Re: [PATCH v7 6/9] rust: sync: atomic: Add the framework of arithmetic operations

[Boqun Feng]

On Wed, Jul 16, 2025 at 07:16:02PM +0200, Benno Lossin wrote:
> On Wed Jul 16, 2025 at 5:48 PM CEST, Boqun Feng wrote:
> > On Wed, Jul 16, 2025 at 05:36:05PM +0200, Benno Lossin wrote:
> > [..]
> >> >
> >> > I have a better solution:
> >> >
> >> > in ops.rs
> >> >
> >> >     pub struct AtomicRepr<T: AtomicImpl>(UnsafeCell<T>)
> >> >
> >> >     impl AtomicArithmeticOps for i32 {
> >> >         // a *safe* function
> >> >         fn atomic_add(a: &AtomicRepr, v: i32) {
> >> > 	    ...
> >> > 	}
> >> >     }
> >> >
> >> > in generic.rs
> >> >
> >> >     pub struct Atomic<T>(AtoimcRepr<T::Repr>);
> >> >
> >> >     impl<T: AtomicAdd> Atomic<T> {
> >> >         fn add(&self, v: .., ...) {
> >> > 	    T::Repr::atomic_add(&self.0, ...);
> >> > 	}
> >> >     }
> >> >
> >> > see:
> >> >
> >> > 	https://git.kernel.org/pub/scm/linux/kernel/git/boqun/linux.git/log/?h=rust-atomic-impl
> >> 
> >> Hmm what does the additional indirection give you?
> >> 
> >
> > What additional indirection you mean? You cannot make atomic_add() safe
> > with only `UnsafeCell<T::Repr>`.
> 
> What is the advantage of making it safe? It just moves the safety

Well, first we in general are in favor of safe functions when we can
make it safe, right? Second, at Atomic<T> level, the major unsafe stuff
comes from the T <-> T::Repr transmutable and making `Atomic<T>` a valid
`T`, the safety of i{32, 64}::atomic_add() would be a bit distraction
when implementing Atomic::add(). With i{32, 64}::atomic_add() being
safe, I can implementation Atomic::add() as:

impl<T: ..> Atomic<T> {
    #[inline(always)]
    pub fn add<Rhs>(&self, v: Rhs, _: ordering::Relaxed)
    where
        T: AtomicAdd<Rhs>,
    {
        let v = T::rhs_into_delta(v);

        // INVARIANT: `self.0` is a valid `T` due to safety requirement of `AtomicAdd`.
        T::Repr::atomic_add(&self.0, v);
    }
}

then all the safety related comments will be focused on why `self.0` is
still a valid `T` after the operation (if you want to be extra detailed
about it, it's fine, and can be done easily)

> comments into `ops.rs` which makes it harder to read due to the macros.

Does it? Add `i32` and `i64` level, you only need the pointer to be a
valid `* mut i{32, 64}`. So the following is pretty clear to me.

    /// Atomic arithmetic operations
    pub trait AtomicArithmeticOps {
        /// Atomic add (wrapping).
        ///
        /// Atomically updates `*a` to `(*a).wrapping_add(v)`.
        fn add[](a: &AtomicRepr<Self>, v: Self::Delta) {
            // SAFETY: `a.as_ptr()` is valid and properly aligned.
            bindings::#call(v, a.as_ptr().cast())
        }
    }

it is at least better than:

            $(
                $(#[$($p_attr)*])*
                $pvis unsafe fn $p_field<E>(
                    self,
                    slot: *mut $p_type,
                    init: impl $crate::PinInit<$p_type, E>,
                ) -> ::core::result::Result<(), E> {
                    // SAFETY: TODO.
                    unsafe { $crate::PinInit::__pinned_init(init, slot) }
                }
            )*

;-)

Regards,
Boqun

> 
> ---
> Cheers,
> Benno