Re: [PATCH v3] sched/deadline: Fix bad accounting of nr_running

[Juri Lelli <juri.lelli@gmail.com>]

On 02/19/2014 09:46 AM, Peter Zijlstra wrote:
> On Tue, Feb 18, 2014 at 09:50:12PM -0500, Steven Rostedt wrote:
>>  
>>> Rationale for this odd behavior is that, when a task is throttled, it
>>> is removed only from the dl_rq, but we keep it on_rq (as this is not
>>> a "full dequeue", that is the task is not actually sleeping). But, it
>>> is also true that, while throttled a task behaves like it is sleeping
>>> (e.g., its timer will fire on a new CPU if the old one is dead). So,
>>> Steven's fix sounds also semantically correct.
>>
>> Actually, it seems that I was hitting it again, but this time getting a
>> negative number. OK, after looking at the code a bit more, I think we
>> should update the runqueue nr_running only when the task is officially
>> enqueued and dequeued, and all accounting within, will not touch that
>> number.

This is a different way to get the same result (mildly tested on my box):

---
 kernel/sched/deadline.c |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index 0dd5e09..675dad3 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -837,7 +837,8 @@ static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags)
        if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
                enqueue_pushable_dl_task(rq, p);

-       inc_nr_running(rq);
+       if (!(flags & ENQUEUE_REPLENISH))
+               inc_nr_running(rq);
 }

 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
--

We touch nr_running only when we don't enqueue back as a consequence
of a replenishment.

> 
> But if the task is throttled it should still very much decrement the
> number. There's places that very much rely on nr_running be exactly the
> number of runnable tasks.
> 

This is a different thing, and V2 seemed to implement this behavior
(that's why I said it looked semantically correct).

Thanks,

- Juri