Re: [PATCH v4 5/8] sched/fair: Take into account latency priority at wakeup

[Hillf Danton <hdanton@sina.com>]

On 16 Sep 2022 15:36:53 +0200 Vincent Guittot <vincent.guittot@linaro.org> wrote:
> 
> Hi Hillf,
> 
> On Fri, 16 Sept 2022 at 14:03, Hillf Danton <hdanton@sina.com> wrote:
> >
> > Hello Vincent
> >
> > On 16 Sep 2022 10:03:02 +0200 Vincent Guittot <vincent.guittot@linaro.org> wrote:
> > >
> > > @@ -4606,6 +4608,7 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> > >
> > >       se = __pick_first_entity(cfs_rq);
> > >       delta = curr->vruntime - se->vruntime;
> > > +     delta -= wakeup_latency_gran(curr, se);
> > >
> > >       if (delta < 0)
> > >               return;
> >
> > What is derived from the latency nice you added is the runtime granulaity
> > which has a role in preempting the current task.
> >
> > Given the same defination of latency nice as the nice, the runtime granularity
> > can be computed without introducing the latency nice.
> >
> > Only for thoughts now.
> >
> > Hillf
> >
> > +++ b/kernel/sched/fair.c
> > @@ -4569,7 +4569,7 @@ dequeue_entity(struct cfs_rq *cfs_rq, st
> >  static void
> >  check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> >  {
> > -       unsigned long ideal_runtime, delta_exec;
> > +       unsigned long ideal_runtime, delta_exec, granu;
> >         struct sched_entity *se;
> >         s64 delta;
> >
> > @@ -4594,6 +4594,14 @@ check_preempt_tick(struct cfs_rq *cfs_rq
> >                 return;
> >
> >         se = __pick_first_entity(cfs_rq);
> > +
> > +       granu = sysctl_sched_min_granularity +
> > +               (ideal_runtime - sysctl_sched_min_granularity) *
> > +               (se->latency_nice + 20) / LATENCY_NICE_WIDTH;
> 
> There is no latency_nice field in se but a latency_offset instead
> 
> Also at this step, we are sure that curr has run at least
> sysctl_sched_min_granularity and we want now to compare curr vruntime
> with first se's one. We take the latency offset into account to make
> sure we will not preempt curr too early
> 
> > +
> > +       if (delta_exec < granu)
> > +               return;
> > +
> >         delta = curr->vruntime - se->vruntime;
> >
> >         if (delta < 0)
		return;

	    if (delta > ideal_runtime)
		resched_curr(rq_of(cfs_rq));

After another look, curr is not preempted without the gap in vruntime
between curr and the first entity growing more than ideal runtime, while
with latency_offset, since the gap becomes larger, preempt happens later
than ideal runtime thoughts IMO.