On Fri, Mar 22, 2024 at 07:57:20PM -0700, Boqun Feng wrote: > On Fri, Mar 22, 2024 at 10:33:13PM -0400, Kent Overstreet wrote: > > On Fri, Mar 22, 2024 at 07:26:28PM -0700, Boqun Feng wrote: > > > On Fri, Mar 22, 2024 at 10:07:31PM -0400, Kent Overstreet wrote: > > > [...] > > > > > Boqun already mentioned the "mixing access sizes", which is actually > > > > > quite fundamental in the kernel, where we play lots of games with that > > > > > (typically around locking, where you find patterns line unlock writing > > > > > a zero to a single byte, even though the whole lock data structure is > > > > > a word). And sometimes the access size games are very explicit (eg > > > > > lib/lockref.c). > > > > > > > > I don't think mixing access sizes should be a real barrier. On the read > > > > > > Well, it actually is, since mixing access sizes is, guess what, > > > an undefined behavior: > > > > > > (example in https://doc.rust-lang.org/std/sync/atomic/#memory-model-for-atomic-accesses) > > > > > > thread::scope(|s| { > > > // This is UB: using different-sized atomic accesses to the same data > > > s.spawn(|| atomic.store(1, Ordering::Relaxed)); > > > s.spawn(|| unsafe { > > > let differently_sized = transmute::<&AtomicU16, &AtomicU8>(&atomic); > > > differently_sized.store(2, Ordering::Relaxed); > > > }); > > > }); > > > > > > Of course, you can say "I will just ignore the UB", but if you have to > > > ignore "compiler rules" to make your code work, why bother use compiler > > > builtin in the first place? Being UB means they are NOT guaranteed to > > > work. > > > > That's not what I'm proposing - you'd need additional compiler support. > > Ah, OK. > > > but the new intrinsic would be no different, semantics wise for the > > compiler to model, than a "lock orb". > > Be ready to be disappointed: > > https://rust-lang.zulipchat.com/#narrow/stream/136281-t-opsem/topic/is.20atomic.20aliasing.20allowed.3F/near/402078545 > https://rust-lang.zulipchat.com/#narrow/stream/136281-t-opsem/topic/is.20atomic.20aliasing.20allowed.3F/near/402082631 > > ;-) > > In fact, if you get a chance to read the previous discussion links I > shared, you will find I was just like you in the beginning: hope we > could extend the model to support more kernel code properly. But my > overall feeling is that it's either very challenging or lack of > motivation to do. That's casting - that doesn't work because compiler people hate aliasing. But intrinsics for e.g. __atomic32_read_u8(atomic_u32_t *a, unsigned byte) __atomic32_write_u8(atomic_u32_t a*, unsigned byte) should be doable - that's perfectly fine for the compiler to model. That would admittedly be ugly to use. But, if Rust ever allowed for marking any struct up to word size as atomic (which we want anyways...), it could use that under the hood for setting a member variable without cmpxchg.
