Zeroable`] is implemented for the struct.
///
/// For instance:
///
/// ```rust
Rust changes for v6.6 In terms of lines, most changes this time are on the pinned-init API and infrastructure. While we have a Rust version upgrade, and thus a bunch of changes from the vendored 'alloc' crate as usual, this time those do not account for many lines. Toolchain and infrastructure: - Upgrade to Rust 1.71.1. This is the second such upgrade, which is a smaller jump compared to the last time. This version allows us to remove the '__rust_*' allocator functions -- the compiler now generates them as expected, thus now our 'KernelAllocator' is used. It also introduces the 'offset_of!' macro in the standard library (as an unstable feature) which we will need soon. So far, we were using a declarative macro as a prerequisite in some not-yet-landed patch series, which did not support sub-fields (i.e. nested structs): #[repr(C)] struct S { a: u16, b: (u8, u8), } assert_eq!(offset_of!(S, b.1), 3); - Upgrade to bindgen 0.65.1. This is the first time we upgrade its version. Given it is a fairly big jump, it comes with a fair number of improvements/changes that affect us, such as a fix needed to support LLVM 16 as well as proper support for '__noreturn' C functions, which are now mapped to return the '!' type in Rust: void __noreturn f(void); // C pub fn f() -> !; // Rust - 'scripts/rust_is_available.sh' improvements and fixes. This series takes care of all the issues known so far and adds a few new checks to cover for even more cases, plus adds some more help texts. All this together will hopefully make problematic setups easier to identify and to be solved by users building the kernel. In addition, it adds a test suite which covers all branches of the shell script, as well as tests for the issues found so far. - Support rust-analyzer for out-of-tree modules too. - Give 'cfg's to rust-analyzer for the 'core' and 'alloc' crates. - Drop 'scripts/is_rust_module.sh' since it is not needed anymore. Macros crate: - New 'paste!' proc macro. This macro is a more flexible version of 'concat_idents!': it allows the resulting identifier to be used to declare new items and it allows to transform the identifiers before concatenating them, e.g. let x_1 = 42; paste!(let [<x _2>] = [<x _1>];); assert!(x_1 == x_2); The macro is then used for several of the pinned-init API changes in this pull. Pinned-init API: - Make '#[pin_data]' compatible with conditional compilation of fields, allowing to write code like: #[pin_data] pub struct Foo { #[cfg(CONFIG_BAR)] a: Bar, #[cfg(not(CONFIG_BAR))] a: Baz, } - New '#[derive(Zeroable)]' proc macro for the 'Zeroable' trait, which allows 'unsafe' implementations for structs where every field implements the 'Zeroable' trait, e.g.: #[derive(Zeroable)] pub struct DriverData { id: i64, buf_ptr: *mut u8, len: usize, } - Add '..Zeroable::zeroed()' syntax to the 'pin_init!' macro for zeroing all other fields, e.g.: pin_init!(Buf { buf: [1; 64], ..Zeroable::zeroed() }); - New '{,pin_}init_array_from_fn()' functions to create array initializers given a generator function, e.g.: let b: Box<[usize; 1_000]> = Box::init::<Error>( init_array_from_fn(|i| i) ).unwrap(); assert_eq!(b.len(), 1_000); assert_eq!(b[123], 123); - New '{,pin_}chain' methods for '{,Pin}Init<T, E>' that allow to execute a closure on the value directly after initialization, e.g.: let foo = init!(Foo { buf <- init::zeroed() }).chain(|foo| { foo.setup(); Ok(()) }); - Support arbitrary paths in init macros, instead of just identifiers and generic types. - Implement the 'Zeroable' trait for the 'UnsafeCell<T>' and 'Opaque<T>' types. - Make initializer values inaccessible after initialization. - Make guards in the init macros hygienic. 'allocator' module: - Use 'krealloc_aligned()' in 'KernelAllocator::alloc' preventing misaligned allocations when the Rust 1.71.1 upgrade is applied later in this pull. The equivalent fix for the previous compiler version (where 'KernelAllocator' is not yet used) was merged into 6.5 already, which added the 'krealloc_aligned()' function used here. - Implement 'KernelAllocator::{realloc, alloc_zeroed}' for performance, using 'krealloc_aligned()' too, which forwards the call to the C API. 'types' module: - Make 'Opaque' be '!Unpin', removing the need to add a 'PhantomPinned' field to Rust structs that contain C structs which must not be moved. - Make 'Opaque' use 'UnsafeCell' as the outer type, rather than inner. Documentation: - Suggest obtaining the source code of the Rust's 'core' library using the tarball instead of the repository. MAINTAINERS: - Andreas and Alice, from Samsung and Google respectively, are joining as reviewers of the "RUST" entry. As well as a few other minor changes and cleanups. -----BEGIN PGP SIGNATURE----- iQIzBAABCgAdFiEEPjU5OPd5QIZ9jqqOGXyLc2htIW0FAmTnzOAACgkQGXyLc2ht IW0RFg/9FKGAn+JNvLUpB7OIXQZFyDVDpXkL14Dy8At0z609ZhkD36pFAxGua4OC BLHpyEQK5bUAQZ4pZ1aexmpFt37z+OPZBMmKoC7eUH2fm8Q277Gm54pno2AzIg3g if9lFhIowQTB8pG1YZRF6YMIdIp5JCmT0m8YuXMrr1XYtWIWnyU4twT/bmfk9UKU DgmuE1GmpHbWQgIf11eYWxbgfIuY9F/QyHzljW8P+Jgln7F4d8WDVJln8Yw0z/Bm w/4kvYv7AHOHQvzjCi971ANvnhsgjeKMSmt2RrcGefn+6t3pNsdZEUYGR9xdAxCz fvcje6nUoGjPr9J4F/JdZPmCb7jwSGpF01OvA//H8YjUwP3+msBwxVhRSH1FA1m3 SVKedXmAUMNAaqtqCNFZmUiNB5LbW4cldFSnNf4CVW9w9bXe2jIKqjjsPi8m57B1 H4zwr1WTtY2s2n2fdYOAtzmOaOJFXa7PIrGo3onj1mSgcyKOVeoMI5+NR/pwxgIR 9Z8633bhTfGVHRyC7p0XpakcZd0jbl0yq+bbvgH2sof+RNWYuoZQ92DJ05/g3zOK Mj54PNjAgY+Z+TqX/vjlEdWs4SoBcnL3cAy9RFKGRDUoGDPeqiW6qa7Y9oAFZHfk PX3oboI0VYn5F9BVGO4i+9cL/CNL4b6sb5FBvL+0EwUBhWTxeKE= =BAP+ -----END PGP SIGNATURE----- Merge tag 'rust-6.6' of https://github.com/Rust-for-Linux/linux Pull rust updates from Miguel Ojeda: "In terms of lines, most changes this time are on the pinned-init API and infrastructure. While we have a Rust version upgrade, and thus a bunch of changes from the vendored 'alloc' crate as usual, this time those do not account for many lines. Toolchain and infrastructure: - Upgrade to Rust 1.71.1. This is the second such upgrade, which is a smaller jump compared to the last time. This version allows us to remove the '__rust_*' allocator functions -- the compiler now generates them as expected, thus now our 'KernelAllocator' is used. It also introduces the 'offset_of!' macro in the standard library (as an unstable feature) which we will need soon. So far, we were using a declarative macro as a prerequisite in some not-yet-landed patch series, which did not support sub-fields (i.e. nested structs): #[repr(C)] struct S { a: u16, b: (u8, u8), } assert_eq!