Re: [PATCH] sched/fair: Check idle_cpu in select_idle_core/cpu()

[Barry Song]

On Mon, Oct 11, 2021 at 3:26 AM Tao Zhou <tao.zhou@xxxxxxxxx> wrote:
>
> Hi Barry,
>
> On Mon, Oct 11, 2021 at 01:19:57AM +1300, Barry Song wrote:
> > On Sun, Oct 10, 2021 at 10:45 PM Tao Zhou <tao.zhou@xxxxxxxxx> wrote:
> > >
> > > Hi Peter,
> > >
> > > On Sun, Oct 10, 2021 at 12:50:57AM +0200, Peter Zijlstra wrote:
> > > > On Sun, Oct 10, 2021 at 02:09:41AM +0800, Tao Zhou wrote:
> > > > > In select_idle_core(), the idle core returned may have no cpu
> > > > > allowed. I think the idle core returned for the task is the one
> > > > > that can be allowed to run. I insist on this semantics.
> > > > >
> > > > > In select_idle_cpu(), if select_idle_core() can not find the
> > > > > idle core, one reason is that the core is not allowed for the
> > > > > task, but the core itself is idle from the point of
> > > > > sds->has_idle_cores. I insist on this semantics.
> > > > >
> > > > > No others, just two additional check.
> > > > > ---
> > > > >  kernel/sched/fair.c | 4 ++--
> > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > >
> > > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > > > > index f6a05d9b5443..a44aca5095d3 100644
> > > > > --- a/kernel/sched/fair.c
> > > > > +++ b/kernel/sched/fair.c
> > > > > @@ -6213,7 +6213,7 @@ static int select_idle_core(struct task_struct *p, int core, struct cpumask *cpu
> > > > >                     *idle_cpu = cpu;
> > > > >     }
> > > > >
> > > > > -   if (idle)
> > > > > +   if (idle && *idle_cpu != -1)
> > > > >             return core;
> > > >
> > > > In that case, core would be nr_cpu_ids (==nr_cpumask_bits), and then the caller checks:
> > > >
> > > >       (unsigned)i < nr_cpumask_bits
> > >
> > > Thank you for reply.
> > >
> > >
> > > If (1)there is no idle core or (2)the idle core has no allowed cpu, we return -1.
> > > Originally, just (1) has happened, we return -1. The (2) is what I want to add.
> >
> > I don't understand (2). before doing
> >         for_each_cpu_wrap(cpu, cpus, target + 1) {
> >                 if (has_idle_core) {
> >                         i = select_idle_core(p, cpu, cpus, &idle_cpu);
> >                         if ((unsigned int)i < nr_cpumask_bits)
> >                                 return i;
> >
> >                 } else {
> >                         if (!--nr)
> >                                 return -1;
> >                         idle_cpu = __select_idle_cpu(cpu, p);
> >                         if ((unsigned int)idle_cpu < nr_cpumask_bits)
> >                                 break;
> >                 }
> >         }
> >
> > to select idle core, we have already done:
> >     cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> >
> > so we are only scanning allowed cpus.
>
> Um.. You read top down.. and you are right.
> The function itself semantics is important to me.
>
> After a secondary recall and not thorough now, I realize that
> cpus_ptr may be changed.
>
>
> See code of this:
>
> static void migrate_disable_switch(struct rq *rq, struct task_struct *p)
> {
>         if (likely(!p->migration_disabled))
>                 return;
>
>         if (p->cpus_ptr != &p->cpus_mask)
>                 return;
>
>         /*
>          * Violates locking rules! see comment in __do_set_cpus_allowed().
>          */
>         __do_set_cpus_allowed(p, cpumask_of(rq->cpu), SCA_MIGRATE_DISABLE);
> }
>
>
> This change is under the light of ->pi_lock.
> That thing is quick to forget to me..
> Not sure I am right. Thank you for remind.
>
> If the cpu_ptr can be changed, you can not depend on the first AND
> operation there.

The explanation doesn't make any sense to me. We are scanning
based on the first AND operation.  select_idle_core() is returning
*idle_cpu based on the cpumask after AND operation.
Even though cpumask can change after select_idle_core() is done
or before select_idle_core() is called, the return value is not wrong.

>
> > >
> > > If we find idle core and has allowed cpu in the core, is it better to return
> > > @*idle_cpu.
> > >
> > >     if (idle && *idle_cpu != -1)
> > >             return *idle_cpu;
> > >
> > > This @*idle_cpu is the allowed cpu in the idle core. We do not promise anything
> > > about the @core(target) is the allowed cpu until we hit in select_task_rq() -->
> > > select_fallback_rq(). And the select_fallback_rq() will return a different cpu
> > > than the @core or @*idle_cpu.
> > >
> > > > >     cpumask_andnot(cpus, cpus, cpu_smt_mask(core));
> > > > > @@ -6324,7 +6324,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> > > > >             }
> > > > >     }
> > > > >
> > > > > -   if (has_idle_core)
> > > > > +   if (has_idle_core && *idle_cpu != -1)
> > > > >             set_idle_cores(target, false);
> > > >
> > > > And this one I'm completely failing, why shouldn't we mark the core as
> > > > non-idle when there is a single idle CPU found? That's just worng.
> > >
> > > When @has_idle_core is true, it implies for all cpu in the core the case
> > > (1) or case (2) has happened. The (1) can be mark as non-idle. I conclude
> > > to contradiction myself last time. The (2) is also seemed to be non-idle.
> > >
> > >
> > > But, I think I am totally wrong because the sds->has_idle_cores is related
> > > to the cpu not task. So, the affinity should not affect the decision of
> > > sds->has_idle_cores.
> > >
> > >
> > >
> > > Thanks,
> > > Tao
> >
> > Thanks
> > barry
>
>
>
> Thanks,
> Tao