Re: [PATCH] sched/deadline: Add update_rq_clock() in yield_task_dl()

[Xunlei Pang]

Hi Kirill,

I've also sent out this similar patch to yours before,
and I agree with you on this point :-)

But with the latest modification which peter made,
( see: cebde6d681aa45f96111cfcffc1544cf2a0454ff )
I think it would be better to add a skip operation
after update_curr_dl(rq) like below:

---
 kernel/sched/deadline.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index e5db8c6..5648e62 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -915,7 +915,11 @@ static void yield_task_dl(struct rq *rq)
                rq->curr->dl.dl_yielded = 1;
                p->dl.runtime = 0;
        }
+
+       update_rq_clock(rq);
        update_curr_dl(rq);
+       /* Will go to schedule(), to avoid another clock update. */
+       rq_clock_skip_update(rq, true);
 }

#ifdef CONFIG_SMP
--
1.9.1

On 19 December 2014 at 00:31, Xunlei Pang <pang.xunlei@xxxxxxxxxx> wrote:
> yield_task_dl() calls update_curr_dl() which calls start_dl_timer()
> to throttle current task. But yield_task_dl() doesn't update the rq
> clock which will cause start_dl_timer() to set the wrong dl_timer
> which may be much later than current deadline time.
>
> For instance, in systems with 100HZ tick, if there're few dl-tasks,
> then rq clock will mainly depend on the tick to update its value.
> If we choose the CONFIG_NO_HZ_FULL_ALL feature, things may get even
> worse.
>
> Moreover, there're some statistics in update_curr_dl() also relying
> on rq clock, like rq_clock_task().
>
> Thus, in yield_task_dl(), we should add update_rq_clock() before
> update_curr_dl().
>
> Signed-off-by: Xunlei Pang <pang.xunlei@xxxxxxxxxx>
> ---
>  kernel/sched/deadline.c | 4 ++++
>  1 file changed, 4 insertions(+)
>
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index e5db8c6..5648e62 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -915,7 +915,11 @@ static void yield_task_dl(struct rq *rq)
>                 rq->curr->dl.dl_yielded = 1;
>                 p->dl.runtime = 0;
>         }
> +
> +       update_rq_clock(rq);
>         update_curr_dl(rq);
> +       /* Will go to schedule(), to avoid another clock update. */
> +       rq->skip_clock_update = 1;
>  }
>
>  #ifdef CONFIG_SMP
> --
> 1.9.1
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/