(offset_of!(S, b.1), 3); - Upgrade to bindgen 0.65.1. This is the first time we upgrade its version. Given it is a fairly big jump, it comes with a fair number of improvements/changes that affect us, such as a fix needed to support LLVM 16 as well as proper support for '__noreturn' C functions, which are now mapped to return the '!' type in Rust: void __noreturn f(void); // C pub fn f() -> !; // Rust - 'scripts/rust_is_available.sh' improvements and fixes. This series takes care of all the issues known so far and adds a few new checks to cover for even more cases, plus adds some more help texts. All this together will hopefully make problematic setups easier to identify and to be solved by users building the kernel. In addition, it adds a test suite which covers all branches of the shell script, as well as tests for the issues found so far. - Support rust-analyzer for out-of-tree modules too. - Give 'cfg's to rust-analyzer for the 'core' and 'alloc' crates. - Drop 'scripts/is_rust_module.sh' since it is not needed anymore. Macros crate: - New 'paste!' proc macro. This macro is a more flexible version of 'concat_idents!': it allows the resulting identifier to be used to declare new items and it allows to transform the identifiers before concatenating them, e.g. let x_1 = 42; paste!(let [<x _2>] = [<x _1>];); assert!(x_1 == x_2); The macro is then used for several of the pinned-init API changes in this pull. Pinned-init API: - Make '#[pin_data]' compatible with conditional compilation of fields, allowing to write code like: #[pin_data] pub struct Foo { #[cfg(CONFIG_BAR)] a: Bar, #[cfg(not(CONFIG_BAR))] a: Baz, } - New '#[derive(Zeroable)]' proc macro for the 'Zeroable' trait, which allows 'unsafe' implementations for structs where every field implements the 'Zeroable' trait, e.g.: #[derive(Zeroable)] pub struct DriverData { id: i64, buf_ptr: *mut u8, len: usize, } - Add '..Zeroable::zeroed()' syntax to the 'pin_init!' macro for zeroing all other fields, e.g.: pin_init!(Buf { buf: [1; 64], ..Zeroable::zeroed() }); - New '{,pin_}init_array_from_fn()' functions to create array initializers given a generator function, e.g.: let b: Box<[usize; 1_000]> = Box::init::<Error>( init_array_from_fn(|i| i) ).unwrap(); assert_eq!(b.len(), 1_000); assert_eq!(b[123], 123); - New '{,pin_}chain' methods for '{,Pin}Init<T, E>' that allow to execute a closure on the value directly after initialization, e.g.: let foo = init!(Foo { buf <- init::zeroed() }).chain(|foo| { foo.setup(); Ok(()) }); - Support arbitrary paths in init macros, instead of just identifiers and generic types. - Implement the 'Zeroable' trait for the 'UnsafeCell<T>' and 'Opaque<T>' types. - Make initializer values inaccessible after initialization. - Make guards in the init macros hygienic. 'allocator' module: - Use 'krealloc_aligned()' in 'KernelAllocator::alloc' preventing misaligned allocations when the Rust 1.71.1 upgrade is applied later in this pull. The equivalent fix for the previous compiler version (where 'KernelAllocator' is not yet used) was merged into 6.5 already, which added the 'krealloc_aligned()' function used here. - Implement 'KernelAllocator::{realloc, alloc_zeroed}' for performance, using 'krealloc_aligned()' too, which forwards the call to the C API. 'types' module: - Make 'Opaque' be '!Unpin', removing the need to add a 'PhantomPinned' field to Rust structs that contain C structs which must not be moved. - Make 'Opaque' use 'UnsafeCell' as the outer type, rather than inner. Documentation: - Suggest obtaining the source code of the Rust's 'core' library using the tarball instead of the repository. MAINTAINERS: - Andreas and Alice, from Samsung and Google respectively, are joining as reviewers of the "RUST" entry. As well as a few other minor changes and cleanups" * tag 'rust-6.6' of https://github.com/Rust-for-Linux/linux: (42 commits) rust: init: update expanded macro explanation rust: init: add `{pin_}chain` functions to `{Pin}Init<T, E>` rust: init: make `PinInit<T, E>` a supertrait of `Init<T, E>` rust: init: implement `Zeroable` for `UnsafeCell<T>` and `Opaque<T>` rust: init: add support for arbitrary paths in init macros rust: init: add functions to create array initializers rust: init: add `..Zeroable::zeroed()` syntax for zeroing all missing fields rust: init: make initializer values inaccessible after initializing rust: init: wrap type checking struct initializers in a closure rust: init: make guards in the init macros hygienic rust: add derive macro for `Zeroable` rust: init: make `#[pin_data]` compatible with conditional compilation of fields rust: init: consolidate init macros docs: rust: clarify what 'rustup override' does docs: rust: update instructions for obtaining 'core' source docs: rust: add command line to rust-analyzer section scripts: generate_rust_analyzer: provide `cfg`s for `core` and `alloc` rust: bindgen: upgrade to 0.65.1 rust: enable `no_mangle_with_rust_abi` Clippy lint rust: upgrade to Rust 1.71.1 ...
2023-08-29 15:19:46 +00:00
/// # use kernel::{macros::{Zeroable, pin_data}, pin_init};
/// # use core::{ptr::addr_of_mut, marker::PhantomPinned};
/// #[pin_data]
/// #[derive(Zeroable)]
/// struct Buf {
/// // `ptr` points into `buf`.
/// ptr: *mut u8,
/// buf: [u8; 64],
/// #[pin]
/// pin: PhantomPinned,
/// }
/// pin_init!(&this in Buf {
/// buf: [0; 64],
/// ptr: unsafe { addr_of_mut!((*this.as_ptr()).buf).cast() },
/// pin: PhantomPinned,
/// });
/// pin_init!(Buf {
/// buf: [1; 64],
/// ..Zeroable::zeroed()
/// });
/// ```
///
/// [`try_pin_init!`]: kernel::try_pin_init
/// [`NonNull<Self>`]: core::ptr::NonNull
// For a detailed example of how this macro works, see the module documentation of the hidden
// module `__internal` inside of `init/__internal.rs`.
#[macro_export]
macro_rules! pin_init {
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)?),
@fields($($fields)*),
@error(::core::convert::Infallible),
@data(PinData, use_data),
@has_data(HasPinData, __pin_data),
@construct_closure(pin_init_from_closure),
@munch_fields($($fields)*),
)
};
}
/// Construct an in-place, fallible pinned initializer for `struct`s.
///
/// If the initialization can complete without error (or [`Infallible`]), then use [`pin_init!`].
///
/// You can use the `?` operator or use `return Err(err)` inside the initializer to stop
/// initialization and return the error.
