Re: [PATCH v7] sched: Consolidate cpufreq updates

[Qais Yousef <qyousef@layalina.io>]

On 08/13/24 12:02, Vincent Guittot wrote:
> >  void wakeup_preempt(struct rq *rq, struct task_struct *p, int flags)
> > @@ -4913,6 +4923,93 @@ static inline void __balance_callbacks(struct rq *rq)
> >
> >  #endif
> >
> > +static __always_inline void
> > +__update_cpufreq_ctx_switch(struct rq *rq, struct task_struct *prev)
> > +{
> > +#ifdef CONFIG_CPU_FREQ
> > +       if (prev && prev->dl.flags & SCHED_FLAG_SUGOV) {
> > +               /* Sugov just did an update, don't be too aggressive */
> > +               return;
> > +       }
> > +
> > +       /*
> > +        * RT and DL should always send a freq update. But we can do some
> > +        * simple checks to avoid it when we know it's not necessary.
> > +        *
> > +        * iowait_boost will always trigger a freq update too.
> > +        *
> > +        * Fair tasks will only trigger an update if the root cfs_rq has
> > +        * decayed.
> > +        *
> > +        * Everything else should do nothing.
> > +        */
> > +       switch (current->policy) {
> > +       case SCHED_NORMAL:
> > +       case SCHED_BATCH:
> > +       case SCHED_IDLE:
> > +               if (unlikely(current->in_iowait)) {
> > +                       cpufreq_update_util(rq, SCHED_CPUFREQ_IOWAIT | SCHED_CPUFREQ_FORCE_UPDATE);
> > +                       return;
> > +               }
> > +
> > +#ifdef CONFIG_SMP
> > +               /*
> > +                * Send an update if we switched from RT or DL as they tend to
> > +                * boost the CPU and we are likely able to reduce the freq now.
> > +                */
> > +               rq->cfs.decayed |= prev && (rt_policy(prev->policy) || dl_policy(prev->policy));
> > +
> > +               if (unlikely(rq->cfs.decayed)) {
> 
> My previous use case of a task non preempting current with large
> util_est is fixed with this version but I'm facing a new one a bit
> similar because of waiting for the context switch and the decay to try
> to update the frequency.
> 
> When the task wakes up on an idle cpu, you wait for the decay to
> update the freq but if the freq is low and the pelt has been updated
> recently (less than 1024us) you can wait a long time before the next
> decay and the freq update. This is a problem if the task's util_est is
> large because you can stay several ms at low frequency before taking
> into account task's util_est

It is a symptom of the same problem. It seems we don't decay and we omit the
cpufreq update.

Why this was not a problem before? AFAICT we only send an update before my
patch if we had a decay and I didn't change this condition. Were we just
getting more lucky or did I change some behavior unwittingly?

The problem with my patch is that I do this unconditional only if we failed
preemption check. But looks like I must enforce a cpufreq update after every
enqueue. I think the overhead of not checking rq->cfs.decayed would be high if
we always call a cpufreq update.

I'll just set rq->cfs.decayaed in util_est_enqueue() which should address both
use cases.

Thanks!