Date:   Mon, 6 Nov 2023 16:37:32 -0500
From:   Joel Fernandes <joel@...lfernandes.org>
To:     Daniel Bristot de Oliveira <bristot@...nel.org>
Cc:     Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Juri Lelli <juri.lelli@...hat.com>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        Daniel Bristot de Oliveira <bristot@...hat.com>,
        Valentin Schneider <vschneid@...hat.com>,
        linux-kernel@...r.kernel.org,
        Luca Abeni <luca.abeni@...tannapisa.it>,
        Tommaso Cucinotta <tommaso.cucinotta@...tannapisa.it>,
        Thomas Gleixner <tglx@...utronix.de>,
        Vineeth Pillai <vineeth@...byteword.org>,
        Shuah Khan <skhan@...uxfoundation.org>,
        Phil Auld <pauld@...hat.com>
Subject: Re: [PATCH v5 6/7] sched/deadline: Deferrable dl server

On Mon, Nov 6, 2023 at 4:32 PM Joel Fernandes <joel@...lfernandes.org> wrote:
>
> On Mon, Nov 6, 2023 at 2:32 PM Joel Fernandes <joel@...lfernandes.org> wrote:
> >
> > Hi Daniel,
> >
> > On Sat, Nov 4, 2023 at 6:59 AM Daniel Bristot de Oliveira
> > <bristot@...nel.org> wrote:
> > >
> > > Among the motivations for the DL servers is the real-time throttling
> > > mechanism. This mechanism works by throttling the rt_rq after
> > > running for a long period without leaving space for fair tasks.
> > >
> > > The base dl server avoids this problem by boosting fair tasks instead
> > > of throttling the rt_rq. The point is that it boosts without waiting
> > > for potential starvation, causing some non-intuitive cases.
> > >
> > > For example, an IRQ dispatches two tasks on an idle system, a fair
> > > and an RT. The DL server will be activated, running the fair task
> > > before the RT one. This problem can be avoided by deferring the
> > > dl server activation.
> > >
> > > By setting the zerolax option, the dl_server will dispatch an
> > > SCHED_DEADLINE reservation with replenished runtime, but throttled.
> > >
> > > The dl_timer will be set for (period - runtime) ns from start time.
> > > Thus boosting the fair rq on its 0-laxity time with respect to
> > > rt_rq.
> > >
> > > If the fair scheduler has the opportunity to run while waiting
> > > for zerolax time, the dl server runtime will be consumed. If
> > > the runtime is completely consumed before the zerolax time, the
> > > server will be replenished while still in a throttled state. Then,
> > > the dl_timer will be reset to the new zerolax time
> > >
> > > If the fair server reaches the zerolax time without consuming
> > > its runtime, the server will be boosted, following CBS rules
> > > (thus without breaking SCHED_DEADLINE).
> > >
> > > Signed-off-by: Daniel Bristot de Oliveira <bristot@...nel.org>
> > > ---
> > >  include/linux/sched.h   |   2 +
> > >  kernel/sched/deadline.c | 100 +++++++++++++++++++++++++++++++++++++++-
> > >  kernel/sched/fair.c     |   3 ++
> > >  3 files changed, 103 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > index 5ac1f252e136..56e53e6fd5a0 100644
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -660,6 +660,8 @@ struct sched_dl_entity {
> > >         unsigned int                    dl_non_contending : 1;
> > >         unsigned int                    dl_overrun        : 1;
> > >         unsigned int                    dl_server         : 1;
> > > +       unsigned int                    dl_zerolax        : 1;
> > > +       unsigned int                    dl_zerolax_armed  : 1;
> > >
> > >         /*
> > >          * Bandwidth enforcement timer. Each -deadline task has its
> > > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > > index 1d7b96ca9011..69ee1fbd60e4 100644
> > > --- a/kernel/sched/deadline.c
> > > +++ b/kernel/sched/deadline.c
> > > @@ -772,6 +772,14 @@ static inline void replenish_dl_new_period(struct sched_dl_entity *dl_se,
> > >         /* for non-boosted task, pi_of(dl_se) == dl_se */
> > >         dl_se->deadline = rq_clock(rq) + pi_of(dl_se)->dl_deadline;
> > >         dl_se->runtime = pi_of(dl_se)->dl_runtime;
> > > +
> > > +       /*
> > > +        * If it is a zerolax reservation, throttle it.
> > > +        */
> > > +       if (dl_se->dl_zerolax) {
> > > +               dl_se->dl_throttled = 1;
> > > +               dl_se->dl_zerolax_armed = 1;
> > > +       }
> > >  }
> > >
> > >  /*
> > > @@ -828,6 +836,7 @@ static inline void setup_new_dl_entity(struct sched_dl_entity *dl_se)
> > >   * could happen are, typically, a entity voluntarily trying to overcome its
> > >   * runtime, or it just underestimated it during sched_setattr().
> > >   */
> > > +static int start_dl_timer(struct sched_dl_entity *dl_se);
> > >  static void replenish_dl_entity(struct sched_dl_entity *dl_se)
> > >  {
> > >         struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
> > > @@ -874,6 +883,28 @@ static void replenish_dl_entity(struct sched_dl_entity *dl_se)
> > >                 dl_se->dl_yielded = 0;
