[PATCH] [sched] Don't account time after deadline twice

[PATCH] [sched] Don't account time after deadline twice

[Zhihui Zhang]

Unless we want to double-penalize an overrun task, the time after the deadline
and before the current time is already accounted in the negative dl_se->runtime
value. So we can leave it as is in the case of dmiss && rorun.

Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
---
 kernel/sched/deadline.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
index fc4f98b1..67df0d6 100644
--- a/kernel/sched/deadline.c
+++ b/kernel/sched/deadline.c
@@ -579,10 +579,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
 	 * the next instance. Thus, if we do not account that, we are
 	 * stealing bandwidth from the system at each deadline miss!
 	 */
-	if (dmiss) {
-		dl_se->runtime = rorun ? dl_se->runtime : 0;
-		dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
-	}
+	if (dmiss && !rorun)
+		dl_se->runtime = dl_se->deadline - rq_clock(rq);
 
 	return 1;
 }
-- 
1.8.1.2

Re: [PATCH] [sched] Don't account time after deadline twice

[Peter Zijlstra]

[-- Attachment #1: Type: text/plain, Size: 1121 bytes --]

On Sun, Jun 29, 2014 at 09:26:10PM -0400, Zhihui Zhang wrote:
> Unless we want to double-penalize an overrun task, the time after the deadline
> and before the current time is already accounted in the negative dl_se->runtime
> value. So we can leave it as is in the case of dmiss && rorun.

Juri?

> Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
> ---
>  kernel/sched/deadline.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index fc4f98b1..67df0d6 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -579,10 +579,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
>  	 * the next instance. Thus, if we do not account that, we are
>  	 * stealing bandwidth from the system at each deadline miss!
>  	 */
> -	if (dmiss) {
> -		dl_se->runtime = rorun ? dl_se->runtime : 0;
> -		dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
> -	}
> +	if (dmiss && !rorun)
> +		dl_se->runtime = dl_se->deadline - rq_clock(rq);
>  
>  	return 1;
>  }
> -- 
> 1.8.1.2
> 

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]

Re: [PATCH] [sched] Don't account time after deadline twice

[Juri Lelli]

On Tue, 1 Jul 2014 15:08:16 +0200
Peter Zijlstra <peterz@infradead.org> wrote:

> On Sun, Jun 29, 2014 at 09:26:10PM -0400, Zhihui Zhang wrote:
> > Unless we want to double-penalize an overrun task, the time after the deadline
> > and before the current time is already accounted in the negative dl_se->runtime
> > value. So we can leave it as is in the case of dmiss && rorun.
> 
> Juri?
> 
> > Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
> > ---
> >  kernel/sched/deadline.c | 6 ++----
> >  1 file changed, 2 insertions(+), 4 deletions(-)
> > 
> > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > index fc4f98b1..67df0d6 100644
> > --- a/kernel/sched/deadline.c
> > +++ b/kernel/sched/deadline.c
> > @@ -579,10 +579,8 @@ int dl_runtime_exceeded(struct rq *rq, struct sched_dl_entity *dl_se)
> >  	 * the next instance. Thus, if we do not account that, we are
> >  	 * stealing bandwidth from the system at each deadline miss!
> >  	 */
> > -	if (dmiss) {
> > -		dl_se->runtime = rorun ? dl_se->runtime : 0;

If we didn't return 0 before, we are going to throttle (or replenish)
the entity, and you want runtime to be <=0. So, this is needed. 

> > -		dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
> > -	}

A little pessimism in some cases, due to the fact that we use both
rq_clock and rq_clock_task (for the budget).

Thanks,

- Juri

> > +	if (dmiss && !rorun)
> > +		dl_se->runtime = dl_se->deadline - rq_clock(rq);
> >  
> >  	return 1;
> >  }
> > -- 
> > 1.8.1.2
> > 

Re: [PATCH] [sched] Don't account time after deadline twice

[Juri Lelli]

On Wed, 2 Jul 2014 19:44:04 -0400
Zhihui Zhang <zzhsuny@gmail.com> wrote:

> My point is that rq_clock(rq) - dl_se->deadline is already part of
> dl_se->runtime, which is decremented before calling dl_runtime_exceeded().

