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

[Barry Song]

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.

>
> 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