rust: Add Per-CPU Variable API

Per-CPU variables are an important tool for reducing lock contention, especially in systems with many processors. They also provide a convenient way to handle data that are logically associated with a particular CPU (e.g., the currently running task). Therefore, add a Rust API to make use of per-CPU variables.

Add a CpuGuard type that disables preemption for its lifetime. Add a PerCpuAllocation type used to track dynamic allocations. Add a define_per_cpu! macro to create static per-CPU allocations. Add DynamicPerCpu and StaticPerCpu to provide a high-level API. Add a PerCpu trait to unify the dynamic and static cases.

Co-developed-by Boqun Feng <[email protected]>

Signed-off-by Boqun Feng <[email protected]>

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/helpers/helpers.c          |   2 +
 rust/helpers/percpu.c           |   9 ++
 rust/helpers/preempt.c          |  14 +++
 rust/kernel/lib.rs              |   3 +
 rust/kernel/percpu.rs           | 223 ++++++++++++++++++++++++++++++++++++++++
 rust/kernel/percpu/cpu_guard.rs |  35 +++++++
 rust/kernel/percpu/dynamic.rs   |  83 +++++++++++++++
 rust/kernel/percpu/static_.rs   | 132 ++++++++++++++++++++++++
 8 files changed, 501 insertions(+)

rust/helpers/helpers.c +2 -0

index 7cf7fe95e41d..2fc8d26cfe66 100644
@@ -31,9 +31,11 @@
 #include "of.c"
 #include "page.c"
 #include "pci.c"
+#include "percpu.c"
 #include "pid_namespace.c"
 #include "platform.c"
 #include "poll.c"
+#include "preempt.c"
 #include "property.c"
 #include "rbtree.c"
 #include "rcu.c"

rust/helpers/percpu.c new file +9 -0

new file mode 100644
index 000000000000..a091389f730f
@@ -0,0 +1,9 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/percpu.h>
+
+void __percpu *rust_helper_alloc_percpu(size_t sz, size_t align)
+{
+	return __alloc_percpu(sz, align);
+}
+

rust/helpers/preempt.c new file +14 -0

new file mode 100644
index 000000000000..2c7529528ddd
@@ -0,0 +1,14 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <linux/preempt.h>
+
+void rust_helper_preempt_disable(void)
+{
+	preempt_disable();
+}
+
+void rust_helper_preempt_enable(void)
+{
+	preempt_enable();
+}
+

rust/kernel/lib.rs +3 -0

index ed53169e795c..ed0d5756dc55 100644
@@ -106,6 +106,9 @@
 pub mod page;
 #[cfg(CONFIG_PCI)]
 pub mod pci;
+// Only x86_64 is supported by percpu for now
+#[cfg(CONFIG_X86_64)]
+pub mod percpu;
 pub mod pid_namespace;
 pub mod platform;
 pub mod prelude;

rust/kernel/percpu.rs new file +54 -0

new file mode 100644
index 000000000000..35afcdba3ccd
@@ -0,0 +1,223 @@
+// SPDX-License-Identifier: GPL-2.0
+//! This module contains abstractions for creating and using per-CPU variables from Rust.
+//! See the define_per_cpu! macro and the DynamicPerCpu<T> type, as well as the PerCpu<T> trait.
+
+pub mod cpu_guard;
+mod dynamic;
+mod static_;
+
+#[doc(inline)]
+pub use dynamic::*;
+#[doc(inline)]
+pub use static_::*;
+
+use bindings::{alloc_percpu, free_percpu};
+
+use crate::alloc::Flags;
+use crate::percpu::cpu_guard::CpuGuard;
+use crate::prelude::*;
+use crate::sync::Arc;
+use crate::types::Opaque;
+use crate::{declare_extern_per_cpu, get_static_per_cpu};
+
+use core::arch::asm;
+use core::cell::{Cell, RefCell, UnsafeCell};
+use core::mem::{align_of, size_of, MaybeUninit};
+
+use ffi::c_void;
+
+/// A per-CPU pointer; that is, an offset into the per-CPU area. Note that this type is NOT a smart
+/// pointer, it does not manage the allocation.
+pub struct PerCpuPtr<T>(*mut MaybeUninit<T>);
+
+/// Represents exclusive access to the memory location pointed at by a particular PerCpu<T>.
+pub struct PerCpuToken<'a, T> {
+    // INVARIANT: the current CPU's memory location associated with the per-CPU variable pointed at
+    // by `ptr` (i.e., the entry in the per-CPU area on the current CPU) has been initialized.
+    _guard: CpuGuard,
+    ptr: &'a PerCpuPtr<T>,
+}
+
+/// Represents access to the memory location pointed at by a particular PerCpu<T> where the type
+/// `T` manages access to the underlying memory to avoid aliaising troubles. (For example, `T`
+/// might be a `Cell` or `RefCell`.)
+pub struct CheckedPerCpuToken<'a, T> {
+    // INVARIANT: the current CPU's memory location associated with the per-CPU variable pointed at
+    // by `ptr` (i.e., the entry in the per-CPU area on the current CPU) has been initialized.
+    _guard: CpuGuard,
+    ptr: &'a PerCpuPtr<T>,
+}
+
+impl<T> PerCpuPtr<T> {
+    /// Makes a new PerCpuPtr from a raw per-CPU pointer.
+    ///
+    /// # Safety

