Re: [PATCH 1/3] sched, timer: Remove usages of ACCESS_ONCE in the scheduler

[Ingo Molnar <mingo@kernel.org>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> On Thu, Apr 16, 2015 at 08:02:27PM +0200, Ingo Molnar wrote:
> > > ACCESS_ONCE() is not a compiler barrier
> > 
> > It's not a general compiler barrier (and I didn't claim so) but it is 
> > still a compiler barrier: it's documented as a weak, variable specific 
> > barrier in Documentation/memor-barriers.txt:
> 
> Ok, fair enough. I just don't generally think of them as 'barriers'.
> 
> > > The 'read' side uses ACCESS_ONCE() for two purposes:
> > >  - to load the value once, we don't want the seq number to change under
> > >    us for obvious reasons
> > >  - to avoid load tearing and observe weird seq numbers
> > > 
> > > The update side uses ACCESS_ONCE() to avoid write tearing, and 
> > > strictly speaking it should also worry about read-tearing since its 
> > > not hard serialized, although its very unlikely to actually have 
> > > concurrency (IIRC).
> 
> > This is what I meant by that there's no harm from this race.
> 
> Ok, but I would still like to preserve the READ one on the usage 
> site and the WRITE one on the update side, if only as documentation 
> that there's something 'special' happening.
> 
> And while the effects here might end up being statistical noise, I 
> have conceptual problems with re-reading seq counts, that's not 
> proper.

Yes ... but that still leaves this weird feeling that it's really 
still a bit wrong because it's not proper parallel code, we just 
reduced the probability of the remaining races radically. And it's not 
like GCC (or any compiler) does load tearing or even store tearing 
under normal -O2 for such code patterns, right?

> 
> And its not like they really cost anything.

That's true.

Would it make sense to add a few comments to the seq field definition 
site(s), about how it's supposed to be accessed - or to the 
READ_ONCE()/WRITE_ONCE() sites, to keep people from wondering?

Thanks,

	Ingo