Re: [PATCH v2 1/2] rust: sync: atomic: Add perfromance-optimal Flag type for atomic booleans

["Gary Guo" <gary@garyguo.net>]

On Thu Jan 29, 2026 at 3:33 PM GMT, Boqun Feng wrote:
> On Thu, Jan 29, 2026 at 02:15:10PM +0000, Gary Guo wrote:
>> On Thu Jan 29, 2026 at 12:26 PM GMT, FUJITA Tomonori wrote:
>> > From: FUJITA Tomonori <fujita.tomonori@gmail.com>
>> >
>> > Add AtomicFlag type for boolean flags.
>> >
>> > Document when AtomicFlag is generally preferable to Atomic<bool>: in
>> > particular, when RMW operations such as xchg()/cmpxchg() may be used
>> > and minimizing memory usage is not the top priority. On some
>> > architectures without byte-sized RMW instructions, Atomic<bool> can be
>> > slower for RMW operations.
>> >
>> > Signed-off-by: FUJITA Tomonori <fujita.tomonori@gmail.com>
>> 
>> Hi Fujita,
>> 
>> Thanks for the patch. I think this looks nice, so from design point of view:
>> 
>> Reviewed-by: Gary Guo <gary@garyguo.net>
>> 
>> However, Boqun reported that the codegen of `.bool_field` may involve a bit
>> masking instruction.
>> 
>
> Yeah, but at the moment, I haven't found any elegant way to reduce that
> see [1], plus I've tried to transmute the 32-bit Flag struct into a
> 32-bit enum, but for example on riscv64 an `sext.w` instruction is still
> generated [2]. That's a sign to me that the micro-optimization here may
> not bring actual performance gain. But of course, open to any
> improvement, let's ship what we have now and improve the codegen later.
>
> [1]: https://rust-for-linux.zulipchat.com/#narrow/channel/288089-General/topic/A.20.60AlwaysZero.60.20type.20for.20padding.3F/near/570631532
> [2]: https://godbolt.org/z/3PMK3EK1r

Interesting! In this case, disabling MIR optimization generates better code for
test2 (-Zmir-opt-level=0). Although, it still has an `andi a0, a0, 1` remaining.

Testing with `-Cno-prepopulate-passes --emit=llvm-ir` it looks like Rust is not
telling LLVM about the fact that `v` can only be 0 or 1... Although, I recall
that previously seeing LLVM codegen issues when Rust does give LLVM additional
unreachable paths.. So the fix isn't going to be straightforward.

With this background I agree we should ship this as is. It's much better than a
LL/SC loop anyway.

