Re: [PATCH 1/2] sched: Rework migrate_tasks()

[Kirill Tkhai <tkhai@yandex.ru>]

В Ср, 11/06/2014 в 17:43 +0400, Kirill Tkhai пишет:
> 
> 11.06.2014, 17:15, "Srikar Dronamraju" <srikar@linux.vnet.ibm.com>:
> >>>  * Kirill Tkhai <ktkhai@parallels.com> [2014-06-11 13:52:10]:
> >>>>   Currently migrate_tasks() skips throttled tasks,
> >>>>   because they are not pickable by pick_next_task().
> >>>  Before migrate_tasks() is called, we do call set_rq_offline(), in
> >>>  migration_call().
> >>>
> >>>  Shouldnt this take care of unthrottling the tasks and making sure that
> >>>  they can be picked by pick_next_task().
> >>  If we do this separate for every class, we'll have to do this 3 times.
> >>  Furthermore, deadline class does not have a list of throttled tasks.
> >>  So we'll have to the same as I did: to lock tasklist_lock and to iterate
> >>  throw all of the tasks in the system just to found deadline tasks.
> >
> > I think you misread my comment.
> >
> > Currently migrate_task() gets called from migration_call() and in the
> > migration_call() before migrate_tasks(), set_rq_offline() should put
> > tasks back using unthrottle_cfs_rq().
> >
> > So my question is: Why are these tasks not getting unthrottled
> > through we are calling set_rq_offline? To me set_rq_offline is
> > calling the actual sched class routines to do the needful.
> >
> > I can understand about deadline tasks, because we don't have a deadline
> > But thats the only tasks that we need to fix.
> 
> Hm, I tested that on fair class tasks. They used to disappear from
> /proc/sched_debug and used to hang. I'll check all once again.
> 
> I'm agree with you, if set_rq_offline() already presents, we should use it.
> 
> /me went to clarify why it does not work in my test.

Ok, it looks like the problem is that unthrottled cfs_rq may become throttled
again ;)

__migrate_task()->dequeue_task() calls update_rq_clock(), which throttles
just unthrottle cfs_rq after a first picked task is moved.

Hope, something like above helps, I'll check if there are no other problem
and write again.

---
 kernel/sched/core.c | 8 ++++++++
 1 file changed, 8 insertions(+)
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index d0d1565..b5fd6f0 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -4843,6 +4843,12 @@ static void migrate_tasks(unsigned int dead_cpu)
 	 */
 	update_rq_clock(rq);
 
+	/*
+	 * Prevent rq->clock update from __migrate_task()->dequeue_task()
+	 * to do not throttle sched_class queues and tasks.
+	 */
+	rq->skip_clock_update = 1;
+
 	for ( ; ; ) {
 		/*
 		 * There's this thread running, bail when that's the only
@@ -4864,6 +4870,8 @@ static void migrate_tasks(unsigned int dead_cpu)
 		raw_spin_lock(&rq->lock);
 	}
 
+	rq->skip_clock_update = 0; /* Reset it back */
+
 	rq->stop = stop;
 }