///
/// IMPORTANT: if you have `unsafe` code inside of the initializer you have to ensure that when
/// initialization fails, the memory can be safely deallocated without any further modifications.
///
/// This macro defaults the error to [`Error`].
///
/// The syntax is identical to [`pin_init!`] with the following exception: you can append `? $type`
/// after the `struct` initializer to specify the error type you want to use.
///
/// # Examples
///
/// ```rust
/// # #![feature(new_uninit)]
/// use kernel::{init::{self, PinInit}, error::Error};
/// #[pin_data]
/// struct BigBuf {
/// big: Box<[u8; 1024 * 1024 * 1024]>,
/// small: [u8; 1024 * 1024],
/// ptr: *mut u8,
/// }
///
/// impl BigBuf {
/// fn new() -> impl PinInit<Self, Error> {
/// try_pin_init!(Self {
/// big: Box::init(init::zeroed())?,
/// small: [0; 1024 * 1024],
/// ptr: core::ptr::null_mut(),
/// }? Error)
/// }
/// }
/// ```
// For a detailed example of how this macro works, see the module documentation of the hidden
// module `__internal` inside of `init/__internal.rs`.
#[macro_export]
macro_rules! try_pin_init {
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)? ),
@fields($($fields)*),
@error($crate::error::Error),
@data(PinData, use_data),
@has_data(HasPinData, __pin_data),
@construct_closure(pin_init_from_closure),
@munch_fields($($fields)*),
)
};
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}? $err:ty) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)? ),
@fields($($fields)*),
@error($err),
@data(PinData, use_data),
@has_data(HasPinData, __pin_data),
@construct_closure(pin_init_from_closure),
@munch_fields($($fields)*),
)
};
}
/// Construct an in-place initializer for `struct`s.
///
/// This macro defaults the error to [`Infallible`]. If you need [`Error`], then use
/// [`try_init!`].
///
/// The syntax is identical to [`pin_init!`] and its safety caveats also apply:
/// - `unsafe` code must guarantee either full initialization or return an error and allow
/// deallocation of the memory.
/// - the fields are initialized in the order given in the initializer.
/// - no references to fields are allowed to be created inside of the initializer.
///
/// This initializer is for initializing data in-place that might later be moved. If you want to
/// pin-initialize, use [`pin_init!`].
rust: upgrade to Rust 1.68.2 This is the first upgrade to the Rust toolchain since the initial Rust merge, from 1.62.0 to 1.68.2 (i.e. the latest). # Context The kernel currently supports only a single Rust version [1] (rather than a minimum) given our usage of some "unstable" Rust features [2] which do not promise backwards compatibility. The goal is to reach a point where we can declare a minimum version for the toolchain. For instance, by waiting for some of the features to be stabilized. Therefore, the first minimum Rust version that the kernel will support is "in the future". # Upgrade policy Given we will eventually need to reach that minimum version, it would be ideal to upgrade the compiler from time to time to be as close as possible to that goal and find any issues sooner. In the extreme, we could upgrade as soon as a new Rust release is out. Of course, upgrading so often is in stark contrast to what one normally would need for GCC and LLVM, especially given the release schedule: 6 weeks for Rust vs. half a year for LLVM and a year for GCC. Having said that, there is no particular advantage to updating slowly either: kernel developers in "stable" distributions are unlikely to be able to use their distribution-provided Rust toolchain for the kernel anyway [3]. Instead, by routinely upgrading to the latest instead, kernel developers using Linux distributions that track the latest Rust release may be able to use those rather than Rust-provided ones, especially if their package manager allows to pin / hold back / downgrade the version for some days during windows where the version may not match. For instance, Arch, Fedora, Gentoo and openSUSE all provide and track the latest version of Rust as they get released every 6 weeks. Then, when the minimum version is reached, we will stop upgrading and decide how wide the window of support will be. For instance, a year of Rust versions. We will probably want to start small, and then widen it over time, just like the kernel did originally for LLVM, see commit 3519c4d6e08e ("Documentation: add minimum clang/llvm version"). # Unstable features stabilized This upgrade allows us to remove the following unstable features since they were stabilized: - `feature(explicit_generic_args_with_impl_trait)` (1.63). - `feature(core_ffi_c)` (1.64). - `feature(generic_associated_types)` (1.65). - `feature(const_ptr_offset_from)` (1.65, *). - `feature(bench_black_box)` (1.66, *). - `feature(pin_macro)` (1.68). The ones marked with `*` apply only to our old `rust` branch, not mainline yet, i.e. only for code that we may potentially upstream. With this patch applied, the only unstable feature allowed to be used outside the `kernel` crate is `new_uninit`, though other code to be upstreamed may increase the list. Please see [2] for details. # Other required changes Since 1.63, `rustdoc` triggers the `broken_intra_doc_links` lint for links pointing to exported (`#[macro_export]`) `macro_rules`. An issue was opened upstream [4], but it turns out it is intended behavior. For the moment, just add an explicit reference for each link. Later we can revisit this if `rustdoc` removes the compatibility measure. Nevertheless, this was helpful to discover a link that was pointing to the wrong place unintentionally. Since that one was actually wrong, it is fixed in a previous commit independently. Another change was the addition of `cfg(no_rc)` and `cfg(no_sync)` in upstream [5], thus remove our original changes for that. Similarly, upstream now tests that it compiles successfully with `#[cfg(not(no_global_oom_handling))]` [6], which allow us to get rid of some changes, such as an `#[allow(dead_code)]`. In addition, remove another `#[allow(dead_code)]` due to new uses within the standard library. Finally, add `try_extend_trusted` and move the code in `spec_extend.rs` since upstream moved it for the infallible version. # `alloc` upgrade and reviewing There are a large amount of changes, but the vast majority of them are due to our `alloc` fork being upgraded at once. There are two kinds of changes to be aware of: the ones coming from upstream, which we should follow as closely as possible, and the updates needed in our added fallible APIs to keep them matching the newer infallible APIs coming from upstream. Instead of taking a look at the diff of this patch, an alternative approach is reviewing a diff of the changes between upstream `alloc` and the kernel's. This allows to easily inspect the kernel additions only, especially to check if the fallible methods we already have still match the infallible ones in the new version coming from upstream. Another approach is reviewing the changes introduced in the additions in the kernel fork between the two versions. This is useful to spot potentially unintended changes to our additions. To apply these approaches, one may follow steps similar to the following to generate a pair of patches that show the differences between upstream Rust and the kernel (for the subset of `alloc` we use) before and after applying this patch: # Get the difference with respect to the old version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > old.patch git -C linux restore rust/alloc # Apply this patch. git -C linux am rust-upgrade.patch # Get the difference with respect to the new version. git -C rust checkout $(linux/scripts/min-tool-version.sh rustc) git -C linux ls-tree -r --name-only HEAD -- rust/alloc | cut -d/ -f3- | grep -Fv README.md | xargs -IPATH cp rust/library/alloc/src/PATH linux/rust/alloc/PATH git -C linux diff --patch-with-stat --summary -R > new.patch git -C linux restore rust/alloc Now one may check the `new.patch` to take a look at the additions (first approach) or at the difference between those two patches (second approach). For the latter, a side-by-side tool is recommended. Link: https://rust-for-linux.com/rust-version-policy [1] Link: https://github.com/Rust-for-Linux/linux/issues/2 [2] Link: https://lore.kernel.org/rust-for-linux/CANiq72mT3bVDKdHgaea-6WiZazd8Mvurqmqegbe5JZxVyLR8Yg@mail.gmail.com/ [3] Link: https://github.com/rust-lang/rust/issues/106142 [4] Link: https://github.com/rust-lang/rust/pull/89891 [5] Link: https://github.com/rust-lang/rust/pull/98652 [6] Reviewed-by: Björn Roy Baron <bjorn3_gh@protonmail.com> Reviewed-by: Gary Guo <gary@garyguo.net> Reviewed-By: Martin Rodriguez Reboredo <yakoyoku@gmail.com> Tested-by: Ariel Miculas <amiculas@cisco.com> Tested-by: David Gow <davidgow@google.com> Tested-by: Boqun Feng <boqun.feng@gmail.com> Link: https://lore.kernel.org/r/20230418214347.324156-4-ojeda@kernel.org [ Removed `feature(core_ffi_c)` from `uapi` ] Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-18 21:43:47 +00:00
///
/// [`try_init!`]: crate::try_init!
