Re: [PATCH] sched/fair: Do not scan non-movable tasks several times

[Konstantin Khorenko]

On 18.12.2023 17:48, Vincent Guittot wrote:
> On Thu, 14 Dec 2023 at 15:42, Konstantin Khorenko
> <khorenko@xxxxxxxxxxxxx> wrote:
> > If busiest rq is small, nr_running < SCHED_NR_MIGRATE_BREAK and all
> > tasks are not movable, detach_tasks() should not iterate more than tasks
> > available in the busiest rq.
> >
> > Previously the (env->loop > env->loop_max) condition prevented us from
>
> It's usually better to give the commit directly when we know it :
> Before commit : b0defa7ae03e ("sched/fair: Make sure to try to detach
> at least one movable task"),
> the (env->loop > env->loop_max) condition prevented us from ...

You are definitely right, added.

> > scanning non-movable tasks more than rq size times, but after we start
> > checking the LBF_ALL_PINNED flag, the "all tasks are not movable" case
> > is under threat.
>
> Fixes: b0defa7ae03e ("sched/fair: Make sure to try to detach at least
> one movable task")

Added this too, thank you.

> > Signed-off-by: Konstantin Khorenko <khorenko@xxxxxxxxxxxxx>
> > ---
> >   kernel/sched/fair.c | 7 ++++++-
> >   1 file changed, 6 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index d7a3c63a2171..faa2a765e899 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -11219,7 +11219,6 @@ static int load_balance(int this_cpu, struct rq *this_rq,
> >                  .dst_rq         = this_rq,
> >                  .dst_grpmask    = group_balance_mask(sd->groups),
> >                  .idle           = idle,
> > -               .loop_break     = SCHED_NR_MIGRATE_BREAK,
> >                  .cpus           = cpus,
> >                  .fbq_type       = all,
> >                  .tasks          = LIST_HEAD_INIT(env.tasks),
> > @@ -11265,6 +11264,12 @@ static int load_balance(int this_cpu, struct rq *this_rq,
> >                   * correctly treated as an imbalance.
> >                   */
> >                  env.loop_max  = min(sysctl_sched_nr_migrate, busiest->nr_running);
> > +               /*
> > +                * If busiest rq is small, nr_running < SCHED_NR_MIGRATE_BREAK
> > +                * and all tasks are not movable, detach_tasks() should not
> > +                * iterate more than tasks available in rq.
> > +                */
> > +               env.loop_break = min(SCHED_NR_MIGRATE_BREAK, busiest->nr_running);
>
> Should it be after more_balance: ?
> In case we do "more_balance:" on a new_dst_cpu and it ends up that
> finally there is no more movable task as we released the lock of
> busiest rq in the meantime ?

Well, both yes and no here.
If we simply move the env.loop_break assignment after "more_balance" label,
we will break the handling LBF_NEED_BREAK case.

But you are right, the new_dst_cpu case should also not just reset loop_break to 
SCHED_NR_MIGRATE_BREAK, but do the same min() calculation.

So i've added one more label before the env.loop_break = min(...) and used it in
the new_dst_cpu case.

> Also you can remove one more superfluous init of loop_break:

Yep, dropped it, thank you.

> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11361,7 +11361,6 @@ static int load_balance(int this_cpu, struct
> rq *this_rq,
>                           */
>                          if (!cpumask_subset(cpus, env.dst_grpmask)) {
>                                  env.loop = 0;
> -                               env.loop_break = SCHED_NR_MIGRATE_BREAK;
>                                  goto redo;
>                          }
>                          goto out_all_pinned;
>
> >   more_balance:
> >                  rq_lock_irqsave(busiest, &rf);
> > --
> > 2.39.3