Re: [PATCH v11 8/8] timers: Exclude isolated cpus from timer migration

[Gabriele Monaco <gmonaco@redhat.com>]

On Tue, 2025-08-12 at 13:20 -0400, Waiman Long wrote:
> On 8/8/25 12:01 PM, Gabriele Monaco wrote:
> > +static bool tmigr_should_isolate_cpu(int cpu, void *ignored)
> > +{
> > +	/*
> > +	 * The tick CPU can be marked as isolated by the cpuset
> > code, however
> > +	 * we cannot mark it as unavailable to avoid having no
> > global migrator
> > +	 * for the nohz_full CPUs.
> > +	 */
> > +	return tick_nohz_cpu_hotpluggable(cpu);
> > +}
> We may have to update the cpuset code to fail isolated partition 
> formation if it includes the nohz_full tick CPU as that CPU cannot be
> fully isolated. That will also make this patch simpler.

Good idea, I can check that!

[...]
> >   /*
> >    * NOHZ can only be enabled after clocksource_done_booting().
> > Don't
> >    * bother trashing the cache in the tree before.
> >    */
> >   static int __init tmigr_late_init(void)
> >   {
> > -	return cpuhp_setup_state(CPUHP_AP_TMIGR_ONLINE,
> > "tmigr:online",
> > -				 tmigr_set_cpu_available,
> > tmigr_clear_cpu_available);
> > +	int cpu, ret;
> > +
> > +	ret = cpuhp_setup_state(CPUHP_AP_TMIGR_ONLINE,
> > "tmigr:online",
> > +				tmigr_set_cpu_available,
> > tmigr_clear_cpu_available);
> > +	if (ret)
> > +		return ret;
> > +	/*
> > +	 * The tick CPU may not be marked as available in the
> > above call, this
> > +	 * can occur only at boot as hotplug handlers are not
> > called on the
> > +	 * tick CPU. Force it enabled here.
> > +	 */
> > +	for_each_possible_cpu(cpu) {
> > +		if (!tick_nohz_cpu_hotpluggable(cpu)) {
> > +			ret = smp_call_function_single(
> > +				cpu, tmigr_cpu_unisolate_force,
> > NULL, 1);
> > +			break;
> > +		}
> > +	}
> > +	return ret;
> >   }
> Can you integrate the
> tick_nohz_cpu_hotpluggable/tmigr_should_isolate_cpu check into 
> tmigr_set_cpu_available() instead of special-casing the tick CPU
> here?


I could but I'm afraid Thomas won't like that [1]:
> This is really horribly confusing. This function is also invoked from
> the CPU offline hotplug callback and in context of CPU hotplug this
> check makes absolutely no sense at all.

The whole thing changed a bit since then, but I think his point still
stands, tick_nohz_cpu_hotpluggable is not meaningful from hotplug
callbacks besides that very first invocation.

Or am I missing your point?

Thanks,
Gabriele

[1] - https://lore.kernel.org/lkml/875xgqqrel.ffs@tglx/