Re: [RFC PATCH 3/4] sched/fair: delay update_tg_load_avg() for cfs_rq's removed load
From: Vincent Guittot
Date: Thu Jul 20 2023 - 09:11:00 EST
On Wed, 19 Jul 2023 at 15:29, Aaron Lu <aaron.lu@xxxxxxxxx> wrote:
>
> On Wed, Jul 19, 2023 at 11:47:06AM +0200, Vincent Guittot wrote:
> > On Wed, 19 Jul 2023 at 10:01, Aaron Lu <aaron.lu@xxxxxxxxx> wrote:
> > >
> > > On Wed, Jul 19, 2023 at 01:18:26PM +0800, Aaron Lu wrote:
> > > > Alternatively, I can remove some callsites of update_tg_load_avg() like
> > > > you suggested below and only call update_tg_load_avg() when cfs_rq is
> > > > decayed(really just decayed, not when it detected it has removed load
> > > > pending or load propagated from its children). This way it would give us
> > > > similar result as above(roughly once per ms).
> > >
> > > Something like this: (I think this is better since it removed those
> > > unnecessary calls to update_tg_load_avg(), although it is inline but
> > > still)
> > >
> > >
> > > From bc749aaefa6bed36aa946921a4006b3dddb69b77 Mon Sep 17 00:00:00 2001
> > > From: Aaron Lu <aaron.lu@xxxxxxxxx>
> > > Date: Wed, 19 Jul 2023 13:54:48 +0800
> > > Subject: [PATCH] sched/fair: only update_tg_load_avg() when cfs_rq decayed
> > >
> > > ---
> > > kernel/sched/fair.c | 22 +++++++---------------
> > > 1 file changed, 7 insertions(+), 15 deletions(-)
> > >
> > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > index a80a73909dc2..7d5b7352b8b5 100644
> > > --- a/kernel/sched/fair.c
> > > +++ b/kernel/sched/fair.c
> > > @@ -3913,16 +3913,16 @@ static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum
> > > }
> > >
> > > /* Update task and its cfs_rq load average */
> > > -static inline int propagate_entity_load_avg(struct sched_entity *se)
> > > +static inline void propagate_entity_load_avg(struct sched_entity *se)
> > > {
> > > struct cfs_rq *cfs_rq, *gcfs_rq;
> > >
> > > if (entity_is_task(se))
> > > - return 0;
> > > + return;
> > >
> > > gcfs_rq = group_cfs_rq(se);
> > > if (!gcfs_rq->propagate)
> > > - return 0;
> > > + return;
> > >
> > > gcfs_rq->propagate = 0;
> > >
> > > @@ -3936,8 +3936,6 @@ static inline int propagate_entity_load_avg(struct sched_entity *se)
> > >
> > > trace_pelt_cfs_tp(cfs_rq);
> > > trace_pelt_se_tp(se);
> > > -
> > > - return 1;
> > > }
> > >
> > > /*
> > > @@ -3974,9 +3972,8 @@ static inline bool skip_blocked_update(struct sched_entity *se)
> > >
> > > static inline void update_tg_load_avg(struct cfs_rq *cfs_rq) {}
> > >
> > > -static inline int propagate_entity_load_avg(struct sched_entity *se)
> > > +static inline void propagate_entity_load_avg(struct sched_entity *se)
> > > {
> > > - return 0;
> > > }
> > >
> > > static inline void add_tg_cfs_propagate(struct cfs_rq *cfs_rq, long runnable_sum) {}
> > > @@ -4086,7 +4083,7 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
> > > {
> > > unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
> > > struct sched_avg *sa = &cfs_rq->avg;
> > > - int decayed = 0;
> > > + int decayed;
> > >
> > > if (cfs_rq->removed.nr) {
> > > unsigned long r;
> > > @@ -4134,11 +4131,9 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
> > > */
> > > add_tg_cfs_propagate(cfs_rq,
> > > -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
> > > -
> > > - decayed = 1;
> > > }
> > >
> > > - decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
> > > + decayed = __update_load_avg_cfs_rq(now, cfs_rq);
> > > u64_u32_store_copy(sa->last_update_time,
> > > cfs_rq->last_update_time_copy,
> > > sa->last_update_time);
> > > @@ -4252,7 +4247,7 @@ static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
> > > __update_load_avg_se(now, cfs_rq, se);
> > >
> > > decayed = update_cfs_rq_load_avg(now, cfs_rq);
> > > - decayed |= propagate_entity_load_avg(se);
> > > + propagate_entity_load_avg(se);
> >
> > but then you also skip the call to cfs_rq_util_change()
>
> Ah right, I missed that, thanks for catching this.
