Re: [RFC PATCH V3 4/6] sched/deadline: Introduce deadline servers

[Joel Fernandes]

On Tue, Jul 4, 2023 at 11:52 AM Daniel Bristot de Oliveira
<bristot@xxxxxxxxxx> wrote:
>
>
> Back from EOSS...
>
> On 6/23/23 18:47, Valentin Schneider wrote:
> > On 08/06/23 17:58, Daniel Bristot de Oliveira wrote:
> >> @@ -2033,9 +2147,20 @@ static struct task_struct *pick_next_task_dl(struct rq *rq)
> >>      struct task_struct *p;
> >>
> >>      p = pick_task_dl(rq);
> >> -    if (p)
> >> +    if (!p)
> >> +            return p;
> >> +
> >> +    /*
> >> +     * XXX: re-check !dl_server, changed from v2 because of
> >> +     * pick_next_task_dl change
> >> +     */
> >> +    if (!dl_server(&p->dl))
> >>              set_next_task_dl(rq, p, true);
> >>
> >
> > Should this be
> >
> >        if (!p->server)
> >
> > instead? AFAICT dl_server(&p->dl) can never be true since there's no
> > pi_se-like link to the server via the dl_se, only via the task_struct, and
> > the server pick cannot return the server itself (as it's a pure sched_entity).
>
> makes sense... I will check that in the v4.

Makes sense to me too. Since p is either a real DL task or a CFS task,
"if (dl_server(&p->dl))" is incorrect.  "if (p->server)" is the right
check.

Optionally, a BUG_ON() as well could be added to make sure the p
returned by pick_task_dl() is always false:
BUG_ON(dl_server(&p->dl));

thanks,

 - Joel



>
> >
> >> +    /* XXX not quite right */
> >> +    if (hrtick_enabled(rq))
> >> +            start_hrtick_dl(rq, &p->dl);
> >> +
> >
> > IIUC that got hauled out of set_next_task_dl() to cover the case where we
> > pick the server (+ the server pick) and want to more thoroughly enforce the
> > server's bandwidth. If so, what's the issue with starting the hrtick here?
>
> I think that the commend was added more as a check if it is correct... it seems it is.
>
> Thanks Vale!
> -- Daniel
>
> >
> >>      return p;
> >>  }
> >>
> >
>