Re: [RFC PATCH v2 1/5] sched/fair: Add ancestors of unthrottled undecayed cfs_rq

[Vincent Guittot]

On Fri, 10 Sept 2021 at 13:35, Michal Koutný <mkoutny@xxxxxxxx> wrote:
>
> Hello Vincent.
>
> Thank you for looking into this!
>
> On Thu, Sep 09, 2021 at 03:57:37PM +0200, Vincent Guittot <vincent.guittot@xxxxxxxxxx> wrote:
> > > +               /* Nothing to run but something to decay? Complete the branch */
> > > +               if (cfs_rq->on_list)
> >
> > Could you use !cfs_rq_is decayed(cfs_rq) ?
>
> What needs to be checked here is whether the list was modified by adding
> the cfs_rq (and branch needs closing).
>
> It'd appear that the equal condition like in tg_unthrottle_up() would
> make do, i.e.
>         !cfs_rq_is_decayed(cfs_rq) || cfs_rq->nr_running
> but the unthrottle_cfs_rq() can be called under a still throttled
> ancestor (i.e. throttle_count not dropping to zero) and in such a case
> cfs_rq should not be added to the list yet.
>
> Therefore, mere !cfs_rq_is_decayed(cfs_rq) doesn't seem correct to me.

fair enough

>
> > > +                       for_each_sched_entity(se) {
> > > +                               if (list_add_leaf_cfs_rq(group_cfs_rq(se)))
> > > +                                       break;
> > > +                       }
> > > +               assert_list_leaf_cfs_rq(rq);
> >
> > Instead of adding  a loop here you should better jump to unthrottle_throttle ?
>
> Oh, that looks a bit clumsy now (it's an artifact I've left when
> reordering the patch in the series to be backport-friendly).
> Jump to unthrottle_throttle seems easier indeed, there would be just the
> additional check
>         if (rq->curr == rq->idle && rq->cfs.nr_running)
> . Besides unnecessary work, it should be harmless.

yes the condition should be always false

>
> Is the jump the preferred form?

yes compared to adding the exact same loop

>
> Michal