Re: [RFC] per-cpu preempt_count

[Ingo Molnar <mingo@kernel.org>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Mon, Aug 12, 2013 at 10:58 AM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > We could still have the advantages of NEED_RESCHED in preempt_count() by
> > realizing that we only rarely actually set/clear need_resched and mostly
> > read it from the highest freq user, the preempt_enable() check.
> >
> > So we could have it atomic, but do atomic_read() in the preempt_enable()
> > hotpath which wouldn't suck donkey balls, right?
> 
> Wrong. The thing is, the common case for preempt is to increment and 
> decrement the count, not testing it. Exactly because we do this for 
> spinlocks and for rcu read-locked regions.

Indeed, I should have realized that immediately ...

> Now, what we *could* do is to say:
> 
>  - we will use the high bit of the preempt count for NEED_RESCHED
> 
>  - when we set/clear that high bit, we *always* use atomic sequences, 
> and we never change any of the other bits.
> 
>  - we will increment/decrement the other counters, we *only* do so on 
> the local CPU, and we don't use atomic accesses.
> 
> Now, the downside of that is that *because* we don't use atomic accesses 
> for the inc/dec parts, the updates to the high bit can get lost. But 
> because the high bit updates are done with atomics, we know that they 
> won't mess up the actual counting bits, so at least the count is never 
> corrupted.
> 
> And the NEED_RESCHED bit getting lost would be very unusual. That 
> clearly would *not* be acceptable for RT, but it it might be acceptable 
> for "in the unusual case where we want to preempt a thread that was not 
> preemtible, *and* we ended up having the extra unsual case that 
> preemption enable ended up missing the preempt bit, we don't get 
> preempted in a timely manner". It's probably impossible to ever see in 
> practice, and considering that for non-RT use the PREEMPT bit is a 
> "strong hint" rather than anything else, it sounds like it might be 
> acceptable.
> 
> It is obviously *not* going to be acceptable for the RT people, though, 
> but since they do different code sequences _anyway_, that's not really 
> much of an issue.

Hm, this could introduce weird artifacts for code like signal delivery 
(via kick_process()), with occasional high - possibly user noticeable - 
signal delivery latencies.

But we could perhaps do something else and push the overhead into 
resched_task(): instead of using atomics we could use the resched IPI to 
set the local preempt_count(). That way preempt_count() will only ever be 
updated CPU-locally and we could merge need_resched into preempt_count() 
just fine.

[ Some care has to be taken with polling-idle threads: those could simply
  use another signalling mechanism, another field in task struct, no need
  to abuse need_resched for that. ]

We could still _read_ the preempt count non-destructively from other CPUs, 
to avoid having to send a resched IPI for already marked tasks. [ OTOH it 
might be faster to never do that and assume that an IPI has to be sent in 
99.9% of the cases - that would have to be re-measured. ]

Using this method we could have a really lightweight, minimal, percpu 
based preempt count mechanism in all the default fastpath cases, both for 
nested and for non-nested preempt_enable()s.

Thanks,

	Ingo