Re: [PATCH v3] sched/core: Tweak wait_task_inactive() to force dequeue sched_delayed tasks

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Apr 29, 2025 at 08:07:26AM -0700, John Stultz wrote:
> It was reported that in 6.12, smpboot_create_threads() was
> taking much longer then in 6.6.
> 
> I narrowed down the call path to:
>  smpboot_create_threads()
>  -> kthread_create_on_cpu()
>     -> kthread_bind()
>        -> __kthread_bind_mask()
>           ->wait_task_inactive()
> 
> Where in wait_task_inactive() we were regularly hitting the
> queued case, which sets a 1 tick timeout, which when called
> multiple times in a row, accumulates quickly into a long
> delay.
> 
> I noticed disabling the DELAY_DEQUEUE sched feature recovered
> the performance, and it seems the newly create tasks are usually
> sched_delayed and left on the runqueue.
> 
> So in wait_task_inactive() when we see the task
> p->se.sched_delayed, manually dequeue the sched_delayed task
> with DEQUEUE_DELAYED, so we don't have to constantly wait a
> tick.

---

(that is, I'll trim the Changelog a this point, seeing how the rest is
'discussion')

> This seems to work, but I've only lightly tested it, so I'd love
> close review and feedback in case I've missed something in
> wait_task_inactive(), or if there is a simpler alternative
> approach.

There might be. I suspect:

	queued = task_on_rq_queued() && !p->se.sched_delayed;

might just work, but that is indeed pushing things quite far. That gets
us into the position of changing ->cpus_allowed while still enqueued,
and while it all might just work out, it is fairly tricky and not worth
the mental pain.

> NOTE: Peter did highlight[1] his general distaste for the
> kthread_bind() through wait_task_inactive() functions, which
> suggests a deeper rework might be better, but I'm not familiar
> enough with all its users to have a sense of how that might be
> done, and this fix seems to address the problem and be more
> easily backported to 6.12-stable, so I wanted to submit it
> again, as a potentially more short-term solution.

Right, so my distaste is with wait_task_inactive() for basically random
waiting for the condition to become true. The neater solution would be a
completion of sorts, but then we need the dequeue path to do a wakeup
and urgh.

So yeah, this is what we have.

> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index c81cf642dba05..b986cd2fb19b7 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2283,6 +2283,12 @@ unsigned long wait_task_inactive(struct task_struct *p, unsigned int match_state
>  		 * just go back and repeat.
>  		 */
>  		rq = task_rq_lock(p, &rf);
> +		/*
> +		 * If task is sched_delayed, force dequeue it, to avoid always
> +		 * hitting the tick timeout in the queued case
> +		 */
> +		if (p->se.sched_delayed)
> +			dequeue_task(rq, p, DEQUEUE_SLEEP | DEQUEUE_DELAYED);
>  		trace_sched_wait_task(p);
>  		running = task_on_cpu(rq, p);
>  		queued = task_on_rq_queued(p);

Lets just do this. I'll to stick it in queue/sched/core.