// For a detailed example of how this macro works, see the module documentation of the hidden
// module `__internal` inside of `init/__internal.rs`.
#[macro_export]
macro_rules! init {
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)?),
@fields($($fields)*),
@error(::core::convert::Infallible),
@data(InitData, /*no use_data*/),
@has_data(HasInitData, __init_data),
@construct_closure(init_from_closure),
@munch_fields($($fields)*),
)
}
}
/// Construct an in-place fallible initializer for `struct`s.
///
/// This macro defaults the error to [`Error`]. If you need [`Infallible`], then use
/// [`init!`].
///
/// The syntax is identical to [`try_pin_init!`]. If you want to specify a custom error,
/// append `? $type` after the `struct` initializer.
/// The safety caveats from [`try_pin_init!`] also apply:
/// - `unsafe` code must guarantee either full initialization or return an error and allow
/// deallocation of the memory.
/// - the fields are initialized in the order given in the initializer.
/// - no references to fields are allowed to be created inside of the initializer.
///
/// # Examples
///
/// ```rust
/// use kernel::{init::{PinInit, zeroed}, error::Error};
/// struct BigBuf {
/// big: Box<[u8; 1024 * 1024 * 1024]>,
/// small: [u8; 1024 * 1024],
/// }
///
/// impl BigBuf {
/// fn new() -> impl Init<Self, Error> {
/// try_init!(Self {
/// big: Box::init(zeroed())?,
/// small: [0; 1024 * 1024],
/// }? Error)
/// }
/// }
/// ```
// For a detailed example of how this macro works, see the module documentation of the hidden
// module `__internal` inside of `init/__internal.rs`.
#[macro_export]
macro_rules! try_init {
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)?),
@fields($($fields)*),
@error($crate::error::Error),
@data(InitData, /*no use_data*/),
@has_data(HasInitData, __init_data),
@construct_closure(init_from_closure),
@munch_fields($($fields)*),
)
};
($(&$this:ident in)? $t:ident $(::<$($generics:ty),* $(,)?>)? {
$($fields:tt)*
}? $err:ty) => {
$crate::__init_internal!(
@this($($this)?),
@typ($t $(::<$($generics),*>)?),
@fields($($fields)*),
@error($err),
@data(InitData, /*no use_data*/),
@has_data(HasInitData, __init_data),
@construct_closure(init_from_closure),
@munch_fields($($fields)*),
)
};
}
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// A pin-initializer for the type `T`.
///
/// To use this initializer, you will need a suitable memory location that can hold a `T`. This can
/// be [`Box<T>`], [`Arc<T>`], [`UniqueArc<T>`] or even the stack (see [`stack_pin_init!`]). Use the
/// [`InPlaceInit::pin_init`] function of a smart pointer like [`Arc<T>`] on this.
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
///
/// Also see the [module description](self).
///
/// # Safety
///
/// When implementing this trait you will need to take great care. Also there are probably very few
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// cases where a manual implementation is necessary. Use [`pin_init_from_closure`] where possible.
///
/// The [`PinInit::__pinned_init`] function:
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// - returns `Ok(())` if it initialized every field of `slot`,
/// - returns `Err(err)` if it encountered an error and then cleaned `slot`, this means:
/// - `slot` can be deallocated without UB occurring,
/// - `slot` does not need to be dropped,
/// - `slot` is not partially initialized.
/// - while constructing the `T` at `slot` it upholds the pinning invariants of `T`.
///
/// [`Arc<T>`]: crate::sync::Arc
/// [`Arc::pin_init`]: crate::sync::Arc::pin_init
#[must_use = "An initializer must be used in order to create its value."]
pub unsafe trait PinInit<T: ?Sized, E = Infallible>: Sized {
/// Initializes `slot`.
///
/// # Safety
///
/// - `slot` is a valid pointer to uninitialized memory.
/// - the caller does not touch `slot` when `Err` is returned, they are only permitted to
/// deallocate.
/// - `slot` will not move until it is dropped, i.e. it will be pinned.
unsafe fn __pinned_init(self, slot: *mut T) -> Result<(), E>;
/// First initializes the value using `self` then calls the function `f` with the initialized
/// value.
///
/// If `f` returns an error the value is dropped and the initializer will forward the error.
///
/// # Examples
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// use kernel::{types::Opaque, init::pin_init_from_closure};
/// #[repr(C)]
/// struct RawFoo([u8; 16]);
/// extern {
/// fn init_foo(_: *mut RawFoo);
/// }
///
/// #[pin_data]
/// struct Foo {
/// #[pin]
/// raw: Opaque<RawFoo>,
/// }
///
/// impl Foo {
/// fn setup(self: Pin<&mut Self>) {
/// pr_info!("Setting up foo");
/// }
/// }
///
/// let foo = pin_init!(Foo {
/// raw <- unsafe {
/// Opaque::ffi_init(|s| {
/// init_foo(s);
/// })
/// },
/// }).pin_chain(|foo| {
/// foo.setup();
/// Ok(())
/// });
/// ```
fn pin_chain<F>(self, f: F) -> ChainPinInit<Self, F, T, E>
where
F: FnOnce(Pin<&mut T>) -> Result<(), E>,
{
ChainPinInit(self, f, PhantomData)
}
}
/// An initializer returned by [`PinInit::pin_chain`].
pub struct ChainPinInit<I, F, T: ?Sized, E>(I, F, __internal::Invariant<(E, Box<T>)>);
// SAFETY: The `__pinned_init` function is implemented such that it
// - returns `Ok(())` on successful initialization,
// - returns `Err(err)` on error and in this case `slot` will be dropped.