rust/kernel/percpu.rs new file +146 -0

+
+    /// Get a `&MaybeUninit<T>` to the per-CPU variable on the current CPU represented by `&self`
+    ///
+    /// # Safety
+    /// The returned `&T` must follow Rust's aliasing rules. That is, no `&mut T` may exist that
+    /// points to the same location in memory. In practice, this means that `get_mut_ref` must not
+    /// be called on another `PerCpuPtr<T>` that is a copy/clone of `&self` for as long as the
+    /// returned reference lives.
+    ///
+    /// CPU preemption must be disabled before calling this function and for the lifetime of the
+    /// returned reference. Otherwise, the returned reference might end up being a reference to a
+    /// different CPU's per-CPU area, causing the potential for a data race.
+    pub unsafe fn get_ref(&self) -> &MaybeUninit<T> {
+        // SAFETY: `self.get_ptr()` returns a valid pointer to a `MaybeUninit<T>` by its contract.
+        // The safety requirements of this function ensure that the returned reference isn't
+        // aliased by a `&mut MaybeUninit<T>`.
+        unsafe { &*self.get_ptr() }
+    }
+
+    /// Get a `*mut MaybeUninit<T>` to the per-CPU variable on the current CPU represented by
+    /// `&self`. Note that if CPU preemption is not disabled before calling this function, use of
+    /// the returned pointer may cause a data race without some other synchronization mechanism.
+    /// Buyer beware!
+    pub fn get_ptr(&self) -> *mut MaybeUninit<T> {
+        let this_cpu_off_pcpu: PerCpuPtr<*mut c_void> = get_static_per_cpu!(this_cpu_off).0;
+        let mut this_cpu_area: *mut c_void;
+        // SAFETY: gs + this_cpu_off_pcpu is guaranteed to be a valid pointer because `gs` points
+        // to the per-CPU area and this_cpu_off_pcpu is a valid per-CPU allocation.
+        unsafe {
+            asm!(
+                "mov {out}, gs:[{off_val}]",
+                off_val = in(reg) this_cpu_off_pcpu.0,
+                out = out(reg) this_cpu_area,
+            )
+        };
+
+        // This_cpu_area + self.0 is guaranteed to be a valid pointer by the per-CPU subsystem and
+        // the invariant that self.0 is a valid offset into the per-CPU area.
+        (this_cpu_area).wrapping_add(self.0 as usize).cast()
+    }
+}
+
+// SAFETY: Sending a `PerCpuPtr<T>` to another thread is safe because as soon as it's sent, the
+// pointer is logically referring to a different place in memory in the other CPU's per-CPU area.
+// In particular, this means that there are no restrictions on the type `T`.
+unsafe impl<T> Send for PerCpuPtr<T> {}
+
+// SAFETY: Two threads concurrently making use of a `PerCpuPtr<T>` will each see the `T` in their
+// own per-CPU area, so there's no potential for a data race (regardless of whether `T` is itself
+// `Sync`).
+unsafe impl<T> Sync for PerCpuPtr<T> {}
+
+impl<T> Clone for PerCpuPtr<T> {
+    fn clone(&self) -> Self {
+        *self
+    }
+}
+
+/// `PerCpuPtr` is just a wrapper around a pointer.
+impl<T> Copy for PerCpuPtr<T> {}
+
+/// A trait representing a per-CPU variable. This is implemented for both `StaticPerCpu<T>` and
+/// `DynamicPerCpu<T>`. The main usage of this trait is to call `get_mut` to get a `PerCpuToken`
+/// that can be used to access the underlying per-CPU variable. See `PerCpuToken::with`.
+pub trait PerCpu<T> {
+    /// Produces a token, asserting that the holder has exclusive access to the underlying memory
+    /// pointed to by `self`
+    ///
+    /// # Safety
+    /// No other `PerCpuToken` or `CheckedPerCpuToken` may exist on the current CPU (which is a
+    /// sensible notion, since we keep a `CpuGuard` around) that is derived from the same
+    /// `PerCpu<T>` or a clone thereof.
+    unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T>;
+}
+
+/// A marker trait for types that are interior mutable. Types that implement this trait can be used
+/// to create "checked" per-CPU variables. See `CheckedPerCpu<T>`.
+pub trait InteriorMutable {}
+
+impl<T> InteriorMutable for Cell<T> {}
+impl<T> InteriorMutable for RefCell<T> {}
+impl<T> InteriorMutable for UnsafeCell<T> {}
+impl<T> InteriorMutable for Opaque<T> {}
+
+/// A trait representing a per-CPU variable that is usable via a `&T`. The unsafety of `PerCpu<T>`
+/// stems from the fact that the holder of a `PerCpuToken` can use it to get a `&mut T` to the
+/// underlying per-CPU variable. This is problematic because the existence of aliaising `&mut T` is
+/// undefined behavior in Rust. This type avoids that issue by only allowing access via a `&T`,
+/// with the tradeoff that then `T` must be interior mutable or the underlying per-CPU variable
+/// must be a constant for the lifetime of the corresponding `CheckedPerCpu<T>`.
+///
+/// Currently, only the case where `T` is interior mutable has first-class support, though a custom
+/// implementation of `PerCpu<T>`/`CheckedPerCpu<T>` could be created for the const case.
+pub trait CheckedPerCpu<T>: PerCpu<T> {
+    /// Produces a token via which the holder can access the underlying per-CPU variable.
+    fn get(&mut self, guard: CpuGuard) -> CheckedPerCpuToken<'_, T>;
+}
+
+impl<'a, T> PerCpuToken<'a, T> {
+    /// # Safety
+    /// No other `PerCpuToken` or `CheckedPerCpuToken` may exist on the current CPU (which is a
+    /// sensible notion, since we keep a `CpuGuard` around) that uses the same `PerCpuPtr<T>`.
+    ///
+    /// The current CPU's memory location associated with the per-CPU variable pointed at by `ptr`
+    /// (i.e., the entry in the per-CPU area on this CPU) must be initialized.
+    pub unsafe fn new(guard: CpuGuard, ptr: &'a PerCpuPtr<T>) -> PerCpuToken<'a, T> {
+        Self { _guard: guard, ptr }
+    }
+
+    /// Immediately invokes `func` with a `&mut T` that points at the underlying per-CPU variable
+    /// that `&mut self` represents.
+    pub fn with<U>(&mut self, func: U)
+    where
+        U: FnOnce(&mut T),
+    {
+        // SAFETY: The existence of a PerCpuToken means that the requirements for get_mut_ref are
+        // satisfied. Likewise, the requirements for assume_init_mut are satisfied because the
+        // invariants of this type ensure that on the current CPU (which is a sensible notion
+        // because we have a CpuGuard), the memory location pointed to by `ptr` is initialized.
+        func(unsafe { self.ptr.get_mut_ref().assume_init_mut() });
+    }
+}
+
+impl<'a, T> CheckedPerCpuToken<'a, T> {
+    /// # Safety
+    /// The current CPU's memory location associated with the per-CPU variable pointed at by `ptr`
+    /// (i.e., the entry in the per-CPU area on this CPU) must be initialized.
+    pub unsafe fn new(guard: CpuGuard, ptr: &'a PerCpuPtr<T>) -> CheckedPerCpuToken<'a, T> {
+        Self { _guard: guard, ptr }
+    }
+
+    /// Immediately invokes `func` with a `&T` that points at the underlying per-CPU variable that
+    /// `&mut self` represents.
+    pub fn with<U>(&mut self, func: U)
+    where
+        U: FnOnce(&T),
+    {
+        // SAFETY: The existence of a CheckedPerCpuToken means that the requirements for get_ref
+        // are satisfied. Likewise, the requirements for assume_init_ref are satisfied because the
+        // invariants of this type ensure that on the current CPU (which is a sensible notion
+        // because we have a CpuGuard), the memory location pointed to by `ptr` is initialized.
+        func(unsafe { self.ptr.get_ref().assume_init_ref() });
+    }
+}
+
+declare_extern_per_cpu!(this_cpu_off: *mut c_void);

rust/kernel/percpu/cpu_guard.rs new file +35 -0

new file mode 100644
index 000000000000..14c04b12e7f0
@@ -0,0 +1,35 @@
+// SPDX-License-Identifier: GPL-2.0
+//! Contains abstractions for disabling CPU preemption. See `CpuGuard`.
+
+/// A RAII guard for bindings::preempt_disable and bindings::preempt_enable. Guarantees preemption
+/// is disabled for as long as this object exists.
+pub struct CpuGuard {
+    // Don't make one without using new()
+    _phantom: (),
+}
+
+impl CpuGuard {
+    /// Create a new CpuGuard. Disables preemption for its lifetime.
+    pub fn new() -> Self {
+        // SAFETY: There are no preconditions required to call preempt_disable
+        unsafe {
+            bindings::preempt_disable();
+        }
+        CpuGuard { _phantom: () }
+    }
+}
+
+impl Default for CpuGuard {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Drop for CpuGuard {
+    fn drop(&mut self) {
+        // SAFETY: There are no preconditions required to call preempt_enable
+        unsafe {
+            bindings::preempt_enable();
+        }
+    }
+}

rust/kernel/percpu/dynamic.rs new file +83 -0

new file mode 100644
index 000000000000..ce95e420f943
@@ -0,0 +1,83 @@
+// SPDX-License-Identifier: GPL-2.0
+//! Dynamically allocated per-CPU variables.
+
+use super::*;
+
+/// Represents a dynamic allocation of a per-CPU variable via alloc_percpu. Calls free_percpu when
+/// dropped.
+pub struct PerCpuAllocation<T>(PerCpuPtr<T>);
+
+impl<T: Zeroable> PerCpuAllocation<T> {
+    /// Dynamically allocates a space in the per-CPU area suitably sized and aligned to hold a `T`,
+    /// initially filled with the zero value for `T`.
+    ///
+    /// Returns `None` under the same circumstances the C function `alloc_percpu` returns `NULL`.
+    pub fn new_zero() -> Option<PerCpuAllocation<T>> {
+        let ptr: *mut MaybeUninit<T> =
+            // SAFETY: No preconditions to call alloc_percpu; MaybeUninit<T> is
+            // `#[repr(transparent)]`, so we can cast a `*mut T` to it.
+            unsafe { alloc_percpu(size_of::<T>(), align_of::<T>()) }.cast();
+        if ptr.is_null() {
+            return None;
+        }
+
+        // alloc_percpu returns zero'ed memory
+        Some(Self(PerCpuPtr(ptr)))
+    }
+}
+
+impl<T> PerCpuAllocation<T> {
+    /// Makes a per-CPU allocation sized and aligned to hold a `T`.
+    ///
+    /// Returns `None` under the same circumstances the C function `alloc_percpu` returns `NULL`.
+    pub fn new_uninit() -> Option<PerCpuAllocation<T>> {
+        let ptr: *mut MaybeUninit<T> =
+            // SAFETY: No preconditions to call alloc_percpu; MaybeUninit<T> is
+            // `#[repr(transparent)]`, so we can cast a `*mut T` to it.
+            unsafe { alloc_percpu(size_of::<T>(), align_of::<T>()) }.cast();
+        if ptr.is_null() {
+            return None;
+        }
+
+        Some(Self(PerCpuPtr(ptr)))
+    }
+}
+
+impl<T> Drop for PerCpuAllocation<T> {
+    fn drop(&mut self) {
+        // SAFETY: self.0.0 was returned by alloc_percpu, and so was a valid pointer into
+        // the percpu area, and has remained valid by the invariants of PerCpuAllocation<T>.
+        unsafe { free_percpu(self.0 .0.cast()) }
+    }
+}
+
+/// Holds a dynamically-allocated per-CPU variable.
+#[derive(Clone)]
+pub struct DynamicPerCpu<T> {
+    // INVARIANT: The memory location in each CPU's per-CPU area pointed at by `alloc.0` has been
+    // initialized.
+    pub(super) alloc: Arc<PerCpuAllocation<T>>,
+}
+
+impl<T: Zeroable> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable
+    ///
+    /// # Arguments
+    /// * `flags` - Flags used to allocate an `Arc` that keeps track of the underlying
+    ///   `PerCpuAllocation`.
+    pub fn new_zero(flags: Flags) -> Option<Self> {
+        let alloc: PerCpuAllocation<T> = PerCpuAllocation::new_zero()?;
+
+        let arc = Arc::new(alloc, flags).ok()?;
+
+        Some(Self { alloc: arc })
+    }
+}
+
+impl<T> PerCpu<T> for DynamicPerCpu<T> {
+    unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
+        // SAFETY: The requirements of `PerCpu::get_mut` and this type's invariant ensure that the
+        // requirements of `PerCpuToken::new` are met.
+        unsafe { PerCpuToken::new(guard, &self.alloc.0) }
+    }
+}

