Re: [PATCH] Make sure timers have migrated before killing migration_thread

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2010-05-19 at 14:35 +0530, Amit K. Arora wrote:
> Problem : In a stress test where some heavy tests were running along with
> regular CPU offlining and onlining, a hang was observed. The system seems to
> be hung at a point where migration_call() tries to kill the migration_thread
> of the dying CPU, which just got moved to the current CPU. This migration
> thread does not get a chance to run (and die) since rt_throttled is set to 1
> on current, and it doesn't get cleared as the hrtimer which is supposed to
> reset the rt bandwidth (sched_rt_period_timer) is tied to the CPU being
> offlined.
> 
> Solution : This patch pushes the killing of migration thread to "CPU_POST_DEAD"
> event. By then all the timers (including sched_rt_period_timer) should have got
> migrated (along with other callbacks).
> 
> Alternate Solution considered : Another option considered was to
> increase the priority of the hrtimer cpu offline notifier, such that it
> gets to run before scheduler's migration cpu offline notifier. In this
> way we are sure that the timers will get migrated before migration_call
> tries to kill migration_thread. But, this can have some non-obvious
> implications, suggested Srivatsa.
> 
> Testing : Without the patch the stress tests didn't last for even 12
> hours. And yes, the problem was reproducible. With the patch applied the
> tests ran successfully for more than 48 hours.
> 
> Thanks!
> --
> Regards,
> Amit Arora
> 
>  Signed-off-by: Amit Arora <aarora@in.ibm.com>
>  Signed-off-by: Gautham R Shenoy <ego@in.ibm.com>
> --
> diff -Nuarp linux-2.6.34.org/kernel/sched.c linux-2.6.34/kernel/sched.c
> --- linux-2.6.34.org/kernel/sched.c	2010-05-18 22:56:21.000000000 -0700
> +++ linux-2.6.34/kernel/sched.c	2010-05-18 22:58:31.000000000 -0700
> @@ -5942,14 +5942,26 @@ migration_call(struct notifier_block *nf
>  		cpu_rq(cpu)->migration_thread = NULL;
>  		break;
>  
> +	case CPU_POST_DEAD:
> +		/*
> +		  Bring the migration thread down in CPU_POST_DEAD event,
> +		  since the timers should have got migrated by now and thus
> +		  we should not see a deadlock between trying to kill the
> +		  migration thread and the sched_rt_period_timer.
> +		*/

Faulty comment style that, please use:

 /*
  * text
  *  goes
  *   here
  */

> +		cpuset_lock();
> +		rq = cpu_rq(cpu);
> +		kthread_stop(rq->migration_thread);
> +		put_task_struct(rq->migration_thread);
> +		rq->migration_thread = NULL;
> +		cpuset_unlock();
> +		break;
> +

The other problem is more urgent though, CPU_POST_DEAD runs outside of
the hotplug lock and thus the above becomes a race where we could
possible kill off the migration thread of a newly brought up cpu:

 cpu0 - down 2
 cpu1 - up 2 (allocs a new migration thread, and leaks the old one)
 cpu0 - post_down 2 - frees the migration thread -- oops!