// - considers `slot` pinned.
unsafe impl<T: ?Sized, E, I, F> PinInit<T, E> for ChainPinInit<I, F, T, E>
where
I: PinInit<T, E>,
F: FnOnce(Pin<&mut T>) -> Result<(), E>,
{
unsafe fn __pinned_init(self, slot: *mut T) -> Result<(), E> {
// SAFETY: All requirements fulfilled since this function is `__pinned_init`.
unsafe { self.0.__pinned_init(slot)? };
// SAFETY: The above call initialized `slot` and we still have unique access.
let val = unsafe { &mut *slot };
// SAFETY: `slot` is considered pinned.
let val = unsafe { Pin::new_unchecked(val) };
(self.1)(val).map_err(|e| {
// SAFETY: `slot` was initialized above.
unsafe { core::ptr::drop_in_place(slot) };
e
})
}
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
}
/// An initializer for `T`.
///
/// To use this initializer, you will need a suitable memory location that can hold a `T`. This can
/// be [`Box<T>`], [`Arc<T>`], [`UniqueArc<T>`] or even the stack (see [`stack_pin_init!`]). Use the
/// [`InPlaceInit::init`] function of a smart pointer like [`Arc<T>`] on this. Because
/// [`PinInit<T, E>`] is a super trait, you can use every function that takes it as well.
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
///
/// Also see the [module description](self).
///
/// # Safety
///
/// When implementing this trait you will need to take great care. Also there are probably very few
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// cases where a manual implementation is necessary. Use [`init_from_closure`] where possible.
///
/// The [`Init::__init`] function:
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// - returns `Ok(())` if it initialized every field of `slot`,
/// - returns `Err(err)` if it encountered an error and then cleaned `slot`, this means:
/// - `slot` can be deallocated without UB occurring,
/// - `slot` does not need to be dropped,
/// - `slot` is not partially initialized.
/// - while constructing the `T` at `slot` it upholds the pinning invariants of `T`.
///
/// The `__pinned_init` function from the supertrait [`PinInit`] needs to execute the exact same
/// code as `__init`.
///
/// Contrary to its supertype [`PinInit<T, E>`] the caller is allowed to
/// move the pointee after initialization.
///
/// [`Arc<T>`]: crate::sync::Arc
#[must_use = "An initializer must be used in order to create its value."]
pub unsafe trait Init<T: ?Sized, E = Infallible>: PinInit<T, E> {
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
/// Initializes `slot`.
///
/// # Safety
///
/// - `slot` is a valid pointer to uninitialized memory.
/// - the caller does not touch `slot` when `Err` is returned, they are only permitted to
/// deallocate.
unsafe fn __init(self, slot: *mut T) -> Result<(), E>;
/// First initializes the value using `self` then calls the function `f` with the initialized
/// value.
///
/// If `f` returns an error the value is dropped and the initializer will forward the error.
///
/// # Examples
///
/// ```rust
/// # #![allow(clippy::disallowed_names)]
/// use kernel::{types::Opaque, init::{self, init_from_closure}};
/// struct Foo {
/// buf: [u8; 1_000_000],
/// }
///
/// impl Foo {
/// fn setup(&mut self) {
/// pr_info!("Setting up foo");
/// }
/// }
///
/// let foo = init!(Foo {
/// buf <- init::zeroed()
/// }).chain(|foo| {
/// foo.setup();
/// Ok(())
/// });
/// ```
fn chain<F>(self, f: F) -> ChainInit<Self, F, T, E>
where
F: FnOnce(&mut T) -> Result<(), E>,
{
ChainInit(self, f, PhantomData)
}
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
}
/// An initializer returned by [`Init::chain`].
pub struct ChainInit<I, F, T: ?Sized, E>(I, F, __internal::Invariant<(E, Box<T>)>);
// SAFETY: The `__init` function is implemented such that it
// - returns `Ok(())` on successful initialization,
// - returns `Err(err)` on error and in this case `slot` will be dropped.
unsafe impl<T: ?Sized, E, I, F> Init<T, E> for ChainInit<I, F, T, E>
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
where
I: Init<T, E>,
F: FnOnce(&mut T) -> Result<(), E>,
{
unsafe fn __init(self, slot: *mut T) -> Result<(), E> {
// SAFETY: All requirements fulfilled since this function is `__init`.
unsafe { self.0.__pinned_init(slot)? };
// SAFETY: The above call initialized `slot` and we still have unique access.
(self.1)(unsafe { &mut *slot }).map_err(|e| {
// SAFETY: `slot` was initialized above.
unsafe { core::ptr::drop_in_place(slot) };
e
})
}
}
// SAFETY: `__pinned_init` behaves exactly the same as `__init`.
unsafe impl<T: ?Sized, E, I, F> PinInit<T, E> for ChainInit<I, F, T, E>
where
I: Init<T, E>,
F: FnOnce(&mut T) -> Result<(), E>,
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
{
unsafe fn __pinned_init(self, slot: *mut T) -> Result<(), E> {
// SAFETY: `__init` has less strict requirements compared to `__pinned_init`.
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
unsafe { self.__init(slot) }
}
}
/// Creates a new [`PinInit<T, E>`] from the given closure.
///
/// # Safety
///
/// The closure:
/// - returns `Ok(())` if it initialized every field of `slot`,
/// - returns `Err(err)` if it encountered an error and then cleaned `slot`, this means:
/// - `slot` can be deallocated without UB occurring,
/// - `slot` does not need to be dropped,
/// - `slot` is not partially initialized.
/// - may assume that the `slot` does not move if `T: !Unpin`,
/// - while constructing the `T` at `slot` it upholds the pinning invariants of `T`.
#[inline]
pub const unsafe fn pin_init_from_closure<T: ?Sized, E>(
f: impl FnOnce(*mut T) -> Result<(), E>,
) -> impl PinInit<T, E> {
__internal::InitClosure(f, PhantomData)
}
/// Creates a new [`Init<T, E>`] from the given closure.
///
/// # Safety
///
/// The closure:
/// - returns `Ok(())` if it initialized every field of `slot`,
/// - returns `Err(err)` if it encountered an error and then cleaned `slot`, this means:
/// - `slot` can be deallocated without UB occurring,
/// - `slot` does not need to be dropped,
/// - `slot` is not partially initialized.
/// - the `slot` may move after initialization.
/// - while constructing the `T` at `slot` it upholds the pinning invariants of `T`.
#[inline]
pub const unsafe fn init_from_closure<T: ?Sized, E>(
f: impl FnOnce(*mut T) -> Result<(), E>,
) -> impl Init<T, E> {
__internal::InitClosure(f, PhantomData)
}
/// An initializer that leaves the memory uninitialized.
///
/// The initializer is a no-op. The `slot` memory is not changed.
#[inline]
pub fn uninit<T, E>() -> impl Init<MaybeUninit<T>, E> {
// SAFETY: The memory is allowed to be uninitialized.
unsafe { init_from_closure(|_| Ok(())) }
}
/// Initializes an array by initializing each element via the provided initializer.
