Re: [PATCH] rust: hrtimer: introduce hrtimer support

[Boqun Feng <boqun.feng@gmail.com>]

On Tue, Apr 30, 2024 at 02:33:50PM +0200, Andreas Hindborg wrote:
[...]
> >
> > Could you see if you can replace this with a `SpinLock<bool>` +
> > `CondVar`? We shouldn't use Rust atomic in kernel now. I know it's
> > unfortunate that LKMM atomics are still work in process, but in real
> > world, you won't do busy waiting for a timer to fire, so a
> > `CondVar::wait` is better for example purpose.
> 
> Since this is only using the atomic from Rust code, it should be fine
> right? There is no mixing of memory models on this memory location.
> 

It's better compared to mixing accesses on a same location, but it's
still not allowed (for now, at least) to avoid mixing memory models on
ordering guarantees, for example:

(assume all memory location is initialized as 0)

	CPU 0				CPU 1
	-----				-----
	x.store(1, RELAXED); // Rust native atomic
	smp_store_release(&y, 1); // LKMM atomic
					let r0 = smp_load_acquire(&y);
					let r1 = x.load(RELAXED);

The smp_store_release() and smp_load_acquire() pairs per LKMM, and
provide certain rel-acq ordering. But to make it (r0 == 1 && r1 == 0),
C11 memory model needs to understand this sort of orderings, but
currently there is no such thing as an "external ordering" to C11 memory
model.

I admit this is much of a theorical concern for code reasoning, in real
world, it must "just work", but "if you want to have fun, start with
one" ;-)

Regards,
Boqun