Re: [PATCH v4] sched/freq: move call to cpufreq_update_util

[Vincent Guittot <vincent.guittot@linaro.org>]

Le Friday 15 Nov 2019 à 11:29:03 (+0100), Vincent Guittot a écrit :
> On Fri, 15 Nov 2019 at 11:18, Vincent Guittot
> <vincent.guittot@linaro.org> wrote:
> >
> > On Fri, 15 Nov 2019 at 10:55, Peter Zijlstra <peterz@infradead.org> wrote:
> > >
> > > On Thu, Nov 14, 2019 at 06:07:31PM +0100, Vincent Guittot wrote:
> > > > update_cfs_rq_load_avg() calls cfs_rq_util_change() everytime pelt decays,
> > > > which might be inefficient when cpufreq driver has rate limitation.
> > > >
> > > > When a task is attached on a CPU, we have call path:
> > > >
> > > > update_load_avg()
> > > >   update_cfs_rq_load_avg()
> > > >     cfs_rq_util_change -- > trig frequency update
> > > >   attach_entity_load_avg()
> > > >     cfs_rq_util_change -- > trig frequency update
> > > >
> > > > The 1st frequency update will not take into account the utilization of the
> > > > newly attached task and the 2nd one might be discard because of rate
> > > > limitation of the cpufreq driver.
> > >
> > > Doesn't this just show that a dumb rate limit in the driver is broken?
> >
> > But the rate limit may come from HW constraints that forces to wait
> > let say 4ms or even 10ms between each frequency update.
> >
> > >
> > > > update_cfs_rq_load_avg() is only called by update_blocked_averages()
> > > > and update_load_avg() so we can move the call to
> > > > cfs_rq_util_change/cpufreq_update_util() into these 2 functions. It's also
> > > > interesting to notice that update_load_avg() already calls directly
> > > > cfs_rq_util_change() for !SMP case.
> > > >
> > > > This changes will also ensure that cpufreq_update_util() is called even
> > > > when there is no more CFS rq in the leaf_cfs_rq_list to update but only
> > > > irq, rt or dl pelt signals.
> > >
> > > I don't think it does that; that is, iirc the return value of
> > > ___update_load_sum() is 1 every time a period lapses. So even if the avg
> > > is 0 and doesn't change, it'll still return 1 on every period.
> > >
> > > Which is what that dumb rate-limit thing wants of course. But I'm still
> > > thinking that it's stupid to do. If nothing changes, don't generate
> > > events.
> >
> > When everything (irq, dl, rt, cfs) is 0, we don't generate events
> > because update_blocked_averages is no more called because
> > rq->has_blocked_load is clear
> >
> > With current implementation, if cfs is 0 but not irq, dl or rt, we
> > don't call cpufreq_update_util because it is only called through cfs
> >
> > >
> > > If anything, update_blocked_avgerages() should look at
> > > @done/others_have_blocked() to emit events for rt,dl,irq.
> >
> > other_have_blocked can be set but no decay happened during the update
> > and we don't need to call cpufreq_update_util
> >
> > >
> > > So why are we making the scheduler code more ugly instead of fixing that
> > > driver?
> 
> Also, I think that calling cfs_rq_util_change in
> attach_entity_load_avg is not optimal because the attach can happen at
> a child level before it has been propagated down to root
> So I'm working on trying to remove it from attach_entity_load_avg and
> keep it in update_load_avg. this would help cleaning the ugly
> 
> -       } else if (decayed && (flags & UPDATE_TG))
> -               update_tg_load_avg(cfs_rq, 0);
> +       } else if (decayed) {
> +               cfs_rq_util_change(cfs_rq, 0);
> +
> +               if (flags & UPDATE_TG)
> +                       update_tg_load_avg(cfs_rq, 0);
> +       }
>  }
>

we can also do this instead :

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index d377a3f..550b6bc 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3614,15 +3614,15 @@ static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
                 *
                 * IOW we're enqueueing a task on a new CPU.
                 */
-               attach_entity_load_avg(cfs_rq, se, SCHED_CPUFREQ_MIGRATION);
+               attach_entity_load_avg(cfs_rq, se, 0);
                update_tg_load_avg(cfs_rq, 0);
+               decayed = 1;
 
-       } else if (decayed) {
-               cfs_rq_util_change(cfs_rq, 0);
+       } else if (decayed && (flags & UPDATE_TG))
+               update_tg_load_avg(cfs_rq, 0);
 
-               if (flags & UPDATE_TG)
-                       update_tg_load_avg(cfs_rq, 0);
-       }
+       if (decayed)
+               cfs_rq_util_change(cfs_rq, 0);
 }



>