Re: [PATCH v8 -tip 08/26] sched/fair: Snapshot the min_vruntime of CPUs on force idle

From: Joel Fernandes
Date: Thu Oct 29 2020 - 22:36:56 EST


On Thu, Oct 29, 2020 at 2:59 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Oct 29, 2020 at 02:24:29PM -0400, Joel Fernandes wrote:
>
> > > @@ -4823,10 +4822,8 @@ pick_next_task(struct rq *rq, struct tas
> > > if (!rq_i->core_pick)
> > > continue;
> > >
> > > - if (is_task_rq_idle(rq_i->core_pick) && rq_i->nr_running &&
> > > - !rq_i->core->core_forceidle) {
> > > - rq_i->core->core_forceidle = true;
> > > - }
> > > + if (!(fi_before && rq->core->core_forceidle))
> > > + task_vruntime_update(rq_i, rq_i->core_pick);
> >
> > Shouldn't this be:
> >
> > if (!fi_before && rq->core->core_forceidle)
> > task_vruntime_update(rq_i, rq_i->core_pick);
> >
> > ?
>
> *groan*, I should've written a comment there :/
>
> When we're not fi, we need to update.
> when we're fi and we were not fi, we must update
> When we're fi and we were already fi, we must not update
>
> Which gives:
>
> fib fi X
>
> 0 0 1
> 0 1 0
> 1 0 1
> 1 1 1
>
> which is: !(!fib && fi) or something.
>

Got it! This is what my initial patch intended to do as well, but
yours is better.

> > > +bool cfs_prio_less(struct task_struct *a, struct task_struct *b)
> > > +{
> > > + struct rq *rq = task_rq(a);
> > > + struct sched_entity *sea = &a->se;
> > > + struct sched_entity *seb = &b->se;
> > > + struct cfs_rq *cfs_rqa;
> > > + struct cfs_rq *cfs_rqb;
> > > + s64 delta;
> > > +
> > > + SCHED_WARN_ON(task_rq(b)->core != rq->core);
> > > +
> > > + while (sea->cfs_rq->tg != seb->cfs_rq->tg) {
> > > + int sea_depth = sea->depth;
> > > + int seb_depth = seb->depth;
> > > +
> > > + if (sea_depth >= seb_depth)
> > > + sea = parent_entity(sea);
> > > + if (sea_depth <= seb_depth)
> > > + seb = parent_entity(seb);
> > > + }
> > > +
> > > + if (rq->core->core_forceidle) {
> > > + se_fi_update(sea, rq->core->core_forceidle_seq, true);
> > > + se_fi_update(seb, rq->core->core_forceidle_seq, true);
> > > + }
> >
> > As we chatted on IRC you mentioned the reason for the sync here is:
> >
> > say we have 2 cgroups (a,b) under root, and we go force-idle in a, then we
> > update a and root. Then we pick and end up in b, but b hasn't been updated
> > yet.
> >
> > One thing I was wondering about that was, if the pick of 'b' happens much
> > later than 'a', then the snapshot might be happening too late right?
>
> No, since this is the first pick in b since fi, it cannot have advanced.
> So by updating to fi_seq before picking, we guarantee it is unchanged
> since we went fi.

Makes complete sense.

I got it to a point where the latencies are much lower, but still not
at a point where it's as good as the initial patch I posted.

There could be more bugs. At the moment, the only one I corrected in
your patch is making the truth table do !(!fib && fi). But there is
still something else going on.

Thanks!

- Joel