Re: [PATCH 1/2] smpboot: Allow restricting per-CPU threads to a CPU mask

[Daniel Lezcano]

Hi Thomas,

thanks for reviewing the patch

On 11/12/26 18:40, Thomas Gleixner wrote:
> On Tue, Nov 03 2026 at 00:09, Daniel Lezcano wrote:
> > smpboot_register_percpu_thread() creates a thread for every online CPU
> > and registers the descriptor for subsequent CPU hotplug operations.
> > There is currently no way to restrict these threads to a subset of
> > CPUs, even when the user only needs work performed on that subset.
> >
> > For example, an idle injection device targets a specific CPU mask.
> > Creating threads outside that mask consumes resources unnecessarily.
> > Supporting multiple such devices also requires each descriptor to
> > manage only the threads associated with its own CPUs.
> >
> > Add an optional CPU mask to struct smp_hotplug_thread and apply it
> > throughout the thread lifecycle. Registration creates threads only for
> > online CPUs in the mask. CPU bring-up creates threads for selected
> > CPUs that were offline at registration, while parking and unparking
> > are restricted to the same mask.
> >
> > Apply the restriction to thread destruction as well, both during
> > unregistration and registration failure cleanup. This is necessary
> > when descriptors with disjoint masks share the same per-CPU task
> > storage: tearing down one descriptor must not stop or clear threads
> > owned by another descriptor.
> >
> > The caller must keep the mask valid and unchanged until the descriptor
> > is unregistered. Callers sharing per-CPU task storage must ensure their
> > masks do not overlap and serialize ownership changes themselves.
>
> This really wants a built-in check which prevents overlapping
> registrations for the same thread storage. It's easy enough to do:
>
> smpboot_register_percpu_thread()
>
> 	if (!pt->cpumask)
> 		pt->cpu_mask = cpu_possible_mask;
>
> 	guard(cpus_read_lock)();
>          guard(mutex)(&smpboot_threads_lock);
>          list_for_each_entry(ht, &hotplug_threads, list) {
> 		if (ht->store != pt->store)
>                  	continue;
>                  if (WARN_ON_ONCE(cpumask_intersects(ht->cpumask, pt->cpumask)))
>                  	return -EBUSY;
>          }
>                  
> Hmm?
Yes, I agree it will prevent strange things happening

Thanks for the suggestion

  -- Daniel