Re: [PATCH 1/2] sched/fair: skip newidle update stats

From: Peter Zijlstra
Date: Fri Nov 12 2021 - 10:30:01 EST


On Fri, Nov 12, 2021 at 03:47:21PM +0100, Vincent Guittot wrote:
> On Fri, 12 Nov 2021 at 15:29, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Nov 12, 2021 at 10:58:56AM +0100, Vincent Guittot wrote:
> > > In case we skip the newly idle LB entirely or we abort it because we are
> > > going to exceed the avg_idle, we have to make sure to not start an update
> > > of the blocked load when entering idle
> > >
> > > Signed-off-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
> > > ---
> > > kernel/sched/fair.c | 18 ++++++++++++++----
> > > 1 file changed, 14 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index 13950beb01a2..a162b0ec8963 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -10861,7 +10861,7 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> > > int this_cpu = this_rq->cpu;
> > > u64 t0, t1, curr_cost = 0;
> > > struct sched_domain *sd;
> > > - int pulled_task = 0;
> > > + int pulled_task = 0, early_stop = 0;
> > >
> > > update_misfit_status(NULL, this_rq);
> > >
> > > @@ -10898,8 +10898,16 @@ static int newidle_balance(struct rq *this_rq, struct rq_flags *rf)
> > > if (!READ_ONCE(this_rq->rd->overload) ||
> > > (sd && this_rq->avg_idle < sd->max_newidle_lb_cost)) {
> > >
> > > - if (sd)
> > > + if (sd) {
> > > update_next_balance(sd, &next_balance);
> > > +
> > > + /*
> > > + * We skip new idle LB because there is not enough
> > > + * time before next wake up. Make sure that we will
> > > + * not kick NOHZ_NEWILB_KICK
> > > + */
> > > + early_stop = 1;
> > > + }
> > > rcu_read_unlock();
> > >
> > > goto out;

> > Anyway, does nohz_newidle_balance() want to loose it's ->avg_idle test
> > with this on?
>
> This test still covers cases with short newly idle balance. Being
> conservative, people never complained that the update of blocked load
> average of idle CPUs doesn't happen often enough. It's most often the
> opposite

Well, per commit c5b0a7eefc70 ("sched/fair: Remove
sysctl_sched_migration_cost condition") combined with the above change,
we no longer call nohz_newidle_balance() in exactly that condition,
right?

Or are we worried about that !overload case?