///
/// # Examples
///
/// ```rust
/// use kernel::{error::Error, init::init_array_from_fn};
/// let array: Box<[usize; 1_000]> = Box::init::<Error>(init_array_from_fn(|i| i)).unwrap();
/// assert_eq!(array.len(), 1_000);
/// ```
pub fn init_array_from_fn<I, const N: usize, T, E>(
mut make_init: impl FnMut(usize) -> I,
) -> impl Init<[T; N], E>
where
I: Init<T, E>,
{
let init = move |slot: *mut [T; N]| {
let slot = slot.cast::<T>();
// Counts the number of initialized elements and when dropped drops that many elements from
// `slot`.
let mut init_count = ScopeGuard::new_with_data(0, |i| {
// We now free every element that has been initialized before.
// SAFETY: The loop initialized exactly the values from 0..i and since we
// return `Err` below, the caller will consider the memory at `slot` as
// uninitialized.
unsafe { ptr::drop_in_place(ptr::slice_from_raw_parts_mut(slot, i)) };
});
for i in 0..N {
let init = make_init(i);
// SAFETY: Since 0 <= `i` < N, it is still in bounds of `[T; N]`.
let ptr = unsafe { slot.add(i) };
// SAFETY: The pointer is derived from `slot` and thus satisfies the `__init`
// requirements.
unsafe { init.__init(ptr) }?;
*init_count += 1;
}
init_count.dismiss();
Ok(())
};
// SAFETY: The initializer above initializes every element of the array. On failure it drops
// any initialized elements and returns `Err`.
unsafe { init_from_closure(init) }
}
/// Initializes an array by initializing each element via the provided initializer.
///
/// # Examples
///
/// ```rust
/// use kernel::{sync::{Arc, Mutex}, init::pin_init_array_from_fn, new_mutex};
/// let array: Arc<[Mutex<usize>; 1_000]> =
/// Arc::pin_init(pin_init_array_from_fn(|i| new_mutex!(i))).unwrap();
/// assert_eq!(array.len(), 1_000);
/// ```
pub fn pin_init_array_from_fn<I, const N: usize, T, E>(
mut make_init: impl FnMut(usize) -> I,
) -> impl PinInit<[T; N], E>
where
I: PinInit<T, E>,
{
let init = move |slot: *mut [T; N]| {
let slot = slot.cast::<T>();
// Counts the number of initialized elements and when dropped drops that many elements from
// `slot`.
let mut init_count = ScopeGuard::new_with_data(0, |i| {
// We now free every element that has been initialized before.
// SAFETY: The loop initialized exactly the values from 0..i and since we
// return `Err` below, the caller will consider the memory at `slot` as
// uninitialized.
unsafe { ptr::drop_in_place(ptr::slice_from_raw_parts_mut(slot, i)) };
});
for i in 0..N {
let init = make_init(i);
// SAFETY: Since 0 <= `i` < N, it is still in bounds of `[T; N]`.
let ptr = unsafe { slot.add(i) };
// SAFETY: The pointer is derived from `slot` and thus satisfies the `__init`
// requirements.
unsafe { init.__pinned_init(ptr) }?;
*init_count += 1;
}
init_count.dismiss();
Ok(())
};
// SAFETY: The initializer above initializes every element of the array. On failure it drops
// any initialized elements and returns `Err`.
unsafe { pin_init_from_closure(init) }
}
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
// SAFETY: Every type can be initialized by-value.
unsafe impl<T, E> Init<T, E> for T {
unsafe fn __init(self, slot: *mut T) -> Result<(), E> {
rust: add pin-init API core This API is used to facilitate safe pinned initialization of structs. It replaces cumbersome `unsafe` manual initialization with elegant safe macro invocations. Due to the size of this change it has been split into six commits: 1. This commit introducing the basic public interface: traits and functions to represent and create initializers. 2. Adds the `#[pin_data]`, `pin_init!`, `try_pin_init!`, `init!` and `try_init!` macros along with their internal types. 3. Adds the `InPlaceInit` trait that allows using an initializer to create an object inside of a `Box<T>` and other smart pointers. 4. Adds the `PinnedDrop` trait and adds macro support for it in the `#[pin_data]` macro. 5. Adds the `stack_pin_init!` macro allowing to pin-initialize a struct on the stack. 6. Adds the `Zeroable` trait and `init::zeroed` function to initialize types that have `0x00` in all bytes as a valid bit pattern. -- In this section the problem that the new pin-init API solves is outlined. This message describes the entirety of the API, not just the parts introduced in this commit. For a more granular explanation and additional information on pinning and this issue, view [1]. Pinning is Rust's way of enforcing the address stability of a value. When a value gets pinned it will be impossible for safe code to move it to another location. This is done by wrapping pointers to said object with `Pin<P>`. This wrapper prevents safe code from creating mutable references to the object, preventing mutable access, which is needed to move the value. `Pin<P>` provides `unsafe` functions to circumvent this and allow modifications regardless. It is then the programmer's responsibility to uphold the pinning guarantee. Many kernel data structures require a stable address, because there are foreign pointers to them which would get invalidated by moving the structure. Since these data structures are usually embedded in structs to use them, this pinning property propagates to the container struct. Resulting in most structs in both Rust and C code needing to be pinned. So if we want to have a `mutex` field in a Rust struct, this struct also needs to be pinned, because a `mutex` contains a `list_head`. Additionally initializing a `list_head` requires already having the final memory location available, because it is initialized by pointing it to itself. But this presents another challenge in Rust: values have to be initialized at all times. There is the `MaybeUninit<T>` wrapper type, which allows handling uninitialized memory, but this requires using the `unsafe` raw pointers and a casting the type to the initialized variant. This problem gets exacerbated when considering encapsulation and the normal safety requirements of Rust code. The fields of the Rust `Mutex<T>` should not be accessible to normal driver code. After all if anyone can modify the fields, there is no way to ensure the invariants of the `Mutex<T>` are upheld. But if the fields are inaccessible, then initialization of a `Mutex<T>` needs to be somehow achieved via a function or a macro. Because the `Mutex<T>` must be pinned in memory, the function cannot return it by value. It also cannot allocate a `Box` to put the `Mutex<T>` into, because that is an unnecessary allocation and indirection which would hurt performance. The solution in the rust tree (e.g. this commit: [2]) that is replaced by this API is to split this function into two parts: 1. A `new` function that returns a partially initialized `Mutex<T>`, 2. An `init` function that requires the `Mutex<T>` to be pinned and that fully initializes the `Mutex<T>`. Both of these functions have to be marked `unsafe`, since a call to `new` needs to be accompanied with a call to `init`, otherwise using the `Mutex<T>` could result in UB. And because calling `init` twice also is not safe. While `Mutex<T>` initialization cannot fail, other structs might also have to allocate memory, which would result in conditional successful initialization requiring even more manual accommodation work. Combine this with the problem of pin-projections -- the way of accessing fields of a pinned struct -- which also have an `unsafe` API, pinned initialization is riddled with `unsafe` resulting in very poor ergonomics. Not only that, but also having to call two functions possibly multiple lines apart makes it very easy to forget it outright or during refactoring. Here is an example of the current way of initializing a struct with two synchronization primitives (see [3] for the full example): struct SharedState { state_changed: CondVar, inner: Mutex<SharedStateInner>, } impl SharedState { fn try_new() -> Result<Arc<Self>> { let mut state = Pin::from(UniqueArc::try_new(Self { // SAFETY: `condvar_init!