>
> Updated:
>
> From 09a649f8111cfca656b7b735da975ef607b00956 Mon Sep 17 00:00:00 2001
> From: Aaron Lu <aaron.lu@xxxxxxxxx>
> Date: Wed, 19 Jul 2023 13:54:48 +0800
> Subject: [PATCH] sched/fair: only update_tg_load_avg() when cfs_rq decayed
>
> ---
> kernel/sched/fair.c | 17 ++++++-----------
> 1 file changed, 6 insertions(+), 11 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index a80a73909dc2..8d4b9e0a19b6 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -4086,7 +4086,7 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
> {
> unsigned long removed_load = 0, removed_util = 0, removed_runnable = 0;
> struct sched_avg *sa = &cfs_rq->avg;
> - int decayed = 0;
> + int decayed;
>
> if (cfs_rq->removed.nr) {
> unsigned long r;
> @@ -4134,11 +4134,9 @@ update_cfs_rq_load_avg(u64 now, struct cfs_rq *cfs_rq)
> */
> add_tg_cfs_propagate(cfs_rq,
> -(long)(removed_runnable * divider) >> SCHED_CAPACITY_SHIFT);
> -
> - decayed = 1;
We need this to propagate the change in other place like cpufreq
> }
>
> - decayed |= __update_load_avg_cfs_rq(now, cfs_rq);
> + decayed = __update_load_avg_cfs_rq(now, cfs_rq);
> u64_u32_store_copy(sa->last_update_time,
> cfs_rq->last_update_time_copy,
> sa->last_update_time);
> @@ -4242,7 +4240,7 @@ static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
> static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
> {
> u64 now = cfs_rq_clock_pelt(cfs_rq);
> - int decayed;
> + int decayed, propagated;
>
> /*
> * Track task load average for carrying it to new CPU after migrated, and
> @@ -4252,7 +4250,7 @@ static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
> __update_load_avg_se(now, cfs_rq, se);
>
> decayed = update_cfs_rq_load_avg(now, cfs_rq);
> - decayed |= propagate_entity_load_avg(se);
> + propagated = propagate_entity_load_avg(se);
>
> if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
>
> @@ -4264,19 +4262,16 @@ 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);
> - update_tg_load_avg(cfs_rq);
> -
> } else if (flags & DO_DETACH) {
> /*
> * DO_DETACH means we're here from dequeue_entity()
> * and we are migrating task out of the CPU.
> */
> detach_entity_load_avg(cfs_rq, se);
> - update_tg_load_avg(cfs_rq);
> - } else if (decayed) {
> + } else if (decayed || propagated) {
> cfs_rq_util_change(cfs_rq, 0);
>
> - if (flags & UPDATE_TG)
> + if (decayed && (flags & UPDATE_TG))
It would be simpler and more readable to clear UPDATE_TG or not set it
from the beginning
IIUC, you rely on the fact that a decay happens every 1024 us of the
cfs_rq_clock_pelt() which is scaled by frequency and cpu compute
capacity. So you can end up with a cfs_rq_clock_pelt() that is far
slower than real clock and the 1ms can easily be extended to dozens of
ms
> update_tg_load_avg(cfs_rq);
> }
> }
> --
> 2.41.0
>
>
> > >
> > > if (!se->avg.last_update_time && (flags & DO_ATTACH)) {
> > >
> > > @@ -4264,15 +4259,12 @@ 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);
> > > - update_tg_load_avg(cfs_rq);
> > > -
> > > } else if (flags & DO_DETACH) {
> > > /*
> > > * DO_DETACH means we're here from dequeue_entity()
> > > * and we are migrating task out of the CPU.
> > > */
> > > detach_entity_load_avg(cfs_rq, se);
> > > - update_tg_load_avg(cfs_rq);
> > > } else if (decayed) {
> > > cfs_rq_util_change(cfs_rq, 0);
> > >
> > > --
> > > 2.41.0
> > >