Best,
Gary
>
>> 
>> > ---
>> >  rust/kernel/sync/atomic.rs           | 125 +++++++++++++++++++++++++++
>> >  rust/kernel/sync/atomic/predefine.rs |  17 ++++
>> >  2 files changed, 142 insertions(+)
>> >
>> > diff --git a/rust/kernel/sync/atomic.rs b/rust/kernel/sync/atomic.rs
>> > index 4aebeacb961a..bfc393d98aa9 100644
>> > --- a/rust/kernel/sync/atomic.rs
>> > +++ b/rust/kernel/sync/atomic.rs
>> > @@ -560,3 +560,128 @@ pub fn fetch_add<Rhs, Ordering: ordering::Ordering>(&self, v: Rhs, _: Ordering)
>> >          unsafe { from_repr(ret) }
>> >      }
>> >  }
>> > +
>> > +#[cfg(any(CONFIG_X86_64, CONFIG_UML, CONFIG_ARM, CONFIG_ARM64))]
>> > +#[repr(C)]
>> > +#[derive(Clone, Copy)]
>> > +struct Flag {
>> > +    bool_field: bool,
>> > +}
>> > +
>> > +/// # Invariants
>> > +///
>> > +/// `padding` must be all zeroes.
>> > +#[cfg(not(any(CONFIG_X86_64, CONFIG_UML, CONFIG_ARM, CONFIG_ARM64)))]
>> > +#[repr(C, align(4))]
>> > +#[derive(Clone, Copy)]
>> > +struct Flag {
>> > +    #[cfg(target_endian = "big")]
>> > +    padding: [u8; 3],
>> > +    bool_field: bool,
>> > +    #[cfg(target_endian = "little")]
>> > +    padding: [u8; 3],
>> > +}
>> > +
>> > +impl Flag {
>> > +    #[inline(always)]
>> > +    const fn new(b: bool) -> Self {
>> > +        // INVARIANT: `padding` is all zeroes.
>> > +        Self {
>> > +            bool_field: b,
>> > +            #[cfg(not(any(CONFIG_X86_64, CONFIG_UML, CONFIG_ARM, CONFIG_ARM64)))]
>> > +            padding: [0; 3],
>> > +        }
>> > +    }
>> > +}
>> > +
>> > +// SAFETY: `Flag` and `Repr` have the same size and alignment, and `Flag` is round-trip
>> > +// transmutable to the selected representation (`i8` or `i32`).
>> > +unsafe impl AtomicType for Flag {
>> > +    #[cfg(any(CONFIG_X86_64, CONFIG_UML, CONFIG_ARM, CONFIG_ARM64))]
>> > +    type Repr = i8;
>> > +    #[cfg(not(any(CONFIG_X86_64, CONFIG_UML, CONFIG_ARM, CONFIG_ARM64)))]
>> > +    type Repr = i32;
>> > +}
>> > +
>> > +/// An atomic flag type intended to be backed by performance-optimal integer type.
>> > +///
>> > +/// The backing integer type is an implementation detail; it may vary by architecture and change
>> > +/// in the future.
>> > +///
>> > +/// [`AtomicFlag`] is generally preferable to [`Atomic<bool>`] when you need read-modify-write
>> > +/// (RMW) operations (e.g. [`Atomic::xchg()`]/[`Atomic::cmpxchg()`]) or when [`Atomic<bool>`] does
>> > +/// not save memory due to padding. On some architectures that do not support byte-sized atomic
>> > +/// RMW operations, RMW operations on [`Atomic<bool>`] are slower.
>> > +///
>> > +/// If you only use [`Atomic::load()`]/[`Atomic::store()`], [`Atomic<bool>`] is fine.
>> > +///
>> > +/// # Examples
>> > +///
>> > +/// ```
>> > +/// use kernel::sync::atomic::{AtomicFlag, Relaxed};
>> > +///
>> > +/// let flag = AtomicFlag::new(false);
>> > +/// assert_eq!(false, flag.load(Relaxed));
>> > +/// flag.store(true, Relaxed);
>> > +/// assert_eq!(true, flag.load(Relaxed));
>> > +/// ```
>> > +pub struct AtomicFlag(Atomic<Flag>);
>> > +
>> > +impl AtomicFlag {
>> > +    /// Creates a new atomic flag.
>> > +    #[inline(always)]
>> > +    pub const fn new(b: bool) -> Self {
>> > +        Self(Atomic::new(Flag::new(b)))
>> > +    }
>> > +
>> > +    /// Returns a mutable reference to the underlying flag as a [`bool`].
>> > +    ///
>> > +    /// This is safe because the mutable reference of the atomic flag guarantees exclusive access.
>> > +    ///
>> > +    /// # Examples
>> > +    ///
>> > +    /// ```
>> > +    /// use kernel::sync::atomic::{AtomicFlag, Relaxed};
>> > +    ///
>> > +    /// let mut atomic_flag = AtomicFlag::new(false);
>> > +    /// assert_eq!(false, atomic_flag.load(Relaxed));
>> > +    /// *atomic_flag.get_mut() = true;
>> > +    /// assert_eq!(true, atomic_flag.load(Relaxed));
>> > +    /// ```
>> > +    #[inline(always)]
>> > +    pub fn get_mut(&mut self) -> &mut bool {
>> > +        &mut self.0.get_mut().bool_field
>> > +    }
>> > +
>> > +    /// Loads the value from the atomic flag.
>> > +    #[inline(always)]
>> > +    pub fn load<Ordering: ordering::AcquireOrRelaxed>(&self, o: Ordering) -> bool {
>> > +        self.0.load(o).bool_field
>> > +    }
>> > +
>> > +    /// Stores a value to the atomic flag.
>> > +    #[inline(always)]
>> > +    pub fn store<Ordering: ordering::ReleaseOrRelaxed>(&self, v: bool, o: Ordering) {
>> > +        self.0.store(Flag::new(v), o);
>> > +    }
>> > +
>> > +    /// Stores a value to the atomic flag and returns the previous value.
>> > +    #[inline(always)]
>> > +    pub fn xchg<Ordering: ordering::Ordering>(&self, new: bool, o: Ordering) -> bool {
>> > +        self.0.xchg(Flag::new(new), o).bool_field
>> > +    }
>> > +
>> > +    /// Store a value to the atomic flag if the current value is equal to `old`.
>> > +    #[inline(always)]
>> > +    pub fn cmpxchg<Ordering: ordering::Ordering>(
>> > +        &self,
>> > +        old: bool,
>> > +        new: bool,
>> > +        o: Ordering,
>> > +    ) -> Result<bool, bool> {
>> > +        match self.0.cmpxchg(Flag::new(old), Flag::new(new), o) {
>> > +            Ok(_) => Ok(old),
>> > +            Err(f) => Err(f.bool_field),
>> > +        }
>> > +    }
>> > +}
>> > diff --git a/rust/kernel/sync/atomic/predefine.rs b/rust/kernel/sync/atomic/predefine.rs
>> > index 42067c6a266c..d14e10544dcf 100644
>> > --- a/rust/kernel/sync/atomic/predefine.rs
>> > +++ b/rust/kernel/sync/atomic/predefine.rs
>> > @@ -215,4 +215,21 @@ fn atomic_bool_tests() {
>> >          assert_eq!(false, x.load(Relaxed));
>> >          assert_eq!(Ok(false), x.cmpxchg(false, true, Full));
>> >      }
>> > +
>> > +    #[test]
>> > +    fn atomic_flag_tests() {
>> > +        let mut flag = AtomicFlag::new(false);
>> > +
>> > +        assert_eq!(false, flag.load(Relaxed));
>> > +
>> > +        *flag.get_mut() = true;
>> > +        assert_eq!(true, flag.load(Relaxed));
>> > +
>> > +        assert_eq!(true, flag.xchg(false, Relaxed));
>> > +        assert_eq!(false, flag.load(Relaxed));
>> > +
>> > +        *flag.get_mut() = true;
>> > +        assert_eq!(Ok(true), flag.cmpxchg(true, false, Full));
>> > +        assert_eq!(false, flag.load(Relaxed));
>> > +    }
>> >  }
>>