` is called below. state_changed: unsafe { CondVar::new() }, // SAFETY: `mutex_init!` is called below. inner: unsafe { Mutex::new(SharedStateInner { token_count: 0 }) }, })?); // SAFETY: `state_changed` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.state_changed) }; kernel::condvar_init!(pinned, "SharedState::state_changed"); // SAFETY: `inner` is pinned when `state` is. let pinned = unsafe { state.as_mut().map_unchecked_mut(|s| &mut s.inner) }; kernel::mutex_init!(pinned, "SharedState::inner"); Ok(state.into()) } } The pin-init API of this patch solves this issue by providing a comprehensive solution comprised of macros and traits. Here is the example from above using the pin-init API: #[pin_data] struct SharedState { #[pin] state_changed: CondVar, #[pin] inner: Mutex<SharedStateInner>, } impl SharedState { fn new() -> impl PinInit<Self> { pin_init!(Self { state_changed <- new_condvar!("SharedState::state_changed"), inner <- new_mutex!( SharedStateInner { token_count: 0 }, "SharedState::inner", ), }) } } Notably the way the macro is used here requires no `unsafe` and thus comes with the usual Rust promise of safe code not introducing any memory violations. Additionally it is now up to the caller of `new()` to decide the memory location of the `SharedState`. They can choose at the moment `Arc<T>`, `Box<T>` or the stack. -- The API has the following architecture: 1. Initializer traits `PinInit<T, E>` and `Init<T, E>` that act like closures. 2. Macros to create these initializer traits safely. 3. Functions to allow manually writing initializers. The initializers (an `impl PinInit<T, E>`) receive a raw pointer pointing to uninitialized memory and their job is to fully initialize a `T` at that location. If initialization fails, they return an error (`E`) by value. This way of initializing cannot be safely exposed to the user, since it relies upon these properties outside of the control of the trait: - the memory location (slot) needs to be valid memory, - if initialization fails, the slot should not be read from, - the value in the slot should be pinned, so it cannot move and the memory cannot be deallocated until the value is dropped. This is why using an initializer is facilitated by another trait that ensures these requirements. These initializers can be created manually by just supplying a closure that fulfills the same safety requirements as `PinInit<T, E>`. But this is an `unsafe` operation. To allow safe initializer creation, the `pin_init!` is provided along with three other variants: `try_pin_init!`, `try_init!` and `init!`. These take a modified struct initializer as a parameter and generate a closure that initializes the fields in sequence. The macros take great care in upholding the safety requirements: - A shadowed struct type is used as the return type of the closure instead of `()`. This is to prevent early returns, as these would prevent full initialization. - To ensure every field is only initialized once, a normal struct initializer is placed in unreachable code. The type checker will emit errors if a field is missing or specified multiple times. - When initializing a field fails, the whole initializer will fail and automatically drop fields that have been initialized earlier. - Only the correct initializer type is allowed for unpinned fields. You cannot use a `impl PinInit<T, E>` to initialize a structurally not pinned field. To ensure the last point, an additional macro `#[pin_data]` is needed. This macro annotates the struct itself and the user specifies structurally pinned and not pinned fields. Because dropping a pinned struct is also not allowed to break the pinning invariants, another macro attribute `#[pinned_drop]` is needed. This macro is introduced in a following commit. These two macros also have mechanisms to ensure the overall safety of the API. Additionally, they utilize a combined proc-macro, declarative macro design: first a proc-macro enables the outer attribute syntax `#[...]` and does some important pre-parsing. Notably this prepares the generics such that the declarative macro can handle them using token trees. Then the actual parsing of the structure and the emission of code is handled by a declarative macro. For pin-projections the crates `pin-project` [4] and `pin-project-lite` [5] had been considered, but were ultimately rejected: - `pin-project` depends on `syn` [6] which is a very big dependency, around 50k lines of code. - `pin-project-lite` is a more reasonable 5k lines of code, but contains a very complex declarative macro to parse generics. On top of that it would require modification that would need to be maintained independently. Link: https://rust-for-linux.com/the-safe-pinned-initialization-problem [1] Link: https://github.com/Rust-for-Linux/linux/tree/0a04dc4ddd671efb87eef54dde0fb38e9074f4be [2] Link: https://github.com/Rust-for-Linux/linux/blob/f509ede33fc10a07eba3da14aa00302bd4b5dddd/samples/rust/rust_miscdev.rs [3] Link: https://crates.io/crates/pin-project [4] Link: https://crates.io/crates/pin-project-lite [5] Link: https://crates.io/crates/syn [6] Co-developed-by: Gary Guo <gary@garyguo.net> Signed-off-by: Gary Guo <gary@garyguo.net> Signed-off-by: Benno Lossin <benno.lossin@proton.me> Reviewed-by: Alice Ryhl <aliceryhl@google.com> Reviewed-by: Wedson Almeida Filho <wedsonaf@gmail.com> Reviewed-by: Andreas Hindborg <a.hindborg@samsung.com> Link: https://lore.kernel.org/r/20230408122429.1103522-7-y86-dev@protonmail.com Signed-off-by: Miguel Ojeda <ojeda@kernel.org>
2023-04-08 12:25:45 +00:00
unsafe { slot.write(self) };
Ok(())
}
}
// SAFETY: Every type can be initialized by-value. `__pinned_init` calls `__init`.
unsafe impl<T, E> PinInit<T, E> for T {
unsafe fn __pinned_init(self, slot: *mut T) -> Result<(), E> {
unsafe { self.__init(slot) }
}
}
/// Smart pointer that can initialize memory in-place.
pub trait InPlaceInit<T>: Sized {
/// Use the given pin-initializer to pin-initialize a `T` inside of a new smart pointer of this
/// type.
///
/// If `T: !Unpin` it will not be able to move afterwards.
fn try_pin_init<E>(init: impl PinInit<T, E>) -> Result<Pin<Self>, E>
where
E: From<AllocError>;
/// Use the given pin-initializer to pin-initialize a `T` inside of a new smart pointer of this
/// type.
///
/// If `T: !Unpin` it will not be able to move afterwards.
fn pin_init<E>(init: impl PinInit<T, E>) -> error::Result<Pin<Self>>
where
Error: From<E>,
{
// SAFETY: We delegate to `init` and only change the error type.
let init = unsafe {
pin_init_from_closure(|slot| init.__pinned_init(slot).map_err(|e| Error::from(e)))
};
Self::try_pin_init(init)
}
/// Use the given initializer to in-place initialize a `T`.
fn try_init<E>(init: impl Init<T, E>) -> Result<Self, E>
where
E: From<AllocError>;
/// Use the given initializer to in-place initialize a `T`.
fn init<E>(init: impl Init<T, E>) -> error::Result<Self>
where
Error: From<E>,
{
// SAFETY: We delegate to `init` and only change the error type.
let init = unsafe {
init_from_closure(|slot| init.__pinned_init(slot).map_err(|e| Error::from(e)))
};
Self::try_init(init)
}
}
impl<T> InPlaceInit<T> for Box<T> {
#[inline]
fn try_pin_init<E>(init: impl PinInit<T, E>) -> Result<Pin<Self>, E>
where
E: From<AllocError>,
{
let mut this = Box::try_new_uninit()?;
let slot = this.as_mut_ptr();
// SAFETY: When init errors/panics, slot will get deallocated but not dropped,
// slot is valid and will not be moved, because we pin it later.
unsafe { init.__pinned_init(slot)? };
// SAFETY: All fields have been initialized.