> > >         if (dl_se->dl_throttled)
> > >                 dl_se->dl_throttled = 0;
> > > +
> > > +       /*
> > > +        * If this is the replenishment of a zerolax reservation,
> > > +        * clear the flag and return.
> > > +        */
> > > +       if (dl_se->dl_zerolax_armed) {
> > > +               dl_se->dl_zerolax_armed = 0;
> > > +               return;
> > > +       }
> > > +
> > > +       /*
> > > +        * A this point, if the zerolax server is not armed, and the deadline
> > > +        * is in the future, throttle the server and arm the zerolax timer.
> > > +        */
> > > +       if (dl_se->dl_zerolax &&
> > > +           dl_time_before(dl_se->deadline - dl_se->runtime, rq_clock(rq))) {
> > > +               if (!is_dl_boosted(dl_se)) {
> > > +                       dl_se->dl_zerolax_armed = 1;
> > > +                       dl_se->dl_throttled = 1;
> > > +                       start_dl_timer(dl_se);
> > > +               }
> > > +       }
> > >  }
> > >
> > >  /*
> > > @@ -1024,6 +1055,13 @@ static void update_dl_entity(struct sched_dl_entity *dl_se)
> > >                 }
> > >
> > >                 replenish_dl_new_period(dl_se, rq);
> > > +       } else if (dl_server(dl_se) && dl_se->dl_zerolax) {
> > > +               /*
> > > +                * The server can still use its previous deadline, so throttle
> > > +                * and arm the zero-laxity timer.
> > > +                */
> > > +               dl_se->dl_zerolax_armed = 1;
> > > +               dl_se->dl_throttled = 1;
> > >         }
> > >  }
> > >
> > > @@ -1056,8 +1094,20 @@ static int start_dl_timer(struct sched_dl_entity *dl_se)
> > >          * We want the timer to fire at the deadline, but considering
> > >          * that it is actually coming from rq->clock and not from
> > >          * hrtimer's time base reading.
> > > +        *
> > > +        * The zerolax reservation will have its timer set to the
> > > +        * deadline - runtime. At that point, the CBS rule will decide
> > > +        * if the current deadline can be used, or if a replenishment
> > > +        * is required to avoid add too much pressure on the system
> > > +        * (current u > U).
> > >          */
> > > -       act = ns_to_ktime(dl_next_period(dl_se));
> > > +       if (dl_se->dl_zerolax_armed) {
> > > +               WARN_ON_ONCE(!dl_se->dl_throttled);
> > > +               act = ns_to_ktime(dl_se->deadline - dl_se->runtime);
> >
> > Just a question, here if dl_se->deadline - dl_se->runtime is large,
> > then does that mean that server activation will be much more into the
> > future? So say I want to give CFS 30%, then it will take 70% of the
> > period before CFS preempts RT thus "starving" CFS for this duration. I
> > think that's Ok for smaller periods and runtimes, though.
> >
> > I think it does reserve the amount of required CFS bandwidth so it is
> > probably OK, though it is perhaps letting RT run more initially (say
> > if CFS tasks are not CPU bound and occasionally wake up, they will
> > always be hit by the 70% latency AFAICS which may be large for large
> > periods and small runtimes).
> >
>
> One more consideration I guess is, because the server is throttled
> till 0-laxity time, it is possible that if CFS sleeps even a bit
> (after the DL-server is unthrottled), then it will be pushed out to a
> full current deadline + period due to CBS. In such a situation,  if
> CFS-server is the only DL task running, it might starve RT for a bit
> more time.
>
> Example, say CFS runtime is 0.3s and period is 1s. At 0.7s, 0-laxity
> timer fires. CFS runs for 0.29s, then sleeps for 0.005s and wakes up
> at 0.295s (its remaining runtime is 0.01s at this point which is < the
> "time till deadline" of 0.005s). Now the runtime of the CFS-server
> will be replenished to the full 3s (due to CBS) and the deadline
> pushed out. The end result is the total runtime that the CFS-server
> actually gets is 0.0595s (though yes it did sleep for 5ms in between,
> still that's tiny -- say if it briefly blocked on a kernel mutex).

Blah, I got lost in decimal points. Here's the example again:

Say CFS-server runtime is 0.3s and period is 1s.

At 0.7s, 0-laxity timer fires. CFS runs for 0.29s, then sleeps for
0.005s and wakes up at 0.295s (its remaining runtime is 0.01s at this
point which is < the "time till deadline" of 0.005s)

Now the runtime of the CFS-server will be replenished to the full 0.3s
(due to CBS) and the deadline
pushed out.

The end result is, the total runtime that the CFS-server actually gets
is 0.595s (though yes it did sleep for 5ms in between, still that's
tiny -- say if it briefly blocked on a kernel mutex). That's almost
double the allocated runtime.

This is just theoretical and I have yet to see if it is actually an
issue in practice.

Thanks.

Powered by blists - more mailing lists