rust/kernel/percpu/static_.rs new file +132 -0

new file mode 100644
index 000000000000..be226dd2c3aa
@@ -0,0 +1,132 @@
+// SPDX-License-Identifier: GPL-2.0
+//! Statically allocated per-CPU variables.
+
+use super::*;
+
+/// A wrapper used for declaring static per-CPU variables. These symbols are "virtual" in that the
+/// linker uses them to generate offsets into each CPU's per-CPU area, but shouldn't be read
+/// from/written to directly. The fact that the statics are immutable prevents them being written
+/// to (generally), this struct having _val be non-public prevents reading from them.
+///
+/// The end-user of the per-CPU API should make use of the define_per_cpu! macro instead of
+/// declaring variables of this type directly. All instances of this type must be `static` and
+/// `#[link_section = ".data..percpu"]` (which the macro handles).
+#[repr(transparent)]
+pub struct StaticPerCpuSymbol<T> {
+    _val: T, // generate a correctly sized type
+}
+
+/// Holds a statically-allocated per-CPU variable.
+#[derive(Clone)]
+pub struct StaticPerCpu<T>(pub(super) PerCpuPtr<T>);
+
+impl<T> StaticPerCpu<T> {
+    /// Creates a `StaticPerCpu<T>` from a `StaticPerCpuSymbol<T>`. You should probably be using
+    /// `get_static_per_cpu!` instead.
+    pub fn new(ptr: *const StaticPerCpuSymbol<T>) -> StaticPerCpu<T> {
+        // SAFETY: `StaticPerCpuSymbol<T>` is `#[repr(transparent)]`, so we can safely cast a
+        // pointer to it into a pointer to `MaybeUninit<T>`. The validity of it as a per-CPU
+        // pointer is guaranteed by the per-CPU subsystem and invariants of the StaticPerCpuSymbol
+        // type.
+        let pcpu_ptr = unsafe { PerCpuPtr::new(ptr.cast_mut().cast()) };
+        Self(pcpu_ptr)
+    }
+}
+
+impl<T> PerCpu<T> for StaticPerCpu<T> {
+    unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
+        // SAFETY: The requirements of `PerCpu::get_mut` and the fact that statically-allocated
+        // per-CPU variables are initialized by the per-CPU subsystem ensure that the requirements
+        // of `PerCpuToken::new` are met.
+        unsafe { PerCpuToken::new(guard, &self.0) }
+    }
+}
+
+impl<T: InteriorMutable> CheckedPerCpu<T> for StaticPerCpu<T> {
+    fn get(&mut self, guard: CpuGuard) -> CheckedPerCpuToken<'_, T> {
+        // SAFETY: The per-CPU subsystem guarantees that each CPU's instance of a
+        // statically allocated variable begins with a copy of the contents of the
+        // corresponding symbol in `.data..percpu`. Thus, the requirements of
+        // `CheckedPerCpuToken::new` are met.
+        unsafe { CheckedPerCpuToken::new(guard, &self.0) }
+    }
+}
+
+/// Gets a `StaticPerCpu<T>` from a symbol declared with `define_per_cpu!` or
+/// `declare_extern_per_cpu!`.
+///
+/// # Arguments
+/// * `ident` - The identifier declared
+#[macro_export]
+macro_rules! get_static_per_cpu {
+    ($id:ident) => {
+        $crate::percpu::StaticPerCpu::new((&raw const $id).cast())
+    };
+}
+
+/// Declares a StaticPerCpuSymbol corresponding to a per-CPU variable defined in C. Be sure to read
+/// the safety requirements of `PerCpu::get`.
+#[macro_export]
+macro_rules! declare_extern_per_cpu {
+    ($id:ident: $ty:ty) => {
+        extern "C" {
+            static $id: StaticPerCpuSymbol<$ty>;
+        }
+    };
+}
+
+/// define_per_cpu! is analogous to the C DEFINE_PER_CPU macro in that it lets you create a
+/// statically allocated per-CPU variable.
+///
+/// # Example
+/// ```
+/// use kernel::define_per_cpu;
+/// use kernel::percpu::StaticPerCpuSymbol;
+///
+/// define_per_cpu!(pub MY_PERCPU: u64 = 0);
+/// ```
+#[macro_export]
+macro_rules! define_per_cpu {
+    ($vis:vis $id:ident: $ty:ty = $expr:expr) => {
+        $crate::macros::paste! {
+            // We might want to have a per-CPU variable that doesn't implement `Sync` (not paying
+            // sync overhead costs is part of the point), but Rust won't let us declare a static of
+            // a `!Sync` type. Of course, we don't actually have any synchronization issues, since
+            // each CPU will see its own copy of the variable, so we cheat a little bit and tell
+            // Rust it's fine.
+            #[doc(hidden)]
+            #[allow(non_camel_case_types)]
+            #[repr(transparent)] // It needs to be the same size as $ty
+            struct [<__PRIVATE_TYPE_ $id>]($ty);
+
+            impl [<__PRIVATE_TYPE_ $id>] {
+                #[doc(hidden)]
+                const fn new(val: $ty) -> Self {
+                    Self(val)
+                }
+            }
+
+            // Expand $expr outside of the unsafe block to avoid silently allowing unsafe code to be
+            // used without a user-facing unsafe block
+            #[doc(hidden)]
+            static [<__INIT_ $id>]: [<__PRIVATE_TYPE_ $id>] = [<__PRIVATE_TYPE_ $id>]::new($expr);
+
+            // SAFETY: This type will ONLY ever be used to declare a `StaticPerCpuSymbol`
+            // (which we then only ever use as input to `&raw`). Reading from the symbol is
+            // already UB, so we won't ever actually have any variables of this type where
+            // synchronization is a concern.
+            #[doc(hidden)]
+            unsafe impl Sync for [<__PRIVATE_TYPE_ $id>] {}
+
+            // SAFETY: StaticPerCpuSymbol<T> is #[repr(transparent)], so we can freely convert from
+            // [<__PRIVATE_TYPE_ $id>], which is also `#[repr(transparent)]` (i.e., everything is
+            // just a `$ty` from a memory layout perspective).
+            #[link_section = ".data..percpu"]
+            $vis static $id: StaticPerCpuSymbol<[<__PRIVATE_TYPE_ $id>]> = unsafe {
+                core::mem::transmute_copy::<
+                    [<__PRIVATE_TYPE_ $id>], StaticPerCpuSymbol<[<__PRIVATE_TYPE_ $id>]>
+                >(&[<__INIT_ $id>])
+            };
+        }
+    };
+}

Signature · 2 lines

-- 
2.34.1

Quoted 15 lines · Mitchell Levy

Per-CPU variables are an important tool for reducing lock contention, especially in systems with many processors. They also provide a convenient way to handle data that are logically associated with a particular CPU (e.g., the currently running task). Therefore, add a Rust API to make use of per-CPU variables.

Add a CpuGuard type that disables preemption for its lifetime. Add a PerCpuAllocation type used to track dynamic allocations. Add a define_per_cpu! macro to create static per-CPU allocations. Add DynamicPerCpu and StaticPerCpu to provide a high-level API. Add a PerCpu trait to unify the dynamic and static cases.

Co-developed-by Boqun Feng <[email protected]>

Signed-off-by Boqun Feng <[email protected]>

Signed-off-by Mitchell Levy <[email protected]>

Quoted 10 lines

---
 rust/helpers/helpers.c          |   2 +
 rust/helpers/percpu.c           |   9 ++
 rust/helpers/preempt.c          |  14 +++
 rust/kernel/lib.rs              |   3 +
 rust/kernel/percpu.rs           | 223 ++++++++++++++++++++++++++++++++++++++++
 rust/kernel/percpu/cpu_guard.rs |  35 +++++++
 rust/kernel/percpu/dynamic.rs   |  83 +++++++++++++++
 rust/kernel/percpu/static_.rs   | 132 ++++++++++++++++++++++++
 8 files changed, 501 insertions(+)

That's quite a massive patch. Can you please split-out C-binders, and maybe make separate patches for each .rs component?

