Re: Severe performance regression w/ 4.4+ on Android due to cgroup locking changes

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Jul 13, 2016 at 02:01:27PM -0700, Dmitry Shmidt wrote:
> On Wed, Jul 13, 2016 at 1:52 PM, Paul E. McKenney
> <paulmck@linux.vnet.ibm.com> wrote:
> > On Wed, Jul 13, 2016 at 10:26:57PM +0200, Peter Zijlstra wrote:
> >> On Wed, Jul 13, 2016 at 04:18:23PM -0400, Tejun Heo wrote:
> >> > Hello, John.
> >> >
> >> > On Wed, Jul 13, 2016 at 01:13:11PM -0700, John Stultz wrote:
> >> > > On Wed, Jul 13, 2016 at 11:33 AM, Tejun Heo <tj@kernel.org> wrote:
> >> > > > On Wed, Jul 13, 2016 at 02:21:02PM -0400, Tejun Heo wrote:
> >> > > >> One interesting thing to try would be replacing it with a regular
> >> > > >> non-percpu rwsem and see how it behaves.  That should easily tell us
> >> > > >> whether this is from actual contention or artifacts from percpu_rwsem
> >> > > >> implementation.
> >> > > >
> >> > > > So, something like the following.  Can you please see whether this
> >> > > > makes any difference?
> >> > >
> >> > > Yea. So this brings it down for me closer to what we're seeing with
> >> > > the Dmitry's patch reverting the two problematic commits, usually
> >> > > 10-50us with one early spike at 18ms.
> >> >
> >> > So, it's a percpu rwsem issue then.  I haven't really followed the
> >> > perpcpu rwsem changes closely.  Oleg, are multi-milisec delay expected
> >> > on down write expected with the current implementation of
> >> > percpu_rwsem?
> >>
> >> There is a synchronize_sched() in there, so sorta. That thing is heavily
> >> geared towards readers, as is the only 'sane' choice for global locks.
> >
> > Then one diagnostic step to take would be to replace that
> > synchronize_sched() with synchronize_sched_expedited(), and see if that
> > gets rid of the delays.
> >
> > Not a particularly real-time-friendly fix, but certainly a good check
> > on our various assumptions.
> 
> All delays <200 us, but one that is 3 ms.

That does indicate that synchronize_sched() is the latency culprit.

Tejun's lglock-like approach seems to me to be the next thing to try.

                                                        Thanx, Paul

> > ------------------------------------------------------------------------
> >
> > diff --git a/kernel/rcu/sync.c b/kernel/rcu/sync.c
> > index be922c9f3d37..211acddc7e21 100644
> > --- a/kernel/rcu/sync.c
> > +++ b/kernel/rcu/sync.c
> > @@ -38,19 +38,19 @@ static const struct {
> >  #endif
> >  } gp_ops[] = {
> >         [RCU_SYNC] = {
> > -               .sync = synchronize_rcu,
> > +               .sync = synchronize_rcu_expedited,
> >                 .call = call_rcu,
> >                 .wait = rcu_barrier,
> >                 __INIT_HELD(rcu_read_lock_held)
> >         },
> >         [RCU_SCHED_SYNC] = {
> > -               .sync = synchronize_sched,
> > +               .sync = synchronize_sched_expedited,
> >                 .call = call_rcu_sched,
> >                 .wait = rcu_barrier_sched,
> >                 __INIT_HELD(rcu_read_lock_sched_held)
> >         },
> >         [RCU_BH_SYNC] = {
> > -               .sync = synchronize_rcu_bh,
> > +               .sync = synchronize_rcu_bh_expedited,
> >                 .call = call_rcu_bh,
> >                 .wait = rcu_barrier_bh,
> >                 __INIT_HELD(rcu_read_lock_bh_held)
> >
>