Re: [PATCH -v9][RFC] mutex: implement adaptive spinning

[Chris Mason <chris.mason@oracle.com>]

On Wed, 2009-01-14 at 12:18 +0100, Dmitry Adamushko wrote:
> 2009/1/14 Chris Mason <chris.mason@oracle.com>:
> > On Tue, 2009-01-13 at 18:21 +0100, Peter Zijlstra wrote:
> >> On Tue, 2009-01-13 at 08:49 -0800, Linus Torvalds wrote:
> >> >
> >> > So do a v10, and ask people to test.
> >>
> >> ---
> >> Subject: mutex: implement adaptive spinning
> >> From: Peter Zijlstra <a.p.zijlstra@chello.nl>
> >> Date: Mon Jan 12 14:01:47 CET 2009
> >>
> >> Change mutex contention behaviour such that it will sometimes busy wait on
> >> acquisition - moving its behaviour closer to that of spinlocks.
> >>
> >
> > I've spent a bunch of time on this one, and noticed earlier today that I
> > still had bits of CONFIG_FTRACE compiling.  I wasn't actually tracing
> > anything, but it seems to have had a big performance hit.
> >
> > The bad news is the simple spin got much much faster, dbench 50 coming
> > in at 1282MB/s instead of 580MB/s.  (other benchmarks give similar
> > results)
> >
> > v10 is better that not spinning, but its in the 5-10% range.  So, I've
> > been trying to find ways to close the gap, just to understand exactly
> > where it is different.
> >
> > If I take out:
> >        /*
> >         * If there are pending waiters, join them.
> >         */
> >        if (!list_empty(&lock->wait_list))
> >                break;
> >
> >
> > v10 pops dbench 50 up to 1800MB/s.  The other tests soundly beat my
> > spinning and aren't less fair.  But clearly this isn't a good solution.
> >
> > I tried a few variations, like only checking the wait list once before
> > looping, which helps some.  Are there other suggestions on better tuning
> > options?
> 
> (some thoughts/speculations)
> 
> Perhaps for highly-contanded mutexes the spinning implementation may
> quickly degrade [*] to the non-spinning one (i.e. the current
> sleep-wait mutex) and then just stay in this state until a moment of
> time when there are no waiters  [**]  -- i.e.
> list_empty(&lock->wait_list) == 1 and waiters can start spinning
> again.

It is actually ok if the highly contention mutexes don't degrade as long
as they are highly contended and the holder isn't likely to schedule.

> 
> what may trigger [*]:
> 
> (1) obviously, an owner scheduling out.
> 
> Even if it happens rarely (otherwise, it's not a target scenario for
> our optimization), due to the [**] it may take quite some time until
> waiters are able to spin again.
> 
> let's say, waiters (almost) never block (and possibly, such cases
> would be better off just using a spinlock after some refactoring, if
> possible)
> 
> (2) need_resched() is triggered for one of the waiters.
> 
> (3) !owner && rt_task(p)
> 
> quite unlikely, but possible (there are 2 race windows).
> 
> Of course, the question is whether it really takes a noticeable amount
> of time to get out of the [**] state.
> I'd imagine it can be a case for highly-contended locks.
> 
> If this is the case indeed, then which of 1,2,3 gets triggered the most.

Sorry, I don't have stats on that.

> 
> Have you tried removing need_resched() checks? So we kind of emulate
> real spinlocks here.

Unfortunately, the need_resched() checks deal with a few of the ugly
corners.  They are more important without the waiter list check.
Basically if we spun without the need_resched() checks, the process who
wants to unlock might not be able to schedule back in.

-chris