Quoted 2 lines

rust/helpers/helpers.c +0 -0

index 7cf7fe95e41d..2fc8d26cfe66 100644

Quoted 61 lines

rust/kernel/percpu.rs new file +58 -0

@@ -0,0 +1,223 @@
+// SPDX-License-Identifier: GPL-2.0
+//! This module contains abstractions for creating and using per-CPU variables from Rust.
+//! See the define_per_cpu! macro and the DynamicPerCpu<T> type, as well as the PerCpu<T> trait.
+
+pub mod cpu_guard;
+mod dynamic;
+mod static_;
+
+#[doc(inline)]
+pub use dynamic::*;
+#[doc(inline)]
+pub use static_::*;
+
+use bindings::{alloc_percpu, free_percpu};
+
+use crate::alloc::Flags;
+use crate::percpu::cpu_guard::CpuGuard;
+use crate::prelude::*;
+use crate::sync::Arc;
+use crate::types::Opaque;
+use crate::{declare_extern_per_cpu, get_static_per_cpu};
+
+use core::arch::asm;
+use core::cell::{Cell, RefCell, UnsafeCell};
+use core::mem::{align_of, size_of, MaybeUninit};
+
+use ffi::c_void;
+
+/// A per-CPU pointer; that is, an offset into the per-CPU area. Note that this type is NOT a smart
+/// pointer, it does not manage the allocation.
+pub struct PerCpuPtr<T>(*mut MaybeUninit<T>);
+
+/// Represents exclusive access to the memory location pointed at by a particular PerCpu<T>.
+pub struct PerCpuToken<'a, T> {
+    // INVARIANT: the current CPU's memory location associated with the per-CPU variable pointed at
+    // by `ptr` (i.e., the entry in the per-CPU area on the current CPU) has been initialized.
+    _guard: CpuGuard,
+    ptr: &'a PerCpuPtr<T>,
+}
+
+/// Represents access to the memory location pointed at by a particular PerCpu<T> where the type
+/// `T` manages access to the underlying memory to avoid aliaising troubles. (For example, `T`
+/// might be a `Cell` or `RefCell`.)
+pub struct CheckedPerCpuToken<'a, T> {
+    // INVARIANT: the current CPU's memory location associated with the per-CPU variable pointed at
+    // by `ptr` (i.e., the entry in the per-CPU area on the current CPU) has been initialized.
+    _guard: CpuGuard,
+    ptr: &'a PerCpuPtr<T>,
+}
+
+impl<T> PerCpuPtr<T> {
+    /// Makes a new PerCpuPtr from a raw per-CPU pointer.
+    ///
+    /// # Safety
+    /// `ptr` must be a valid per-CPU pointer.
+    pub unsafe fn new(ptr: *mut MaybeUninit<T>) -> Self {
+        Self(ptr)
+    }

rust/kernel/percpu.rs line 64

Quoted 6 lines

rust/kernel/percpu.rs new file +6 -0

+
+    /// Get a `&mut MaybeUninit<T>` to the per-CPU variable on the current CPU represented by `&self`
+    ///
+    /// # Safety
+    /// The returned `&mut T` must follow Rust's aliasing rules. That is, no other `&(mut) T` may
+    /// exist that points to the same location in memory. In practice, this means that `get_(mut_)ref`

How long is this line?

rust/kernel/percpu.rs line 77

Quoted 13 lines

rust/kernel/percpu.rs new file +13 -0

+    /// must not be called on another `PerCpuPtr<T>` that is a copy/clone of `&self` for as long as
+    /// the returned reference lives.
+    ///
+    /// CPU preemption must be disabled before calling this function and for the lifetime of the
+    /// returned reference. Otherwise, the returned reference might end up being a reference to a
+    /// different CPU's per-CPU area, causing the potential for a data race.
+    #[allow(clippy::mut_from_ref)] // Safety requirements prevent aliasing issues
+    pub unsafe fn get_mut_ref(&self) -> &mut MaybeUninit<T> {
+        // SAFETY: `self.get_ptr()` returns a valid pointer to a `MaybeUninit<T>` by its contract,
+        // and the safety requirements of this function ensure that the returned reference is
+        // exclusive.
+        unsafe { &mut *(self.get_ptr()) }
+    }

Here and everywhere: would it make sense to enforce it by testing the CPU with preemptible() before returning a reference?

Thanks,
Yury

Quoted 1 line · Yury Norov

On Thu, Aug 28, 2025 at 12:00:08PM -0700, Mitchell Levy wrote:

Quoted 7 lines

rust/kernel/percpu.rs line 64

Quoted 5 lines

rust/kernel/percpu.rs new file +5 -0

+    /// Get a `&mut MaybeUninit<T>` to the per-CPU variable on the current CPU represented by `&self`
+    ///
+    /// # Safety
+    /// The returned `&mut T` must follow Rust's aliasing rules. That is, no other `&(mut) T` may
+    /// exist that points to the same location in memory. In practice, this means that `get_(mut_)ref`

How long is this line?

102 chars, or 103 if you include the newline. rustfmt doesn't break the line, so I left it as-is for this patch. Happy to change it if it poses a problem, though.

Quoted 14 lines

rust/kernel/percpu.rs line 77

Quoted 13 lines

rust/kernel/percpu.rs new file +13 -0

+    /// must not be called on another `PerCpuPtr<T>` that is a copy/clone of `&self` for as long as
+    /// the returned reference lives.
+    ///
+    /// CPU preemption must be disabled before calling this function and for the lifetime of the
+    /// returned reference. Otherwise, the returned reference might end up being a reference to a
+    /// different CPU's per-CPU area, causing the potential for a data race.
+    #[allow(clippy::mut_from_ref)] // Safety requirements prevent aliasing issues
+    pub unsafe fn get_mut_ref(&self) -> &mut MaybeUninit<T> {
+        // SAFETY: `self.get_ptr()` returns a valid pointer to a `MaybeUninit<T>` by its contract,
+        // and the safety requirements of this function ensure that the returned reference is
+        // exclusive.
+        unsafe { &mut *(self.get_ptr()) }
+    }

Quoted 2 lines

Here and everywhere: would it make sense to enforce it by testing the CPU with preemptible() before returning a reference?

The only thing we could do would be to panic, which I don't 100% love. Another alternative would be to take a &'a CpuGuard and bound the lifetime of the returned reference to 'a, and then we don't need to do any run-time checking at all.

Originally, I had left this to the caller because it might make sense down the line for some complex behavior based on per-CPU (e.g., per-CPU refcount) to do all its own management of per-CPU variables using PerCpuPtr as a core primitive. In these cases, I believe there are some times where being non-preemptible wouldn't actually be required (that said, my thoughts on this aren't well reflected in the safety comment, since I said it must be disabled... gah). But, the more I think about it, the more I think these use cases would be better served by just using get_ptr --- conjuring &mut references seems like it would be a big footgun. And the safety comment already actually reflects these thoughts somewhat :)

For v4 I will probably have this function take a &'a CpuGuard and use that to bound the liftetime of the returned reference, unless there are other thoughts on this point.

Quoted 2 lines · Mitchell Levy

Quoted 1 line · Yury Norov

On Thu, Aug 28, 2025 at 12:00:08PM -0700, Mitchell Levy wrote:

...

Quoted 6 lines

Quoted 6 lines

Quoted 5 lines

rust/kernel/percpu.rs new file +5 -0

+    /// Get a `&mut MaybeUninit<T>` to the per-CPU variable on the current CPU represented by `&self`
+    ///
+    /// # Safety
+    /// The returned `&mut T` must follow Rust's aliasing rules. That is, no other `&(mut) T` may
+    /// exist that points to the same location in memory. In practice, this means that `get_(mut_)ref`

Quoted 5 lines

Quoted 1 line

How long is this line?

102 chars, or 103 if you include the newline. rustfmt doesn't break the line, so I left it as-is for this patch. Happy to change it if it poses a problem, though.

Then don't use that tool - it's broken. In kernel we used to have 80-chars limit for the lines, recently relaxed to 100.

Quoted 22 lines

Quoted 14 lines

rust/kernel/percpu.rs line 77

Quoted 13 lines

rust/kernel/percpu.rs new file +13 -0

+    /// must not be called on another `PerCpuPtr<T>` that is a copy/clone of `&self` for as long as
+    /// the returned reference lives.
+    ///
+    /// CPU preemption must be disabled before calling this function and for the lifetime of the
+    /// returned reference. Otherwise, the returned reference might end up being a reference to a
+    /// different CPU's per-CPU area, causing the potential for a data race.
+    #[allow(clippy::mut_from_ref)] // Safety requirements prevent aliasing issues
+    pub unsafe fn get_mut_ref(&self) -> &mut MaybeUninit<T> {
+        // SAFETY: `self.get_ptr()` returns a valid pointer to a `MaybeUninit<T>` by its contract,
+        // and the safety requirements of this function ensure that the returned reference is
+        // exclusive.
+        unsafe { &mut *(self.get_ptr()) }
+    }

Quoted 2 lines

Here and everywhere: would it make sense to enforce it by testing the CPU with preemptible() before returning a reference?

The only thing we could do would be to panic, which I don't 100% love. Another alternative would be to take a &'a CpuGuard and bound the lifetime of the returned reference to 'a, and then we don't need to do any run-time checking at all.

