On Fri, Jun 14, 2024 at 11:51:24AM +0200, Peter Zijlstra wrote:
> On Thu, Jun 13, 2024 at 09:30:26AM -0700, Boqun Feng wrote:
>
> > We can always add a layer on top of what we have here to provide the
> > generic `Atomic<T>`. However, I personally don't think generic
> > `Atomic<T>` is a good idea, for a few reasons:
> >
> > * I'm not sure it will bring benefits to users, the current atomic
> > users in kernel are pretty specific on the size of atomic they
> > use, so they want to directly use AtomicI32 or AtomicI64 in
> > their type definitions rather than use a `Atomic<T>` where their
> > users can provide type later.
> >
> > * I can also see the future where we have different APIs on
> > different types of atomics, for example, we could have a:
> >
> > impl AtomicI64 {
> > pub fn split(&self) -> (&AtomicI32, &AtomicI32)
> > }
> >
> > which doesn't exist for AtomicI32. Note this is not a UB because
> > we write our atomic implementation in asm, so it's perfectly
> > fine for mix-sized atomics.
> >
> > So let's start with some basic and simple until we really have a need
> > for generic `Atomic<T>`. Thoughts?
>
> Not on the generic thing, but on the lack of long. atomic_long_t is
> often used when we have pointers with extra bits on. Then you want a
> number type in order to be able to manipulate the low bits.
I mentioned my plan on AtomicPtr, but I think I should have clarified
this more. My plan is:
pub struct AtomicIsize {
#[cfg(CONFIG_64BIT)]
inner: AtomicI64
#[cfg(not(CONFIG_64BIT))]
inner: AtomicI32
}
i.e. building AtomicIsize (Rust's atomic_long_t) based on AtomicI64 and
AtomicI32. And we can a AtomicPtr type on it:
pub struct AtomicPtr<T> {
inner: AtomicIsize,
_type: PhantomData<*mut T>
}
Of course, I need to do some code generating work for AtomicIsize and
AtomicPtr, I plan to do that in Rust not in scripts, this will keep the
rust/kernel/sync/atomic/impl.rs relatively small (i.e. the Rust/C
interface is smaller). I can include this part in the next version, if
you want to see it.
Regards,
Boqun
