Re: [PATCH v3 1/4] sched/rt: Check to push the task away after its affinity was changed

[Peter Zijlstra]

On Tue, May 12, 2015 at 10:46:41PM +0800, Xunlei Pang wrote:
> @@ -2278,6 +2279,20 @@ static void set_cpus_allowed_rt(struct task_struct *p,
>  	}
>  
>  	update_rt_migration(&rq->rt);
> +
> +check_push:
> +	if (weight > 1 &&
> +	    !task_running(rq, p) &&
> +	    !test_tsk_need_resched(rq->curr) &&
> +	    !cpumask_subset(new_mask, &p->cpus_allowed)) {
> +		/* Update new affinity and try to push. */
> +		cpumask_copy(&p->cpus_allowed, new_mask);
> +		p->nr_cpus_allowed = weight;
> +		push_rt_tasks(rq);
> +		return true;
> +	}
> +
> +	return false;
>  }

I think this is broken; push_rt_tasks() will do double_rq_lock() which
will drop rq->lock.

This means load-balancing can come in and move our task p; in fact,
push_rt_task() can do exactly that -- after all that was the point of
this patch.

_However_ this means that after calling ->set_cpus_allowed() we must not
assume @p is on @rt, yet we do. Look at __set_cpus_allowed_ptr(), we'll
call move_queued_task() if (!running || waking) && on_rq, and
move_queued_task() happily calls dequeue_task(rq, p), which will go
*boom*.

I currently do not have a better idea than to repurpose the PUSH_IPI
stuff, that is, send a self IPI to go do the push or somesuch. Lemme
stare at this a little more.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/