public inbox for linux-kernel@vger.kernel.org
 help / color / mirror / Atom feed
From: Mitchell Levy <levymitchell0@gmail.com>
To: "Miguel Ojeda" <ojeda@kernel.org>,
	"Alex Gaynor" <alex.gaynor@gmail.com>,
	"Gary Guo" <gary@garyguo.net>,
	"Björn Roy Baron" <bjorn3_gh@protonmail.com>,
	"Andreas Hindborg" <a.hindborg@kernel.org>,
	"Alice Ryhl" <aliceryhl@google.com>,
	"Trevor Gross" <tmgross@umich.edu>,
	"Andrew Morton" <akpm@linux-foundation.org>,
	"Dennis Zhou" <dennis@kernel.org>, "Tejun Heo" <tj@kernel.org>,
	"Christoph Lameter" <cl@linux.com>,
	"Danilo Krummrich" <dakr@kernel.org>,
	"Benno Lossin" <lossin@kernel.org>,
	"Yury Norov" <yury.norov@gmail.com>,
	"Viresh Kumar" <viresh.kumar@linaro.org>,
	"Boqun Feng" <boqun@kernel.org>
Cc: Tyler Hicks <code@tyhicks.com>,
	Allen Pais <apais@linux.microsoft.com>,
	 linux-kernel@vger.kernel.org, rust-for-linux@vger.kernel.org,
	 linux-mm@kvack.org, Mitchell Levy <levymitchell0@gmail.com>
Subject: [PATCH v5 5/8] rust: percpu: introduce a rust API for dynamic per-CPU variables
Date: Fri, 10 Apr 2026 14:35:35 -0700	[thread overview]
Message-ID: <20260410-rust-percpu-v5-5-4292380d7a41@gmail.com> (raw)
In-Reply-To: <20260410-rust-percpu-v5-0-4292380d7a41@gmail.com>

Dynamically allocated per-CPU variables are core to many of the
use-cases of per-CPU variables (e.g., ref counting). Add support for
them using the core `PerCpuPtr<T>` primitive, implementing the
`PerCpu<T>` and `CheckedPerCpu<T>` traits.

Co-developed-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Boqun Feng <boqun@kernel.org>
Signed-off-by: Mitchell Levy <levymitchell0@gmail.com>
---
 rust/helpers/percpu.c         |  12 +++
 rust/kernel/percpu.rs         |  30 +++++-
 rust/kernel/percpu/dynamic.rs | 217 ++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 255 insertions(+), 4 deletions(-)

diff --git a/rust/helpers/percpu.c b/rust/helpers/percpu.c
index 463186b8af9d..3b2f69a96c66 100644
--- a/rust/helpers/percpu.c
+++ b/rust/helpers/percpu.c
@@ -7,3 +7,15 @@ void __percpu *rust_helper_alloc_percpu(size_t sz, size_t align)
 {
 	return __alloc_percpu(sz, align);
 }
+
+__rust_helper
+void *rust_helper_per_cpu_ptr(void __percpu *ptr, unsigned int cpu)
+{
+	return per_cpu_ptr(ptr, cpu);
+}
+
+__rust_helper
+void rust_helper_on_each_cpu(smp_call_func_t func, void *info, int wait)
+{
+	on_each_cpu(func, info, wait);
+}
diff --git a/rust/kernel/percpu.rs b/rust/kernel/percpu.rs
index 615859d4daad..0ec1245038bb 100644
--- a/rust/kernel/percpu.rs
+++ b/rust/kernel/percpu.rs
@@ -1,15 +1,20 @@
 // SPDX-License-Identifier: GPL-2.0
 //! Per-CPU variables.
 //!
-//! See the [`crate::define_per_cpu!`] macro and the [`PerCpu<T>`] trait.
+//! See the [`crate::define_per_cpu!`] macro, the [`DynamicPerCpu`] type, and the [`PerCpu<T>`]
+//! trait.
 
 pub mod cpu_guard;
+mod dynamic;
 mod static_;
 
+#[doc(inline)]
+pub use dynamic::*;
 #[doc(inline)]
 pub use static_::*;
 
 use crate::{
+    cpu::CpuId,
     declare_extern_per_cpu,
     percpu::cpu_guard::CpuGuard,
     types::Opaque, //
@@ -132,6 +137,23 @@ 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>`](MaybeUninit) 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.
+    ///
+    /// # Returns
+    /// - 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
+    ///   [`kernel::cpumask::Cpumask::possible_cpus()`].
+    pub fn get_remote_ptr(&self, cpu: CpuId) -> *mut MaybeUninit<T> {
+        // SAFETY: `bindings::per_cpu_ptr` is just doing pointer arithmetic. The returned pointer
+        // may not be valid (under the conditions specified in this function's documentation), but
+        // the act of producing the pointer 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
@@ -155,9 +177,9 @@ impl<T> Copy for PerCpuPtr<T> {}
 
