Re: [PATCH]sched/rt: Do not try to push tasks if pinned task switches to RT

[Kirill Tkhai <tkhai@yandex.ru>]

12.03.2014, 14:39, "Nicholas Mc Guire" <der.herr@hofr.at>:
> On Wed, 12 Mar 2014, Steven Rostedt wrote:
>
>>  Peter,
>>
>>  I'm going through my inbox (over a year old), and found this patch from
>>  Kirill. It looks fine to me. You can apply it with my
>>
>>    Acked-by: Steven Rostedt <rostedt@goodmis.org>
>>
>>  -- Steve
>>
>>  [PATCH]sched/rt: Do not try to push tasks if pinned task switches to RT
>>
>>  Just switched pinned task is not able to be pushed. If the rq had had
>>  several RT tasks before they have already been considered as candidates
>>  to be pushed (or pulled).
>>
>>  Signed-off-by: Kirill V Tkhai <tkhai@yandex.ru>
>>  CC: Steven Rostedt <rostedt@goodmis.org>
>>  CC: Ingo Molnar <mingo@kernel.org>
>>  CC: Peter Zijlstra <peterz@infradead.org>
>>  CC: linux-rt-users <linux-rt-users@vger.kernel.org>
>>  ---
>>   kernel/sched/rt.c |    4 ++--
>>   1 files changed, 2 insertions(+), 2 deletions(-)
>>  diff --git a/kernel/sched/rt.c b/kernel/sched/rt.c
>>  index 4e8f0f4..5aab032 100644
>>  --- a/kernel/sched/rt.c
>>  +++ b/kernel/sched/rt.c
>>  @@ -1925,9 +1925,9 @@ static void switched_to_rt(struct rq *rq, struct task_struct *p)
>>            */
>>           if (p->on_rq && rq->curr != p) {
>>   #ifdef CONFIG_SMP
>>  - if (rq->rt.overloaded && push_rt_task(rq) &&
>>  + if (p->nr_cpus_allowed > 1 && rq->rt.overloaded &&
>>                       /* Don't resched if we changed runqueues */
>>  -    rq != task_rq(p))
>>  +    push_rt_task(rq) && rq != task_rq(p))
>>                           check_resched = 0;
>>   #endif /* CONFIG_SMP */
>>                   if (check_resched && p->prio < rq->curr->prio)
>
> would there not need to be a check for p->migrate_disable ?
> push_rt_task() is not checking and so a high prio RT task
> preemting a low prio RT task in a migrate_disable() section
> would actually push it off this cpu ? atleast I did not
> find why that would not happen.

Hi, Nicholas!

p is not rq->curr, so its p->migrate_disable state is already updated and
it can't be pushed (nr_cpus_allowed == 1 and it's not pushable).

(If I understand right, that you worry about this).

Kirill

> thx!
> hofrat