Quoted 11 lines

Originally, I had left this to the caller because it might make sense down the line for some complex behavior based on per-CPU (e.g., per-CPU refcount) to do all its own management of per-CPU variables using PerCpuPtr as a core primitive. In these cases, I believe there are some times where being non-preemptible wouldn't actually be required (that said, my thoughts on this aren't well reflected in the safety comment, since I said it must be disabled... gah). But, the more I think about it, the more I think these use cases would be better served by just using get_ptr --- conjuring &mut references seems like it would be a big footgun. And the safety comment already actually reflects these thoughts somewhat :)

If you think that in future there will be a user who will not need to disable preemption before dereferencing a percpu pointer, then you can add another less restricted flavor of the helper.

Quoted 3 lines

For v4 I will probably have this function take a &'a CpuGuard and use that to bound the liftetime of the returned reference, unless there are other thoughts on this point.

I don't want you to panic just because of invlid user call, but whatever you call in comment must be enforced in code, right?

You can use the guard, if it guarantees the preemption disabled; or you can return None; you can create CONFIG_RUST_PERCPU_HARDENED for panics.

Please refer the recent bitmap API wrapper, and how erroneous request is handled there.

https://lore.kernel.org/all/[email protected]/

Add an iterator for Cpumask making use of C's cpumask_next.

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/helpers/cpumask.c |  5 +++++
 rust/kernel/cpumask.rs | 48 +++++++++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 52 insertions(+), 1 deletion(-)

Add getters for the global cpumasks documented in
include/linux/cpumask.h, specifically:
- cpu_possible_mask
- cpu_online_mask
- cpu_enabled_mask
- cpu_present_mask
- cpu_active_mask

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/kernel/cpumask.rs | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

Add getters for the global cpumasks documented in
include/linux/cpumask.h, specifically:
- cpu_possible_mask
- cpu_online_mask
- cpu_enabled_mask
- cpu_present_mask
- cpu_active_mask

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/kernel/cpumask.rs | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

rust/kernel/cpumask.rs +0 -0

index b7401848f59e..ca9aa6875b4a 100644

rust/kernel/cpumask.rs +39 -0

+
+    /// Get a CPU mask representing online CPUs; has bit `cpu` set iff cpu available to the
+    /// scheduler
+    #[inline]
+    pub fn online() -> &'static Self {
+        // SAFETY: `__cpu_online_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_online_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_online_mask) }
+    }
+
+    /// Get a CPU mask representing enabled CPUs; has bit `cpu` set iff cpu can be brought online
+    #[inline]
+    pub fn enabled() -> &'static Self {
+        // SAFETY: `__cpu_enabled_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_enabled_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_enabled_mask) }
+    }
+
+    /// Get a CPU mask representing present CPUs; has bit `cpu` set iff cpu is populated
+    #[inline]
+    pub fn present() -> &'static Self {
+        // SAFETY: `__cpu_present_mask` is a valid global provided by the kernel that lives
+        // forever. Since we wrap the returned pointer in an `Opaque`, it's ok that
+        // `__cpu_present_mask` may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_present_mask) }
+    }
+
+    /// Get a CPU mask representing active CPUs; has bit `cpu` set iff cpu is available to
+    /// migration.
+    #[inline]
+    pub fn active() -> &'static Self {
+        // SAFETY: `__cpu_active_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_active_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_active_mask) }
+    }
+
     /// Obtain the raw `struct cpumask` pointer.
     pub fn as_raw(&self) -> *mut bindings::cpumask {
         let this: *const Self = self;

Signature · 2 lines

-- 
2.34.1

Quoted 21 lines · Mitchell Levy

Add getters for the global cpumasks documented in
include/linux/cpumask.h, specifically:
- cpu_possible_mask
- cpu_online_mask
- cpu_enabled_mask
- cpu_present_mask
- cpu_active_mask

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/kernel/cpumask.rs | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

rust/kernel/cpumask.rs +0 -0

index b7401848f59e..ca9aa6875b4a 100644
@@ -77,6 +77,52 @@ pub unsafe fn as_ref<'a>(ptr: *const bindings::cpumask) -> &'a Self {
         unsafe { &*ptr.cast() }
     }
 

rust/kernel/cpumask.rs pub unsafe fn as_ref<'a>(ptr: *const bindings::cpumask) -> &'a Self { · line 86

Quoted 7 lines

rust/kernel/cpumask.rs +7 -0

+    /// Get a CPU mask representing possible CPUs; has bit `cpu` set iff cpu is populatable
+    #[inline]
+    pub fn possible() -> &'static Self {
+        // SAFETY: `__cpu_possible_mask` is a valid global provided by the kernel that lives
+        // forever.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_possible_mask) }
+    }

This function and the following look like returning a boolean. Maybe possible_cpus()?

The general rule is to keep the rust naming as close to C as possible, but in this case Cpumask::cpu_possible_mask is, OK, somewhat excessive.

Thanks,
Yury

Quoted 45 lines

rust/kernel/cpumask.rs +39 -0

+
+    /// Get a CPU mask representing online CPUs; has bit `cpu` set iff cpu available to the
+    /// scheduler
+    #[inline]
+    pub fn online() -> &'static Self {
+        // SAFETY: `__cpu_online_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_online_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_online_mask) }
+    }
+
+    /// Get a CPU mask representing enabled CPUs; has bit `cpu` set iff cpu can be brought online
+    #[inline]
+    pub fn enabled() -> &'static Self {
+        // SAFETY: `__cpu_enabled_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_enabled_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_enabled_mask) }
+    }
+
+    /// Get a CPU mask representing present CPUs; has bit `cpu` set iff cpu is populated
+    #[inline]
+    pub fn present() -> &'static Self {
+        // SAFETY: `__cpu_present_mask` is a valid global provided by the kernel that lives
+        // forever. Since we wrap the returned pointer in an `Opaque`, it's ok that
+        // `__cpu_present_mask` may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_present_mask) }
+    }
+
+    /// Get a CPU mask representing active CPUs; has bit `cpu` set iff cpu is available to
+    /// migration.
+    #[inline]
+    pub fn active() -> &'static Self {
+        // SAFETY: `__cpu_active_mask` is a valid global provided by the kernel that lives forever.
+        // Since we wrap the returned pointer in an `Opaque`, it's ok that `__cpu_active_mask`
+        // may change its value.
+        unsafe { Cpumask::as_ref(&raw const bindings::__cpu_active_mask) }
+    }
+
     /// Obtain the raw `struct cpumask` pointer.
     pub fn as_raw(&self) -> *mut bindings::cpumask {
         let this: *const Self = self;

Signature · 2 lines

-- 
2.34.1

Add functionality to PerCpuPtr to compute pointers to per-CPU variable slots on other CPUs. Use this facility to initialize per-CPU variables on all possible CPUs when a dynamic per-CPU variable is created with a non-zeroable type. Since RefCell and other Cell-like types fall into this category, impl CheckedPerCpu on DynamicPerCpu for these InteriorMutable types since they can now be used. Add examples of these usages to samples/rust/rust_percpu.rs.

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/helpers/percpu.c         |  5 +++
 rust/kernel/percpu.rs         | 15 +++++++
 rust/kernel/percpu/dynamic.rs | 40 +++++++++++++++++
 samples/rust/rust_percpu.rs   | 99 ++++++++++++++++++++++++++++++++++++++++---
 4 files changed, 152 insertions(+), 7 deletions(-)

rust/helpers/percpu.c +5 -0

index 8cc01d094752..8d83b6b86106 100644
@@ -8,6 +8,11 @@ void __percpu *rust_helper_alloc_percpu(size_t sz, size_t align)
 	return __alloc_percpu(sz, align);
 }
 
+void *rust_helper_per_cpu_ptr(void __percpu *ptr, unsigned int cpu)
+{
+	return per_cpu_ptr(ptr, cpu);
+}
+
 void rust_helper_on_each_cpu(smp_call_func_t func, void *info, int wait)
 {
 	on_each_cpu(func, info, wait);

rust/kernel/percpu.rs +1 -0

index 35afcdba3ccd..c68c7520b67f 100644
@@ -14,6 +14,7 @@
 use bindings::{alloc_percpu, free_percpu};
 
 use crate::alloc::Flags;
+use crate::cpu::CpuId;
 use crate::percpu::cpu_guard::CpuGuard;
 use crate::prelude::*;
 use crate::sync::Arc;

@@ -115,6 +116,20 @@ pub fn get_ptr(&self) -> *mut MaybeUninit<T> {
         // the invariant that self.0 is a valid offset into the per-CPU area.
         (this_cpu_area).wrapping_add(self.0 as usize).cast()
     }

rust/kernel/percpu.rs +4 -0

+    pub unsafe fn get_remote_ptr(&self, cpu: CpuId) -> *mut MaybeUninit<T> {
+        // SAFETY: The requirements of this function ensure this call is safe.
+        unsafe { bindings::per_cpu_ptr(self.0.cast(), cpu.as_u32()) }.cast()
+    }
 }
 
