Re: [PATCH 1/2] sched/fair: optimization of update_blocked_averages()

[Vincent Guittot]

On Fri, 8 Feb 2019 at 17:30, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Fri, Feb 08, 2019 at 04:44:53PM +0100, Vincent Guittot wrote:
> > On Fri, 8 Feb 2019 at 16:40, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >
> > >
> > > Argh head hurts!!
> > >
> > > On Wed, Feb 06, 2019 at 05:14:21PM +0100, Vincent Guittot wrote:
> > > > @@ -4438,6 +4450,10 @@ static int tg_unthrottle_up(struct task_group *tg, void *data)
> > > >               /* adjust cfs_rq_clock_task() */
> > > >               cfs_rq->throttled_clock_task_time += rq_clock_task(rq) -
> > > >                                            cfs_rq->throttled_clock_task;
> > > > +
> > > > +             /* Add cfs_rq with already running entity in the list */
> > > > +             if (cfs_rq->nr_running >= 1)
> > > > +                     list_add_leaf_cfs_rq(cfs_rq);
> > > >       }
> > > >
> > > >       return 0;
> > >
> > > Do we want the below to go with the above change?
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 38d4669aa2ef..0bd80a891025 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -4536,6 +4536,8 @@ void unthrottle_cfs_rq(struct cfs_rq *cfs_rq)
> > >         /* update hierarchical throttle state */
> > >         walk_tg_tree_from(cfs_rq->tg, tg_nop, tg_unthrottle_up, (void *)rq);
> > >
> > > +       assert_list_leaf_cfs_rq(rq);
> > > +
> >
> > Good point but this should go after the for_each_sched_entity() loop
>
> Indeed, but that loop does enqueue and can throttle again, should that
> not also get that additional list_add_leaf_cfs_rq() loop we added to
> enqueue_task_fair() to finish the add?

Initially, I added this additional loop but finally removed it because
I didn't hit it during my tests. IIRC, we are protected by
throttle_count in such case, which is not the case when we enqueue a
task