Ok(unsafe { this.assume_init() }.into())
}
#[inline]
fn try_init<E>(init: impl Init<T, E>) -> Result<Self, E>
where
E: From<AllocError>,
{
let mut this = Box::try_new_uninit()?;
let slot = this.as_mut_ptr();
// SAFETY: When init errors/panics, slot will get deallocated but not dropped,
// slot is valid.
unsafe { init.__init(slot)? };
// SAFETY: All fields have been initialized.
Ok(unsafe { this.assume_init() })
}
}
impl<T> InPlaceInit<T> for UniqueArc<T> {
#[inline]
fn try_pin_init<E>(init: impl PinInit<T, E>) -> Result<Pin<Self>, E>
where
E: From<AllocError>,
{
let mut this = UniqueArc::try_new_uninit()?;
let slot = this.as_mut_ptr();
// SAFETY: When init errors/panics, slot will get deallocated but not dropped,
// slot is valid and will not be moved, because we pin it later.
unsafe { init.__pinned_init(slot)? };
// SAFETY: All fields have been initialized.
Ok(unsafe { this.assume_init() }.into())
}
#[inline]
fn try_init<E>(init: impl Init<T, E>) -> Result<Self, E>
where
E: From<AllocError>,
{
let mut this = UniqueArc::try_new_uninit()?;
let slot = this.as_mut_ptr();
// SAFETY: When init errors/panics, slot will get deallocated but not dropped,
// slot is valid.
unsafe { init.__init(slot)? };
// SAFETY: All fields have been initialized.
Ok(unsafe { this.assume_init() })
}
}
/// Trait facilitating pinned destruction.
///
/// Use [`pinned_drop`] to implement this trait safely:
///
/// ```rust
/// # use kernel::sync::Mutex;
/// use kernel::macros::pinned_drop;
/// use core::pin::Pin;
/// #[pin_data(PinnedDrop)]
/// struct Foo {
/// #[pin]
/// mtx: Mutex<usize>,
/// }
///
/// #[pinned_drop]
/// impl PinnedDrop for Foo {
/// fn drop(self: Pin<&mut Self>) {
/// pr_info!("Foo is being dropped!");
/// }
/// }
/// ```
///
/// # Safety
///
/// This trait must be implemented via the [`pinned_drop`] proc-macro attribute on the impl.
///
/// [`pinned_drop`]: kernel::macros::pinned_drop
pub unsafe trait PinnedDrop: __internal::HasPinData {
/// Executes the pinned destructor of this type.
///
/// While this function is marked safe, it is actually unsafe to call it manually. For this
/// reason it takes an additional parameter. This type can only be constructed by `unsafe` code
/// and thus prevents this function from being called where it should not.
///
/// This extra parameter will be generated by the `#[pinned_drop]` proc-macro attribute
/// automatically.
fn drop(self: Pin<&mut Self>, only_call_from_drop: __internal::OnlyCallFromDrop);
}
/// Marker trait for types that can be initialized by writing just zeroes.
///
/// # Safety
///
/// The bit pattern consisting of only zeroes is a valid bit pattern for this type. In other words,
/// this is not UB:
///
/// ```rust,ignore
/// let val: Self = unsafe { core::mem::zeroed() };
/// ```
pub unsafe trait Zeroable {}
/// Create a new zeroed T.
///
/// The returned initializer will write `0x00` to every byte of the given `slot`.
#[inline]
pub fn zeroed<T: Zeroable>() -> impl Init<T> {
// SAFETY: Because `T: Zeroable`, all bytes zero is a valid bit pattern for `T`
// and because we write all zeroes, the memory is initialized.
unsafe {
init_from_closure(|slot: *mut T| {
slot.write_bytes(0, 1);
Ok(())
})
}
}
macro_rules! impl_zeroable {
($($({$($generics:tt)*})? $t:ty, )*) => {
$(unsafe impl$($($generics)*)? Zeroable for $t {})*
};
}
impl_zeroable! {
// SAFETY: All primitives that are allowed to be zero.
bool,
char,
u8, u16, u32, u64, u128, usize,
i8, i16, i32, i64, i128, isize,
f32, f64,
// SAFETY: These are ZSTs, there is nothing to zero.
{<T: ?Sized>} PhantomData<T>, core::marker::PhantomPinned, Infallible, (),
// SAFETY: Type is allowed to take any value, including all zeros.
{<T>} MaybeUninit<T>,
// SAFETY: Type is allowed to take any value, including all zeros.
{<T>} Opaque<T>,
// SAFETY: `T: Zeroable` and `UnsafeCell` is `repr(transparent)`.
{<T: ?Sized + Zeroable>} UnsafeCell<T>,
// SAFETY: All zeros is equivalent to `None` (option layout optimization guarantee).
Option<NonZeroU8>, Option<NonZeroU16>, Option<NonZeroU32>, Option<NonZeroU64>,
Option<NonZeroU128>, Option<NonZeroUsize>,
Option<NonZeroI8>, Option<NonZeroI16>, Option<NonZeroI32>, Option<NonZeroI64>,
Option<NonZeroI128>, Option<NonZeroIsize>,
// SAFETY: All zeros is equivalent to `None` (option layout optimization guarantee).
//
// In this case we are allowed to use `T: ?Sized`, since all zeros is the `None` variant.
{<T: ?Sized>} Option<NonNull<T>>,
{<T: ?Sized>} Option<Box<T>>,
// SAFETY: `null` pointer is valid.
//
// We cannot use `T: ?Sized`, since the VTABLE pointer part of fat pointers is not allowed to be
// null.
//
// When `Pointee` gets stabilized, we could use
// `T: ?Sized where <T as Pointee>::Metadata: Zeroable`
{<T>} *mut T, {<T>} *const T,
// SAFETY: `null` pointer is valid and the metadata part of these fat pointers is allowed to be
// zero.
{<T>} *mut [T], {<T>} *const [T], *mut str, *const str,
// SAFETY: `T` is `Zeroable`.
{<const N: usize, T: Zeroable>} [T; N], {<T: Zeroable>} Wrapping<T>,
}
macro_rules! impl_tuple_zeroable {
($(,)?) => {};
($first:ident, $($t:ident),* $(,)?) => {
// SAFETY: All elements are zeroable and padding can be zero.
unsafe impl<$first: Zeroable, $($t: Zeroable),*> Zeroable for ($first, $($t),*) {}
impl_tuple_zeroable!($($t),* ,);
}
}
impl_tuple_zeroable!(A, B, C, D, E, F, G, H, I, J);