 // SAFETY: Sending a `PerCpuPtr<T>` to another thread is safe because as soon as it's sent, the

rust/kernel/percpu/dynamic.rs +4 -0

index ce95e420f943..64f04cef3705 100644
@@ -3,6 +3,8 @@
 
 use super::*;
 
+use crate::cpumask::Cpumask;
+
 /// Represents a dynamic allocation of a per-CPU variable via alloc_percpu. Calls free_percpu when
 /// dropped.
 pub struct PerCpuAllocation<T>(PerCpuPtr<T>);

@@ -74,6 +76,36 @@ pub fn new_zero(flags: Flags) -> Option<Self> {
     }
 }
 
+impl<T: Clone> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable

rust/kernel/percpu/dynamic.rs +26 -0

+            // SAFETY: `ptr` is a valid allocation, and `cpu` appears in `Cpumask::possible()`
+            let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
+            // SAFETY: Each CPU's slot corresponding to `ptr` is currently uninitialized, and no
+            // one else has a reference to it. Therefore, we can freely write to it without
+            // worrying about the need to drop what was there or whether we're racing with someone
+            // else. `PerCpuPtr::get_remote_ptr` guarantees that the pointer is valid since we
+            // derived it from a valid allocation and `cpu`.
+            unsafe {
+                (*remote_ptr).write(val.clone());
+            }
+        }
+
+        let arc = Arc::new(alloc, flags).ok()?;
+
+        Some(Self { alloc: arc })
+    }
+}
+
 impl<T> PerCpu<T> for DynamicPerCpu<T> {
     unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
         // SAFETY: The requirements of `PerCpu::get_mut` and this type's invariant ensure that the

@@ -81,3 +113,11 @@ unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
         unsafe { PerCpuToken::new(guard, &self.alloc.0) }
     }
 }
+
+impl<T: InteriorMutable> CheckedPerCpu<T> for DynamicPerCpu<T> {
+    fn get(&mut self, guard: CpuGuard) -> CheckedPerCpuToken<'_, T> {
+        // SAFETY: By the invariant of `DynamicPerCpu`, the memory location in each CPU's
+        // per-CPU area corresponding to this variable has been initialized.
+        unsafe { CheckedPerCpuToken::new(guard, &self.alloc.0) }
+    }
+}

samples/rust/rust_percpu.rs +92 -7

index 98ca1c781b6b..06b322019134 100644
@@ -130,13 +130,72 @@ fn init(_module: &'static ThisModule) -> Result<Self, Error> {
 
         // SAFETY: No prerequisites for on_each_cpu.
         unsafe {
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 1);
-            on_each_cpu(Some(check_percpu), (&raw mut test).cast(), 1);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 1);
+            on_each_cpu(Some(check_percpu_u64), (&raw mut test).cast(), 1);
         }
 
+        let mut checked: DynamicPerCpu<RefCell<u64>> =
+            DynamicPerCpu::new_with(RefCell::new(100), GFP_KERNEL).unwrap();
+
+        // SAFETY: No prerequisites for on_each_cpu.
+        unsafe {
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 1);
+            on_each_cpu(Some(check_percpu_refcell_u64), (&raw mut checked).cast(), 1);
+        }
+
+        checked.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+            assert!(*val.borrow() == 104);
+
+            let mut checked_native = 0;
+            *val.borrow_mut() = 0;
+
+            checked_native += 1;
+            *val.borrow_mut() += 1;
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == 1);
+
+            checked_native = checked_native.wrapping_add((-1i64) as u64);
+            val.replace_with(|old: &mut u64| old.wrapping_add((-1i64) as u64));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == 0);
+
+            checked_native = checked_native.wrapping_add((-1i64) as u64);
+            val.replace_with(|old: &mut u64| old.wrapping_add((-1i64) as u64));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == (-1i64) as u64);
+
+            checked_native = 0;
+            *val.borrow_mut() = 0;
+
+            checked_native = checked_native.wrapping_sub(1);
+            val.replace_with(|old: &mut u64| old.wrapping_sub(1));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == (-1i64) as u64);
+            assert!(checked_native == *val.borrow() && checked_native == u64::MAX);
+        });
+
         pr_info!("rust dynamic percpu test done\n");
 
         // Return Err to unload the module

@@ -144,7 +203,7 @@ fn init(_module: &'static ThisModule) -> Result<Self, Error> {
     }
 }
 
-extern "C" fn inc_percpu(info: *mut c_void) {
+extern "C" fn inc_percpu_u64(info: *mut c_void) {
     // SAFETY: We know that info is a void *const DynamicPerCpu<u64> and DynamicPerCpu<u64> is Send.
     let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<u64>)).clone() };
     pr_info!("Incrementing on {}\n", CpuId::current().as_u32());

@@ -153,7 +212,7 @@ extern "C" fn inc_percpu(info: *mut c_void) {
     unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val: &mut u64| *val += 1);
 }
 
-extern "C" fn check_percpu(info: *mut c_void) {
+extern "C" fn check_percpu_u64(info: *mut c_void) {
     // SAFETY: We know that info is a void *const DynamicPerCpu<u64> and DynamicPerCpu<u64> is Send.
     let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<u64>)).clone() };
     pr_info!("Asserting on {}\n", CpuId::current().as_u32());

@@ -161,3 +220,29 @@ extern "C" fn check_percpu(info: *mut c_void) {
     // SAFETY: We don't have multiple clones of pcpu in scope
     unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val: &mut u64| assert!(*val == 4));
 }
+
+extern "C" fn inc_percpu_refcell_u64(info: *mut c_void) {
+    // SAFETY: We know that info is a void *const DynamicPerCpu<RefCell<u64>> and
+    // DynamicPerCpu<RefCell<u64>> is Send.
+    let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<RefCell<u64>>)).clone() };
+    // SAFETY: smp_processor_id has no preconditions
+    pr_info!("Incrementing on {}\n", CpuId::current().as_u32());
+
+    pcpu.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+        let mut val = val.borrow_mut();
+        *val += 1;
+    });
+}
+
+extern "C" fn check_percpu_refcell_u64(info: *mut c_void) {
+    // SAFETY: We know that info is a void *const DynamicPerCpu<RefCell<u64>> and
+    // DynamicPerCpu<RefCell<u64>> is Send.
+    let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<RefCell<u64>>)).clone() };
+    // SAFETY: smp_processor_id has no preconditions
+    pr_info!("Asserting on {}\n", CpuId::current().as_u32());
+
+    pcpu.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+        let val = val.borrow();
+        assert!(*val == 104);
+    });
+}

Signature · 2 lines

-- 
2.34.1

Quoted 54 lines · Mitchell Levy

Add functionality to PerCpuPtr to compute pointers to per-CPU variable slots on other CPUs. Use this facility to initialize per-CPU variables on all possible CPUs when a dynamic per-CPU variable is created with a non-zeroable type. Since RefCell and other Cell-like types fall into this category, impl CheckedPerCpu on DynamicPerCpu for these InteriorMutable types since they can now be used. Add examples of these usages to samples/rust/rust_percpu.rs.

Signed-off-by Mitchell Levy <[email protected]>

---
 rust/helpers/percpu.c         |  5 +++
 rust/kernel/percpu.rs         | 15 +++++++
 rust/kernel/percpu/dynamic.rs | 40 +++++++++++++++++
 samples/rust/rust_percpu.rs   | 99 ++++++++++++++++++++++++++++++++++++++++---
 4 files changed, 152 insertions(+), 7 deletions(-)

rust/helpers/percpu.c +5 -0

index 8cc01d094752..8d83b6b86106 100644
@@ -8,6 +8,11 @@ void __percpu *rust_helper_alloc_percpu(size_t sz, size_t align)
 	return __alloc_percpu(sz, align);
 }
 
+void *rust_helper_per_cpu_ptr(void __percpu *ptr, unsigned int cpu)
+{
+	return per_cpu_ptr(ptr, cpu);
+}
+
 void rust_helper_on_each_cpu(smp_call_func_t func, void *info, int wait)
 {
 	on_each_cpu(func, info, wait);

rust/kernel/percpu.rs +7 -0

index 35afcdba3ccd..c68c7520b67f 100644
@@ -14,6 +14,7 @@
 use bindings::{alloc_percpu, free_percpu};
 
 use crate::alloc::Flags;
+use crate::cpu::CpuId;
 use crate::percpu::cpu_guard::CpuGuard;
 use crate::prelude::*;
 use crate::sync::Arc;

@@ -115,6 +116,20 @@ pub fn get_ptr(&self) -> *mut MaybeUninit<T> {
         // the invariant that self.0 is a valid offset into the per-CPU area.
         (this_cpu_area).wrapping_add(self.0 as usize).cast()
     }
+
+    /// Get a `*mut MaybeUninit<T>` to the per-CPU variable on the CPU represented by `cpu`. Note
+    /// that without some kind of synchronization, use of the returned pointer may cause a data
+    /// race. It is the caller's responsibility to use the returned pointer in a reasonable way.
+    ///
+    /// # Safety

rust/kernel/percpu.rs pub fn get_ptr(&self) -> *mut MaybeUninit<T> { · line 128

Quoted 4 lines

rust/kernel/percpu.rs +4 -0

+    /// - The returned pointer is valid only if `self` is (that is, it points to a live allocation
+    ///   correctly sized and aligned to hold a `T`)
+    /// - The returned pointer is valid only if the bit corresponding to `cpu` is set in
+    ///   `Cpumask::possible()`.

