Re: [RFC PATCH 0/2] Add queue_*() functions and prefer per-cpu workqueue and flag

[Frederic Weisbecker <frederic@kernel.org>]

Le Wed, Jun 24, 2026 at 09:46:55AM -1000, Tejun Heo a écrit :
> Hello,
> 
> On Tue, Jun 23, 2026 at 05:13:31PM +0200, Marco Crivellari wrote:
> > > - At boot time, allow selecting whether to back them with percpu wqs or
> > >   WQ_AFFN_X unbound ones. Maybe we can even experiment with default to
> > >   WQ_AFFN_CPU.
> > 
> > You mentioned here "at boot time". What about making this also dynamic,
> > "moving" the WQs with WQ_PREFER_PERCPU away from CPU N when N is
> > isolated through a cgroup isolated partition?
> 
> Yeah, being dynamic is better but switching dynamically between percpu and
> unbound workqueues feels like it's going to be complicated. I can't think of
> a simple way to do that. If you can, please be my guest.

I fear it's necessary to have dynamic isolation working correctly.
One way could be to make wq::cpu_pwq point to a pair of pwqs. One
for the isolated -> unbound configuration and another one for the
non_isolated -> per_cpu configuration.

Luckily it's fetched under RCU, so we can have that on one side:

housekeeping_update()
   rcu_assign_pointer(housekeeping.cpumasks[HK_TYPE_DOMAIN], trial);
   housekeeping.flags |= HK_FLAG_DOMAIN; // for housekeeping_enabled()
   synchronize_rcu()
   flush_workqueue(wq_prefer_percpu); // flush both wq->cpu_unbound_pwq and
                                         wq->cpu_pwq

And that on the other side:

__queue_work()
   rcu_read_lock()
   unbound = false;
   if (req_cpu == WORK_CPU_UNBOUND) {
      if (wq->flags & WQ_UNBOUND || (housekeeping_enabled(HK_TYPE_DOMAIN) &&
          wq->flags & WQ_PREFER_PERCPU)) {
         unbound = true;
         cpu = wq_select_unbound_cpu(raw_smp_processor_id());
      } else {
         cpu = raw_smp_processor_id();
      }
   }
   if (unbound)
      pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_unbound_pwq, cpu));
   else
      pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
   ...
   rcu_read_unlock()
   
Thanks.

-- 
Frederic Weisbecker
SUSE Labs