 /// A trait representing a per-CPU variable.
 ///
-/// This is implemented for [`StaticPerCpu<T>`]. The main usage of this trait is to call
-/// [`Self::get_mut`] to get a [`PerCpuToken`] that can be used to access the underlying per-CPU
-/// variable.
+/// This is implemented for both [`StaticPerCpu<T>`] and [`DynamicPerCpu<T>`]. The main usage of
+/// this trait is to call [`Self::get_mut`] to get a [`PerCpuToken`] that can be used to access the
+/// underlying per-CPU variable.
 ///
 /// See [`PerCpuToken::with`].
 pub trait PerCpu<T> {
diff --git a/rust/kernel/percpu/dynamic.rs b/rust/kernel/percpu/dynamic.rs
new file mode 100644
index 000000000000..40514704b3d0
--- /dev/null
+++ b/rust/kernel/percpu/dynamic.rs
@@ -0,0 +1,217 @@
+// SPDX-License-Identifier: GPL-2.0
+//! Dynamically allocated per-CPU variables.
+
+use super::*;
+
+use crate::{
+    alloc::Flags,
+    bindings::{
+        alloc_percpu,
+        free_percpu, //
+    },
+    cpumask::Cpumask,
+    prelude::*,
+    sync::Arc, //
+};
+
+use core::mem::{
+    align_of,
+    size_of,
+    MaybeUninit, //
+};
+
+/// Represents a dynamic allocation of a per-CPU variable via `alloc_percpu`. Calls `free_percpu`
+/// when dropped.
+///
+/// # Contents
+/// Note that the allocated memory need not be initialized, and this type does not track when/if
+/// the memory location on any particular CPU has been initialized. This means that it cannot tell
+/// whether it should drop the *contents* of the allocation when it is dropped. It is up to the
+/// user to do this via something like [`core::ptr::drop_in_place`].
+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`. The pointer is sized/aligned for a
+            // `T`, and `MaybeUninit<T>` is `#[repr(transparent)]` (and thus same size/align), so
+            // we can cast this `*mut c_void` 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::new(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`. The pointer is sized/aligned for a
+            // `T`, and `MaybeUninit<T>` is `#[repr(transparent)]` (and thus same size/align), so
+            // we can cast this `*mut c_void` to it.
+            unsafe { alloc_percpu(size_of::<T>(), align_of::<T>()) }.cast();
+        if ptr.is_null() {
+            return None;
+        }
+
+        Some(Self(PerCpuPtr::new(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.
+///
+/// # Examples
+/// ```
+/// # use kernel::prelude::*;
+/// # use kernel::percpu::*;
+/// # use kernel::percpu::cpu_guard::CpuGuard;
+///
+/// let mut pcpu: DynamicPerCpu<u32> =
+///   DynamicPerCpu::new_zero(GFP_KERNEL).expect("failed to allocate per-CPU variable");
+/// {
+///   let _guard = CpuGuard::new();
+///   // SAFETY: No other `pcpu` reference
+///   unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val| *val = 42);
+///   // SAFETY: No other `pcpu` reference
+///   unsafe { pcpu.get_mut(CpuGuard::new()) }.with(|val| assert!(*val == 42));
+/// }
+/// ```
+#[derive(Clone)]
+pub struct DynamicPerCpu<T> {
+    // INVARIANT: `alloc` is `Some` unless this object is in the process of being dropped.
+    // INVARIANT: The allocation held by `alloc` is sized and aligned for a `T`.
+    // INVARIANT: The memory location in each CPU's per-CPU area pointed at by the alloc is
+    // initialized.
+    alloc: Option<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: Some(arc) })
+    }
+}
+
+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> {
+        Self::new_from(|_| val.clone(), flags)
+    }
+}
+
+impl<T> DynamicPerCpu<T> {
+    /// Allocates a new per-CPU variable
+    ///
+    /// # Arguments
+    /// * `initer` - A function that takes a `CpuId` and returns the initial value of the per-CPU
+    ///   variable on the corresponding CPU. Called once for each possible CPU (i.e.,
+    ///   `Cpumask::possible_cpus().weight()` times).
+    /// * `flags` - Flags used to allocate an [`Arc`] that keeps track of the underlying
+    ///   [`PerCpuAllocation`].
+    pub fn new_from(mut initer: impl FnMut(CpuId) -> T, flags: Flags) -> Option<Self> {
+        let alloc: PerCpuAllocation<T> = PerCpuAllocation::new_uninit()?;
+        let ptr = alloc.0;
+
+        let arc = Arc::new(alloc, flags).ok()?;
+
+        for cpu in Cpumask::possible_cpus().iter() {
+            let remote_ptr = ptr.get_remote_ptr(cpu);
+            // SAFETY: `remote_ptr` is valid because `ptr` points to a live allocation and `cpu`
+            // appears in `Cpumask::possible_cpus()`.
+            //
+            // 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.
+            unsafe {
+                (*remote_ptr).write(initer(cpu));
+            }
+        }
+
+        Some(Self { alloc: Some(arc) })
+    }
+}
+
+impl<T> PerCpu<T> for DynamicPerCpu<T> {
+    unsafe fn get_mut(&mut self, guard: CpuGuard) -> PerCpuToken<'_, T> {
+        // SAFETY:
+        // 1. Invariants of this type assure that `alloc` is `Some`.
+        // 2. The requirements of `PerCpu::get_mut` ensure that no other `[Checked]PerCpuToken`
+        //    exists on the current CPU.
+        // 3. The invariants of `DynamicPerCpu` ensure that the contents of the allocation are
+        //    initialized on each CPU.
+        // 4. The existence of a reference to the `PerCpuAllocation` ensures that the allocation is
+        //    live.
+        // 5. The invariants of `DynamicPerCpu` ensure that the allocation is sized and aligned for
+        //    a `T`.
+        unsafe { PerCpuToken::new(guard, &self.alloc.as_ref().unwrap_unchecked().0) }
+    }
+}
+
+impl<T: InteriorMutable> CheckedPerCpu<T> for DynamicPerCpu<T> {
+    fn get(&self, guard: CpuGuard) -> CheckedPerCpuToken<'_, T> {
+        // SAFETY:
+        // 1. Invariants of this type assure that `alloc` is `Some`.
+        // 2. The invariants of `DynamicPerCpu` ensure that the contents of the allocation are
+        //    initialized on each CPU.
+        // 3. The existence of a reference to the `PerCpuAllocation` ensures that the allocation is
+        //    live.
+        // 4. The invariants of `DynamicPerCpu` ensure that the allocation is sized and aligned for
+        //    a `T`.
+        unsafe { CheckedPerCpuToken::new(guard, &self.alloc.as_ref().unwrap_unchecked().0) }
+    }
+}
+
+impl<T> Drop for DynamicPerCpu<T> {
+    fn drop(&mut self) {
+        // SAFETY: This type's invariant ensures that `self.alloc` is `Some`.
+        let alloc = unsafe { self.alloc.take().unwrap_unchecked() };
+        if let Some(unique_alloc) = Arc::into_unique_or_drop(alloc) {
+            let ptr = unique_alloc.0;
+            for cpu in Cpumask::possible_cpus().iter() {
+                let remote_ptr = ptr.get_remote_ptr(cpu);
+                // SAFETY: `remote_ptr` is valid because the allocation it points to is still live,
+                // `cpu` appears in `Cpumask::possible_cpus()`, and the original allocation was
+                // sized and aligned for a `T`.
+                //
+                // This type's invariant ensures that the memory location in each CPU's per-CPU
+                // area pointed at by `alloc.0` has been initialized. We have a `UniqueArc`, so we
+                // know we're the only ones with a reference to the memory. These two facts
+                // together satisfy the requirements for `assume_init_drop`.
+                unsafe {
+                    (*remote_ptr).assume_init_drop();
+                }
+            }
+        }
+    }
+}

