Re: [PATCH v3] sched/deadline: Change the time to replenish runtime for sleep tasks

[Byungchul Park]

On Tue, Feb 28, 2017 at 11:35:15AM +0000, Juri Lelli wrote:
> Hi,
> 
> On 23/02/17 15:14, Byungchul Park wrote:
> > Let's consider the following example.
> > 
> > timeline : o...................o.........o.......o..o
> >            ^                   ^         ^       ^  ^
> >            |                   |         |       |  |
> >        start                   |         |       |  |
> >                 original runtime         |       |  |
> >                      sleep with (-)runtime       |  |
> >                                  original deadline  |
> >                                               wake up
> > 
> > When this task is woken up, a negative runtime should be considered,
> > which means that the task should get penalized when assigning runtime,
> > becasue it already spent more than expected. Current code handles this
> > by replenishing a runtime in hrtimer callback for deadline. But this
> > approach has room for improvement:
> > 
> >    It will be replenished twice unnecessarily if the task sleeps for
> >    long time so that the deadline, assigned in the hrtimer callback,
> >    also passed. In other words, one happens in the callback and the
> >    other happens in update_dl_entiry() when waking it up.
> > 
> > So force to replenish it for sleep tasks when waking it up.
> > 
> > Signed-off-by: Byungchul Park <byungchul.park@xxxxxxx>
> > ---
> >  kernel/sched/deadline.c | 13 ++++++-------
> >  1 file changed, 6 insertions(+), 7 deletions(-)
> > 
> > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > index 27737f3..cb43ce9 100644
> > --- a/kernel/sched/deadline.c
> > +++ b/kernel/sched/deadline.c
> > @@ -498,8 +498,9 @@ static void update_dl_entity(struct sched_dl_entity *dl_se,
> >  	struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
> >  	struct rq *rq = rq_of_dl_rq(dl_rq);
> >  
> > -	if (dl_time_before(dl_se->deadline, rq_clock(rq)) ||
> > -	    dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
> > +	if (dl_time_before(dl_se->deadline, rq_clock(rq)))
> > +		replenish_dl_entity(dl_se, pi_se);
> > +	else if (dl_entity_overflow(dl_se, pi_se, rq_clock(rq))) {
> >  		dl_se->deadline = rq_clock(rq) + pi_se->dl_deadline;
> >  		dl_se->runtime = pi_se->dl_runtime;
> >  	}
> > @@ -621,13 +622,11 @@ static enum hrtimer_restart dl_task_timer(struct hrtimer *timer)
> >  	 *         __dequeue_task_dl()
> >  	 *     prev->on_rq = 0;
> >  	 *
> > -	 * We can be both throttled and !queued. Replenish the counter
> > -	 * but do not enqueue -- wait for our wakeup to do that.
> > +	 * We can be both throttled and !queued. Wait for our wakeup to
> > +	 * replenish runtime and enqueue p.
> >  	 */
> > -	if (!task_on_rq_queued(p)) {
> > -		replenish_dl_entity(dl_se, dl_se);
> 
> Hasn't this patch the same problem we discussed a couple of weeks ago?

No. This patch solves the problem by calling replenish_dl_entity() when
a dl task is woken up.

The problem was that it cannot consider negative runtime if we replenish
the task when it's woken up. So I made replenish_dl_entity() called even
on wake-up path, instead of simple assignment.

IMHO, this patch avoids double-replenishing properly, but adds additional
condition on wake-up path to acheive it. To be honest, I don't think it's
worth as I expected.

Thank you,
Byungchul

> 
> https://marc.info/?l=linux-kernel&m=148699950802995
> 
> Thanks,
> 
> - Juri