Instead of explaining those rules in comments, can you just enforce them in code? Not sure about the 1st rule, but the 2nd one looks like a trivial check.

Quoted 33 lines

rust/kernel/percpu.rs +4 -0

+    pub unsafe fn get_remote_ptr(&self, cpu: CpuId) -> *mut MaybeUninit<T> {
+        // SAFETY: The requirements of this function ensure this call is safe.
+        unsafe { bindings::per_cpu_ptr(self.0.cast(), cpu.as_u32()) }.cast()
+    }
 }
 
 // SAFETY: Sending a `PerCpuPtr<T>` to another thread is safe because as soon as it's sent, the

rust/kernel/percpu/dynamic.rs +11 -0

index ce95e420f943..64f04cef3705 100644
@@ -3,6 +3,8 @@
 
 use super::*;
 
+use crate::cpumask::Cpumask;
+
 /// Represents a dynamic allocation of a per-CPU variable via alloc_percpu. Calls free_percpu when
 /// dropped.
 pub struct PerCpuAllocation<T>(PerCpuPtr<T>);

@@ -74,6 +76,36 @@ pub fn new_zero(flags: Flags) -> Option<Self> {
     }
 }
 
+impl<T: Clone> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable
+    ///
+    /// # Arguments
+    /// * `val` - The initial value of the per-CPU variable on all CPUs.
+    /// * `flags` - Flags used to allocate an `Arc` that keeps track of the underlying
+    ///   `PerCpuAllocation`.
+    pub fn new_with(val: T, flags: Flags) -> Option<Self> {
+        let alloc: PerCpuAllocation<T> = PerCpuAllocation::new_uninit()?;

rust/kernel/percpu/dynamic.rs pub fn new_zero(flags: Flags) -> Option<Self> { · line 90

Quoted 3 lines

rust/kernel/percpu/dynamic.rs +3 -0

+        let ptr = alloc.0;
+
+        for cpu in Cpumask::possible().iter() {

In C we've got the 'for_each_possible_cpu()'. Is there any way to preserve that semantics in rust? I really believe that similar semantics on higher level on both sides will help _a_lot_ for those transitioning into the rust world (like me).

Thanks,
Yury

Quoted 166 lines

rust/kernel/percpu/dynamic.rs +26 -0

+            // SAFETY: `ptr` is a valid allocation, and `cpu` appears in `Cpumask::possible()`
+            let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
+            // SAFETY: Each CPU's slot corresponding to `ptr` is currently uninitialized, and no
+            // one else has a reference to it. Therefore, we can freely write to it without
+            // worrying about the need to drop what was there or whether we're racing with someone
+            // else. `PerCpuPtr::get_remote_ptr` guarantees that the pointer is valid since we
+            // derived it from a valid allocation and `cpu`.
+            unsafe {
+                (*remote_ptr).write(val.clone());
+            }
+        }
+
+        let arc = Arc::new(alloc, flags).ok()?;
+
+        Some(Self { alloc: arc })
+    }
+}
+
 impl<T> PerCpu<T> for DynamicPerCpu<T> {
     unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
         // SAFETY: The requirements of `PerCpu::get_mut` and this type's invariant ensure that the

@@ -81,3 +113,11 @@ unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
         unsafe { PerCpuToken::new(guard, &self.alloc.0) }
     }
 }
+
+impl<T: InteriorMutable> CheckedPerCpu<T> for DynamicPerCpu<T> {
+    fn get(&mut self, guard: CpuGuard) -> CheckedPerCpuToken<'_, T> {
+        // SAFETY: By the invariant of `DynamicPerCpu`, the memory location in each CPU's
+        // per-CPU area corresponding to this variable has been initialized.
+        unsafe { CheckedPerCpuToken::new(guard, &self.alloc.0) }
+    }
+}

samples/rust/rust_percpu.rs +92 -7

index 98ca1c781b6b..06b322019134 100644
@@ -130,13 +130,72 @@ fn init(_module: &'static ThisModule) -> Result<Self, Error> {
 
         // SAFETY: No prerequisites for on_each_cpu.
         unsafe {
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 0);
-            on_each_cpu(Some(inc_percpu), (&raw mut test).cast(), 1);
-            on_each_cpu(Some(check_percpu), (&raw mut test).cast(), 1);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 0);
+            on_each_cpu(Some(inc_percpu_u64), (&raw mut test).cast(), 1);
+            on_each_cpu(Some(check_percpu_u64), (&raw mut test).cast(), 1);
         }
 
+        let mut checked: DynamicPerCpu<RefCell<u64>> =
+            DynamicPerCpu::new_with(RefCell::new(100), GFP_KERNEL).unwrap();
+
+        // SAFETY: No prerequisites for on_each_cpu.
+        unsafe {
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 0);
+            on_each_cpu(Some(inc_percpu_refcell_u64), (&raw mut checked).cast(), 1);
+            on_each_cpu(Some(check_percpu_refcell_u64), (&raw mut checked).cast(), 1);
+        }
+
+        checked.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+            assert!(*val.borrow() == 104);
+
+            let mut checked_native = 0;
+            *val.borrow_mut() = 0;
+
+            checked_native += 1;
+            *val.borrow_mut() += 1;
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == 1);
+
+            checked_native = checked_native.wrapping_add((-1i64) as u64);
+            val.replace_with(|old: &mut u64| old.wrapping_add((-1i64) as u64));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == 0);
+
+            checked_native = checked_native.wrapping_add((-1i64) as u64);
+            val.replace_with(|old: &mut u64| old.wrapping_add((-1i64) as u64));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == (-1i64) as u64);
+
+            checked_native = 0;
+            *val.borrow_mut() = 0;
+
+            checked_native = checked_native.wrapping_sub(1);
+            val.replace_with(|old: &mut u64| old.wrapping_sub(1));
+            pr_info!(
+                "Checked native: {}, *checked: {}\n",
+                checked_native,
+                val.borrow()
+            );
+            assert!(checked_native == *val.borrow() && checked_native == (-1i64) as u64);
+            assert!(checked_native == *val.borrow() && checked_native == u64::MAX);
+        });
+
         pr_info!("rust dynamic percpu test done\n");
 
         // Return Err to unload the module

@@ -144,7 +203,7 @@ fn init(_module: &'static ThisModule) -> Result<Self, Error> {
     }
 }
 
-extern "C" fn inc_percpu(info: *mut c_void) {
+extern "C" fn inc_percpu_u64(info: *mut c_void) {
     // SAFETY: We know that info is a void *const DynamicPerCpu<u64> and DynamicPerCpu<u64> is Send.
     let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<u64>)).clone() };
     pr_info!("Incrementing on {}\n", CpuId::current().as_u32());

@@ -153,7 +212,7 @@ extern "C" fn inc_percpu(info: *mut c_void) {
     unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val: &mut u64| *val += 1);
 }
 
-extern "C" fn check_percpu(info: *mut c_void) {
+extern "C" fn check_percpu_u64(info: *mut c_void) {
     // SAFETY: We know that info is a void *const DynamicPerCpu<u64> and DynamicPerCpu<u64> is Send.
     let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<u64>)).clone() };
     pr_info!("Asserting on {}\n", CpuId::current().as_u32());

@@ -161,3 +220,29 @@ extern "C" fn check_percpu(info: *mut c_void) {
     // SAFETY: We don't have multiple clones of pcpu in scope
     unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val: &mut u64| assert!(*val == 4));
 }
+
+extern "C" fn inc_percpu_refcell_u64(info: *mut c_void) {
+    // SAFETY: We know that info is a void *const DynamicPerCpu<RefCell<u64>> and
+    // DynamicPerCpu<RefCell<u64>> is Send.
+    let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<RefCell<u64>>)).clone() };
+    // SAFETY: smp_processor_id has no preconditions
+    pr_info!("Incrementing on {}\n", CpuId::current().as_u32());
+
+    pcpu.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+        let mut val = val.borrow_mut();
+        *val += 1;
+    });
+}
+
+extern "C" fn check_percpu_refcell_u64(info: *mut c_void) {
+    // SAFETY: We know that info is a void *const DynamicPerCpu<RefCell<u64>> and
+    // DynamicPerCpu<RefCell<u64>> is Send.
+    let mut pcpu = unsafe { (*(info as *const DynamicPerCpu<RefCell<u64>>)).clone() };
+    // SAFETY: smp_processor_id has no preconditions
+    pr_info!("Asserting on {}\n", CpuId::current().as_u32());
+
+    pcpu.get(CpuGuard::new()).with(|val: &RefCell<u64>| {
+        let val = val.borrow();
+        assert!(*val == 104);
+    });
+}

