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 Wed, Apr 15, 2015 at 09:46:01AM +0200, Ingo Molnar wrote:
> 
>  > @@ -2088,7 +2088,7 @@ void task_numa_fault(int last_cpupid, int mem_node, int pages, int flags)
>  >  
>  >  static void reset_ptenuma_scan(struct task_struct *p)
>  >  {
>  > -	ACCESS_ONCE(p->mm->numa_scan_seq)++;
>  > +	WRITE_ONCE(p->mm->numa_scan_seq, READ_ONCE(p->mm->numa_scan_seq) + 1);
>  
> vs
> 
> 	seq = ACCESS_ONCE(p->mm->numa_scan_seq);
> 	if (p->numa_scan_seq == seq)
> 		return;
> 	p->numa_scan_seq = seq;
> 
> 
> > So the original ACCESS_ONCE() barriers were misguided to begin with: I 
> > think they tried to handle races with the scheduler balancing softirq 
> > and tried to avoid having to use atomics for the sequence counter 
> > (which would be overkill), but things like ACCESS_ONCE(x)++ never 
> > guaranteed atomicity (or even coherency) of the update.
> > 
> > But since in reality this is only statistical sampling code, all these 
> > compiler barriers can be removed I think. Peter, Mel, Rik, do you 
> > agree?
> 
> 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:

  COMPILER BARRIER
  ----------------

  The Linux kernel has an explicit compiler barrier function that  prevents the
  compiler from moving the memory accesses either side of it to the  other side:

        barrier();

  This is a general barrier -- there are no read-read or write-write variants
  of barrier().  However, ACCESS_ONCE() can be thought of as a weak form
  for barrier() that affects only the specific accesses flagged by the
  ACCESS_ONCE().

 [...]

> 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).

So what bad effects can there be from the very unlikely read and write 
tearing?

AFAICS nothing particularly bad. On the read side:

        seq = ACCESS_ONCE(p->mm->numa_scan_seq);
        if (p->numa_scan_seq == seq)
                return;
        p->numa_scan_seq = seq;

If p->mm->numa_scan_seq gets loaded twice (very unlikely), and two 
different values happen, then we might get a 'double' NUMA placement 
run - i.e. statistical noise.

On the update side:

        ACCESS_ONCE(p->mm->numa_scan_seq)++;
        p->mm->numa_scan_offset = 0;

If the compiler tears that up we might skip an update - again 
statistical noise at worst.

Nor is compiler tearing the only theoretical worry here: in theory, 
with long cache coherency latencies we might get two updates 'mixed 
up' and resulting in a (single) missed update.

Only atomics would solve all the races, but I think that would be 
overdoing it.

This is what I meant by that there's no harm from this race.

Thanks,

	Ingo