lkml.org 
[lkml]   [2026]   [May]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 2/5] sched/fair: Attach sched_domain_shared to sd_asym_cpucapacity
On Wed, 6 May 2026 at 12:20, K Prateek Nayak <kprateek.nayak@amd.com> wrote:
>
> Hello Vincent,
>
> On 5/6/2026 3:15 PM, Vincent Guittot wrote:
> >> @@ -7925,7 +7925,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> >> struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_rq_mask);
> >> int i, cpu, idle_cpu = -1, nr = INT_MAX;
> >>
> >> - if (sched_feat(SIS_UTIL)) {
> >> + if (sched_feat(SIS_UTIL) && sd->shared) {
> >
> > If shared is attached to sd_asym_cpucapacity instead of sd_llc we
> > should never reach this point. Or I'm missing a case ?
>
> So the hotpulg might race with a wakeup like:
>
> claim_asym_sched_domain_shared()
> init_sched_domain_shared(d, sd_asym);
> return true;
> update_top_cache_domain()
> rcu_assign_pointer(sd_llc, sd);
> ... select_idle_sibling()
> sd = rcu_dereference_all(sd_asym_cpucapacity)
> /* sd_asym_cpucapacity still hasn't been updated */
> if (sd /* NULL */) { ... }
> sd = rcu_dereference_all(sd_llc); /* Valid */
> select_idle_cpu(sd)
> rcu_assign_pointer(sd_asym_cpucapacity, sd); sd->shared /* NULL */
>
>
> This prevents that rare race where a remote CPU will see sd_llc
> before sd_asym is published and take the !ASYM wakeup route only
> to find sd->shared is NULL since sd_asym has claimed it.

fair enough

>
> >
> >> /*
> >> * Increment because !--nr is the condition to stop scan.
> >> *
>
> --
> Thanks and Regards,
> Prateek
>

\
 
 \ /
  Last update: 2026-05-06 12:31    [W:0.070 / U:1.571 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and my Meterkast|Read the blog