Signature · 2 lines

-- 
2.34.1

Quoted 1 line · Yury Norov

On Thu, Aug 28, 2025 at 12:00:12PM -0700, Mitchell Levy wrote:

Quoted 9 lines

Quoted 9 lines

rust/kernel/percpu/dynamic.rs +9 -0

+impl<T: Clone> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable
+    ///
+    /// # Arguments
+    /// * `val` - The initial value of the per-CPU variable on all CPUs.
+    /// * `flags` - Flags used to allocate an `Arc` that keeps track of the underlying
+    ///   `PerCpuAllocation`.
+    pub fn new_with(val: T, flags: Flags) -> Option<Self> {
+        let alloc: PerCpuAllocation<T> = PerCpuAllocation::new_uninit()?;

Quoted 8 lines

rust/kernel/percpu/dynamic.rs pub fn new_zero(flags: Flags) -> Option<Self> { · line 90

Quoted 3 lines

rust/kernel/percpu/dynamic.rs +3 -0

+        let ptr = alloc.0;
+
+        for cpu in Cpumask::possible().iter() {

In C we've got the 'for_each_possible_cpu()'. Is there any way to preserve that semantics in rust? I really believe that similar semantics on higher level on both sides will help _a_lot_ for those transitioning into the rust world (like me).

I'm not sure I understand what you mean --- I believe the semantics should be the same here (cpu takes on each value in cpu_possible_mask). Could you please clarify?

Quoted 2 lines · Mitchell Levy

Quoted 1 line · Yury Norov

On Thu, Aug 28, 2025 at 12:00:12PM -0700, Mitchell Levy wrote:

...

Quoted 18 lines

Quoted 9 lines

Quoted 9 lines

rust/kernel/percpu/dynamic.rs +9 -0

+impl<T: Clone> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable
+    ///
+    /// # Arguments
+    /// * `val` - The initial value of the per-CPU variable on all CPUs.
+    /// * `flags` - Flags used to allocate an `Arc` that keeps track of the underlying
+    ///   `PerCpuAllocation`.
+    pub fn new_with(val: T, flags: Flags) -> Option<Self> {
+        let alloc: PerCpuAllocation<T> = PerCpuAllocation::new_uninit()?;

Quoted 8 lines

rust/kernel/percpu/dynamic.rs pub fn new_zero(flags: Flags) -> Option<Self> { · line 90

Quoted 3 lines

rust/kernel/percpu/dynamic.rs +3 -0

+        let ptr = alloc.0;
+
+        for cpu in Cpumask::possible().iter() {

In C we've got the 'for_each_possible_cpu()'. Is there any way to preserve that semantics in rust? I really believe that similar semantics on higher level on both sides will help _a_lot_ for those transitioning into the rust world (like me).

Quoted 4 lines

I'm not sure I understand what you mean --- I believe the semantics should be the same here (cpu takes on each value in cpu_possible_mask). Could you please clarify?

I mean:

        for_each_possible_cpu(cpu) {
                let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
                unsafe { (*remote_ptr).write(val.clone()); }
                let arc = Arc::new(alloc, flags).ok()?;
                Some(Self { alloc: arc })
        }

Is it possible to do the above in rust?

Quoted 9 lines

        for_each_possible_cpu(cpu) {
                let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
                unsafe { (*remote_ptr).write(val.clone()); }
                let arc = Arc::new(alloc, flags).ok()?;
                Some(Self { alloc: arc })
        }

Is it possible to do the above in rust?

Ah, I see.

The syntax would be slightly different, probably something like

        use cpu::for_each_possible_cpu;

        for_each_possible_cpu(|&cpu| {
                let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
                // ...
        })

it *might* also be possible to use a macro and dispense with the need for a closure, though I'm not familiar enough with proc macros to say for sure. That would probably look like

        for_each_possible_cpu!(cpu) {
                let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
                // ...
        }

though personally I think the first one is better (simpler implementation without too much syntactic overhead, especially since closures are already used some within R4L).

it *might* also be possible to use a macro and dispense with the need for a closure, though I'm not familiar enough with proc macros to say for sure. That would probably look like

        for_each_possible_cpu!(cpu) {
                let remote_ptr = unsafe { ptr.get_remote_ptr(cpu) };
                // ...
        }

though personally I think the first one is better (simpler implementation without too much syntactic overhead, especially since closures are already used some within R4L).

+ /// # Safety

rust/kernel/percpu.rs pub fn get_ptr(&self) -> *mut MaybeUninit<T> { · line 122

Quoted 4 lines · Mitchell Levy

rust/kernel/percpu.rs +3 -0

+    /// Get a `*mut MaybeUninit<T>` to the per-CPU variable on the CPU represented by `cpu`. Note
+    /// that without some kind of synchronization, use of the returned pointer may cause a data
+    /// race. It is the caller's responsibility to use the returned pointer in a reasonable way.

Please try to make the first paragraph ("short description" / title) smaller.

Does "reasonable" mean anything different than any other raw pointer?

Quoted 1 line

+ /// # Safety

Newline after section headers (also elsewhere).

rust/kernel/percpu.rs pub fn get_ptr(&self) -> *mut MaybeUninit<T> { · line 128

Quoted 4 lines

rust/kernel/percpu.rs +4 -0

+    /// - The returned pointer is valid only if `self` is (that is, it points to a live allocation
+    ///   correctly sized and aligned to hold a `T`)
+    /// - The returned pointer is valid only if the bit corresponding to `cpu` is set in
+    ///   `Cpumask::possible()`.

It sounds like the returned pointer can be invalid without triggering UB -- could you please clarify why the method is unsafe?

In addition, please use intra-doc links wherever possible (e.g. there a was also an Arc elsewhere).

rust/kernel/percpu.rs pub fn get_ptr(&self) -> *mut MaybeUninit<T> { · line 131

Quoted 2 lines

rust/kernel/percpu.rs +2 -0

+        // SAFETY: The requirements of this function ensure this call is safe.
+        unsafe { bindings::per_cpu_ptr(self.0.cast(), cpu.as_u32()) }.cast()

Please try to explain why, not just that it is. It isn't clear how the safety preconditions, which only seem to talk about the returned pointer, make this OK.

Thanks!

Cheers,
Miguel

Will do.

Yes, this is true; strictly speaking, calling this function without dereferencing the returned pointer is always safe. I suppose I find i

it to return a valid pointer, the "safe-ness" has more to do with the functions preconditions than the act of dereferencing the pointer.

In this case, the pointer shouldn't be going very far, but I think this logic applies especially well in cases where pointers might be getting stored away for later (and so the validity of the dereference later on might rely on non-local conditions).

This flows from the first requirement (that self is a live allocation, which is necessary for per_cpu_ptr to return a valid pointer). Though, as above, simply possessing this pointer isn't UB, so it's arguable that any call to per_cpu_ptr (on x86 at least, I'm not sure how it's implemented on other arches) is always safe. Regardless, I agree this comment should be more clear (even if the function is safe, it's probably worth noting why the pointer returned is valid when the function preconditions are met); will fix.

Quoted 2 lines · Mitchell Levy

Will do.

By the way, sorry -- just to clarify, I didn't mean to remove anything, but rather split it (likely at the first full stop).

Quoted 2 lines

Yes, this is true; strictly speaking, calling this function without dereferencing the returned pointer is always safe. I suppose I find i

t> clearer that, when a function has preconditions that are necessary for

Quoted 2 lines

it to return a valid pointer, the "safe-ness" has more to do with the functions preconditions than the act of dereferencing the pointer.

In that case, I would make it safe -- we don't use safety preconditions to mark "dangerous" operations, so to speak.

Quoted 4 lines

In this case, the pointer shouldn't be going very far, but I think this logic applies especially well in cases where pointers might be getting stored away for later (and so the validity of the dereference later on might rely on non-local conditions).

But that applies to any returned raw pointer, no?

Quoted 8 lines

This flows from the first requirement (that self is a live allocation, which is necessary for per_cpu_ptr to return a valid pointer). Though, as above, simply possessing this pointer isn't UB, so it's arguable that any call to per_cpu_ptr (on x86 at least, I'm not sure how it's implemented on other arches) is always safe. Regardless, I agree this comment should be more clear (even if the function is safe, it's probably worth noting why the pointer returned is valid when the function preconditions are met); will fix.

Sounds good, thanks!

If you have cases you think other architectures may have different requirements, and you think it is likely one could enable the support for other arches and break it, then I would suggest adding a comment about it.

Or, ideally, try to see if other architectures are fine already, even if you still don't enable the code in other arches.

Cheers,
Miguel