On Sun, Jun 16, 2024 at 10:36:13PM -0700, Boqun Feng wrote:
> On Sun, Jun 16, 2024 at 08:06:05AM -0700, Boqun Feng wrote:
> [...]
> > >
> > > Note that crossbeam's AtomicCell is also generic, and crossbeam is used
> > > by tons of crates. As Miguel mentioned, I think it's very likely that in
> > > the future we want be able to do atomics on new types (e.g. for
> > > seqlocks perhaps). We probably don't need the non-lock-free fallback of
> >
> > Good, another design bit, thank you!
> >
> > What's our overall idea on sub-word types, like Atomic<u8> and
> > Atomic<u16>, do we plan to say no to them, or they could have a limited
> > APIs? IIUC, some operations on them are relatively sub-optimal on some
> > architectures, supporting the same set of API as i32 and i64 is probably
> > a bad idea.
> >
> > Another thing in my mind is making `Atomic<T>`
> >
> > pub struct Atomic<T: Send + ...> { ... }
> >
> > so that `Atomic<T>` will always be `Sync`, because quite frankly, an
> > atomic type that cannot `Sync` is pointless.
> >
>
> Also, how do we avoid this issue [1] in kernel?
>
> `atomic_load()` in C is implemented as READ_ONCE() and it's, at most
> time, a volatile read, so the eventual code is:
>
> let a: (u8, u16) = (1, 2);
> let b = unsafe { core::ptr::read_volatile::<i32>(&a as *const _ as *const i32) };
>
^^^^ this line should really be:
let b: (u8, u16) = unsafe { transmute_copy(&read_volatile::<i32>(&a as *const _ as *const i32)) };
but you get the idea.
Regards,
Boqun
> I know we probably ignore data race here and treat `read_volatile` as a
> dependency read per LKMM [2]. But this is an using of uninitialized
> data, so it's a bit different.
>
> We can do what https://crates.io/crates/atomic does:
>
> pub struct Atomic<T: NoUninit + ..> { ... }
>
> , where `NoUinit` means no internal padding bytes, but it loses the
> ability to put a
>
> #[repr(u32)]
> pub enum Foo { .. }
>
> into `Atomic<T>`, right? Which is probably a case you want to support?
>
> Regards,
> Boqun
>
> [1]: https://github.com/crossbeam-rs/crossbeam/issues/748#issuecomment-1133926617
> [2]: tools/memory-model/Documentation/access-marking.txt
>
> > Regards,
> > Boqun
> >
> > > crossbeam's AtomicCell, but the lock-free subset with newtype support
> > > is desirable.
> > >
> > > People in general don't use the `atomic` crate because it provides no
> > > additional feature compared to the standard library. But it doesn't
> > > really mean that the standard library's atomic design is good.
> > >
> > > People decided to use AtomicT and NonZeroT instead of Atomic<T> or
> > > NonZero<T> long time ago, but many now thinks the decision was bad.
> > > Introduction of NonZero<T> is a good example of it. NonZeroT are now
> > > all type aliases of NonZero<T>.
> > >
> > > I also don't see any downside in using generics. We can provide type
> > > aliases so people can use `AtomicI32` and `AtomicI64` when they want
> > > their code to be compatible with userspace Rust can still do so.
> > >
> > > `Atomic<i32>` is also just aesthetically better than `AtomicI32` IMO.
> > > When all other types look like `NonZero<i32>`, `Wrapping<i32>`, I don't
> > > think we should have `AtomicI32` just because "it's done this way in
> > > Rust std". Our alloc already deviates a lot from Rust std.
> > >
> > > Best,
> > > Gary
