On 22/07/19 15:21, Dietmar Eggemann wrote:
> On 7/22/19 2:28 PM, Juri Lelli wrote:
> > On 22/07/19 13:07, Dietmar Eggemann wrote:
> >> On 7/19/19 3:59 PM, Juri Lelli wrote:
> >>
> >> [...]
> >>
> >>> @@ -557,6 +558,38 @@ static struct rq *dl_task_offline_migration(struct rq *rq, struct task_struct *p
> >>> double_lock_balance(rq, later_rq);
> >>> }
> >>>
> >>> + if (p->dl.dl_non_contending || p->dl.dl_throttled) {
> >>> + /*
> >>> + * Inactive timer is armed (or callback is running, but
> >>> + * waiting for us to release rq locks). In any case, when it
> >>> + * will file (or continue), it will see running_bw of this
> >>
> >> s/file/fire ?
> >
> > Yep.
> >
> >>> + * task migrated to later_rq (and correctly handle it).
> >>
> >> Is this because of dl_task_timer()->enqueue_task_dl()->task_contending()
> >> setting dl_se->dl_non_contending = 0 ?
> >
> > No, this is related to inactive_task_timer() callback. Since the task is
> > migrated (by this function calling set_task_cpu()) because a CPU hotplug
> > operation happened, we need to reflect this w.r.t. running_bw, or
> > inactive_task_timer() might sub from the new CPU and cause running_bw to
> > underflow.
>
> I was more referring to the '... it will see running_bw of thus task
> migrated to later_rq ...) and specifically to the HOW the timer
> callback can detect this.
Oh, it actually doesn't "actively" detect this condition. The problem is
that if it still sees dl_non_contending == 1, it will sub (from the
"new" rq to which task's running_bw hasn't been added - w/o this fix)
and cause the underflow.
