Re: [PATCH v6 1/4] sched/fair: Fix attaching task sched avgs twice when switching to fair or changing task group

[Vincent Guittot]

On 17 June 2016 at 04:12, Yuyang Du <yuyang.du@xxxxxxxxx> wrote:
>
> On Thu, Jun 16, 2016 at 11:21:55PM +0200, Vincent Guittot wrote:
> > On 16 June 2016 at 22:07, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > On Thu, Jun 16, 2016 at 09:00:57PM +0200, Vincent Guittot wrote:
> > >> On 16 June 2016 at 20:51, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > >> > On Thu, Jun 16, 2016 at 06:30:13PM +0200, Vincent Guittot wrote:
> > >> >> With patch [1] for the init of cfs_rq side, all use cases will be
> > >> >> covered regarding the issue linked to a last_update_time set to 0 at
> > >> >> init
> > >> >> [1] https://lkml.org/lkml/2016/5/30/508
> > >> >
> > >> > Aah, wait, now I get it :-)
> > >> >
> > >> > Still, we should put cfs_rq_clock_task(cfs_rq) in it, not 1. And since
> > >> > we now acquire rq->lock on init this should well be possible. Lemme sort
> > >> > that.
> > >>
> > >> yes with the rq->lock we can use cfs_rq_clock_task which is make more
> > >> sense than 1.
> > >> But the delta can be still significant between the creation of the
> > >> task group and the 1st task that will be attach to the cfs_rq
> > >
> > > Ah, I think I've spotted more fail.
> > >
> > > And I think you're right, it doesn't matter, in fact, 0 should have been
> > > fine too!
> > >
> > > enqueue_entity()
> > >   enqueue_entity_load_avg()
> > >     update_cfs_rq_load_avg()
> > >       now = clock()
> > >       __update_load_avg(&cfs_rq->avg)
> > >         cfs_rq->avg.last_load_update = now
> > >         // ages 0 load/util for: now - 0
> > >     if (migrated)
> > >       attach_entity_load_avg()
> > >         se->avg.last_load_update = cfs_rq->avg.last_load_update; // now != 0
> > >
> > > So I don't see how it can end up being attached again.
> >
> > In fact it has already been attached during the sched_move_task. The
> > sequence for the 1st task that is attached to a cfs_rq is :
> >
> > sched_move_task()
> >   task_move_group_fair()
> >     detach_task_cfs_rq()
> >     set_task_rq()
> >     attach_task_cfs_rq()
> >       attach_entity_load_avg()
> >         se->avg.last_load_update = cfs_rq->avg.last_load_update == 0
> >
>
> Then again, does this fix it?

Peter's proposal to update the cfs_rq before attaching/detaching the
task,  fixes the problem

>
>
> static void task_move_group_fair(struct task_struct *p)
> {
>         detach_task_cfs_rq(p);
>         set_task_rq(p, task_cpu(p));
>         attach_task_cfs_rq(p);
>         /*
>          * If the cfs_rq's last_update_time is 0, attach the sched avgs
>          * won't be anything useful, as it will be decayed to 0 when any
>          * sched_entity is enqueued to that cfs_rq.
>          *
>          * On the other hand, if the cfs_rq's last_update_time is 0, we
>          * must reset the task's last_update_time to ensure we will attach
>          * the sched avgs when the task is enqueued.
>          */
>         if (!cfs_rq_of(&p->se)->avg.last_update_time)
>                 reset_task_last_update_time(p);
>         else
>                 attach_entity_load_avg(cfs_rq_of(&p->se), &p->se);
> }