Re: [PATCH v4] hrtimers: Force migrate away hrtimers queued after CPUHP_AP_HRTIMERS_DYING

[Frederic Weisbecker]

Le Wed, Apr 02, 2025 at 06:53:24AM +0000, Kuyo Chang (張建文) a écrit :
> Hi,
> 
> By review the get_nohz_timer_target(), it's probably making an offline
> CPU visible at timer candidates, maybe this patch could fix it?
> 
> 
> [PATCH] sched/core: Exclude offline CPUs from the timer candidates
> 
> The timer target is chosen from the HK_TYPE_KERNEL_NOISE.
> However,the candidate may be an offline CPU,
> so exclude offline CPUs and choose only from online CPUs.
> 
> Signed-off-by: kuyo chang <kuyo.chang@xxxxxxxxxxxx>
> ---
>  kernel/sched/core.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index cfaca3040b2f..efcc2576e622 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -1182,7 +1182,7 @@ int get_nohz_timer_target(void)
>         struct sched_domain *sd;
>         const struct cpumask *hk_mask;
> 
> -       if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE)) {
> +       if (housekeeping_cpu(cpu, HK_TYPE_KERNEL_NOISE) &&
> cpu_online(cpu)) {
>                 if (!idle_cpu(cpu))
>                         return cpu;
>                 default_cpu = cpu;

It can't choose the current CPU because get_target_base() prevents that:

	if (!hrtimer_base_is_online(base)) {
		int cpu = cpumask_any_and(cpu_online_mask, housekeeping_cpumask(HK_TYPE_TIMER));

		return &per_cpu(hrtimer_bases, cpu);
	}

> @@ -1197,13 +1197,16 @@ int get_nohz_timer_target(void)
>                         if (cpu == i)
>                                 continue;
> 
> -                       if (!idle_cpu(i))
> +                       if (!idle_cpu(i) && cpu_online(i))
>                                 return i;

CPUs within the domain hierarchy are guaranteed to be online.
sched_cpu_deactivate() -> cpuset_cpu_inactive(cpu) is supposed to
take care of that. Unless there is another bug lurking here, which is
my suspicion. But it's hard to know as we are dealing with a kernel
with out of tree patches.

>                 }
>         }
> 
> -       if (default_cpu == -1)
> +       if (default_cpu == -1) {
>                 default_cpu =
> housekeeping_any_cpu(HK_TYPE_KERNEL_NOISE);
> +               if (!cpu_online(default_cpu))
> +                       default_cpu = cpumask_any(cpu_online_mask);

housekeeping_any_cpu() only returns online CPUs.

Thanks.