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 <[email protected]>
---
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
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 <[email protected]> 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 <[email protected]>
> ---
> 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
>
On 4 February 2015 at 21:28, Xunlei Pang <[email protected]> wrote:
> 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 )
My mistake, the commit should be:
9edfbfed3f544a7830d99b341f0c175995a02950 (tip tree)
Thanks,
Xunlei
Hi, Xunlei,
В Ср, 04/02/2015 в 21:28 +0800, Xunlei Pang пишет:
> I've also sent out this similar patch to yours before,
> and I agree with you on this point :-)
I missed your patch, thanks for the pointing to it.
> 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 <[email protected]> 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 <[email protected]>
> > ---
> > 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
> >
Kirill