Re: [RFC PATCH] workqueue: Unbind workers before sending them to exit()

[Valentin Schneider]

On 21/07/22 11:35, Lai Jiangshan wrote:
>> @@ -1999,6 +2011,16 @@ static void destroy_worker(struct worker *worker)
>>
>>         list_del_init(&worker->entry);
>>         worker->flags |= WORKER_DIE;
>> +
>> +       /*
>> +        * We're sending that thread off to die, so any CPU would do. This is
>> +        * especially relevant for pcpu kworkers affined to an isolated CPU:
>> +        * we'd rather not interrupt an isolated CPU just for a kworker to
>> +        * do_exit().
>> +        */
>> +       if (!(worker->flags & WORKER_UNBOUND))
>> +               unbind_worker(worker);
>> +
>>         wake_up_process(worker->task);
>>  }
>
> destroy_worker() is called with raw_spin_lock_irq(pool->lock), so
> it cannot call the sleepable set_cpus_allowed_ptr().
>
> From __set_cpus_allowed_ptr:
>> * NOTE: the caller must have a valid reference to the task, the
>> * task must not exit() & deallocate itself prematurely. The
>> * call is not atomic; no spinlocks may be held.
>

Heh, I somehow forgot that this is blocking. Now in this particular case I
think pcpu kworkers are "safe" - they shouldn't be running when
destroy_worker() is invoked on them (though AFAICT that is not a "hard"
guarantee), and it doesn't make any sense for them to use
migrate_disable(). Still, yeah, not ideal.

> I think it needs something like task_set_cpumask_possible() which is
> documented as being usable in (raw) spinlocks and set the task's cpumask
> to cpu_possible_mask and let the later ttwu help migrate it to a
> proper non-isolated CPU or let it keep running.
>

I'll see what I can come up with, thanks for the suggestion.