-- 
2.34.1


  parent reply	other threads:[~2026-04-10 21:36 UTC|newest]

Thread overview: 10+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-04-10 21:35 [PATCH v5 0/8] rust: Add Per-CPU Variable API Mitchell Levy
2026-04-10 21:35 ` [PATCH v5 1/8] rust: cpumask: Add a `Cpumask` iterator Mitchell Levy
2026-04-10 21:35 ` [PATCH v5 2/8] rust: cpumask: Add getters for globally defined cpumasks Mitchell Levy
2026-04-10 21:35 ` [PATCH v5 3/8] rust: percpu: Add C bindings for per-CPU variable API Mitchell Levy
2026-04-10 21:35 ` [PATCH v5 4/8] rust: percpu: introduce a rust API for static per-CPU variables Mitchell Levy
2026-04-10 21:35 ` Mitchell Levy [this message]
2026-04-10 21:35 ` [PATCH v5 6/8] rust: percpu: add a rust per-CPU variable sample Mitchell Levy
2026-04-10 21:35 ` [PATCH v5 7/8] rust: percpu: Add pin-hole optimizations for numerics Mitchell Levy
2026-04-11  3:06   ` Yury Norov
2026-04-10 21:35 ` [PATCH v5 8/8] rust: percpu: cache per-CPU pointers in the dynamic case Mitchell Levy

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20260410-rust-percpu-v5-5-4292380d7a41@gmail.com \
    --to=levymitchell0@gmail.com \
    --cc=a.hindborg@kernel.org \
    --cc=akpm@linux-foundation.org \
    --cc=alex.gaynor@gmail.com \
    --cc=aliceryhl@google.com \
    --cc=apais@linux.microsoft.com \
    --cc=bjorn3_gh@protonmail.com \
    --cc=boqun@kernel.org \
    --cc=cl@linux.com \
    --cc=code@tyhicks.com \
    --cc=dakr@kernel.org \
    --cc=dennis@kernel.org \
    --cc=gary@garyguo.net \
    --cc=linux-kernel@vger.kernel.org \
    --cc=linux-mm@kvack.org \
    --cc=lossin@kernel.org \
    --cc=ojeda@kernel.org \
    --cc=rust-for-linux@vger.kernel.org \
    --cc=tj@kernel.org \
    --cc=tmgross@umich.edu \
    --cc=viresh.kumar@linaro.org \
    --cc=yury.norov@gmail.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox