On Thu, 12 Dec 2019 at 11:23, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Thu, Dec 12, 2019 at 11:14:24AM +0100, Peter Zijlstra wrote:
> > On Thu, Dec 12, 2019 at 11:10:31AM +0100, Peter Zijlstra wrote:
> > > @@ -4156,13 +4159,13 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
> > > if (delta_exec < sysctl_sched_min_granularity)
> > > return;
> > >
> > > - se = __pick_first_entity(cfs_rq);
> > > + se = __pick_next_entity(cfs_rq, NULL);
> > > delta = curr->vruntime - se->vruntime;
> > >
> > > if (delta < 0)
> > > return;
> >
> > What I mean with the below comment is, when this isn't enough, try
> > something like:
> >
> > if (se == cfs_rq->next)
> > ideal_runtime /= 2;
> >
> > to make it yield sooner to 'next' buddies. Sadly, due to the whole
> > cgroup mess, we can't say what actual task is on the end of it (without
> > doing a full hierarchy pick, which is more expensive still).
>
> Just for giggles, that'd look something like:
>
> while (!entity_is_task(se) {
> cfs_rq = group_cfs_rq(se);
> se = pick_next_entity(cfs_rq, cfs_rq->curr);
> }
> p = task_of(se);
>
> if (is_per_cpu_kthread(p))
> ideal_runtime /= 2;
>
> the core-scheduling patch set includes the right primitive for this I
> think, pick_task_fair().
why not only updating wan_gran() which is the only function which uses
sysctl_sched_wakeup_granularity ?
For per cpu kthread, we could set the gran to sched_min_granularity
instead of scaling it with thread's priority so per cpu kthread can
still preempt current ask even if sysctl_sched_wakeup_granularity is
large
>
> > > - if (delta > ideal_runtime)
> > > + if (delta > ideal_runtime) // maybe frob this too ?
> > > resched_curr(rq_of(cfs_rq));
> > > }