But, we decrement dl_se->runtime looking at rq_clock_task(rq), that is
in general <= rq_clock(rq), that we use to handle deadlines. So, if we
do like you suggest, in some cases we could end up stealing some
bandwidth from the system. Indeed, we prefer some pessimism here.

Thanks,

- Juri

> So the following line is not needed in the case of both overrun and missing
> deadline:
> 
>     dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
> 
> Or did I miss anything?
> 
> thanks,
> 
> 
> On Tue, Jul 1, 2014 at 9:59 AM, Juri Lelli <juri.lelli@gmail.com> wrote:
> 
> > On Tue, 1 Jul 2014 15:08:16 +0200
> > Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > > On Sun, Jun 29, 2014 at 09:26:10PM -0400, Zhihui Zhang wrote:
> > > > Unless we want to double-penalize an overrun task, the time after the
> > deadline
> > > > and before the current time is already accounted in the negative
> > dl_se->runtime
> > > > value. So we can leave it as is in the case of dmiss && rorun.
> > >
> > > Juri?
> > >
> > > > Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
> > > > ---
> > > >  kernel/sched/deadline.c | 6 ++----
> > > >  1 file changed, 2 insertions(+), 4 deletions(-)
> > > >
> > > > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> > > > index fc4f98b1..67df0d6 100644
> > > > --- a/kernel/sched/deadline.c
> > > > +++ b/kernel/sched/deadline.c
> > > > @@ -579,10 +579,8 @@ int dl_runtime_exceeded(struct rq *rq, struct
> > sched_dl_entity *dl_se)
> > > >      * the next instance. Thus, if we do not account that, we are
> > > >      * stealing bandwidth from the system at each deadline miss!
> > > >      */
> > > > -   if (dmiss) {
> > > > -           dl_se->runtime = rorun ? dl_se->runtime : 0;
> >
> > If we didn't return 0 before, we are going to throttle (or replenish)
> > the entity, and you want runtime to be <=0. So, this is needed.
> >
> > > > -           dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
> > > > -   }
> >
> > A little pessimism in some cases, due to the fact that we use both
> > rq_clock and rq_clock_task (for the budget).
> >
> > Thanks,
> >
> > - Juri
> >
> > > > +   if (dmiss && !rorun)
> > > > +           dl_se->runtime = dl_se->deadline - rq_clock(rq);
> > > >
> > > >     return 1;
> > > >  }
> > > > --
> > > > 1.8.1.2
> > > >
> >

Re: [PATCH] [sched] Don't account time after deadline twice

[Zhihui Zhang]

We calculate difference between two readings of a clock to see how
much time has elapsed.  Part of the time between rq_clock(rq) -
dl_se->deadline can indeed be accounted for by reading a different
clock
(i.e., rq_clock_task()) if the task was running during the period.
And that is how dl_se->runtime is obtained. After all, both clocks are
running independently, right? Furthermore, the caller of
dl_runtime_exceeded() will still use rq_clock() and dl_se->deadline to
determine if we throttle or replenish. Anyway, I have failed to see
any steal of time.  Could you please give a concrete example (perhaps
with numbers)?

thanks,

-Zhihui

On Thu, Jul 3, 2014 at 5:50 AM, Juri Lelli <juri.lelli@gmail.com> wrote:
> On Wed, 2 Jul 2014 19:44:04 -0400
> Zhihui Zhang <zzhsuny@gmail.com> wrote:
>
>> My point is that rq_clock(rq) - dl_se->deadline is already part of
>> dl_se->runtime, which is decremented before calling dl_runtime_exceeded().
>
> But, we decrement dl_se->runtime looking at rq_clock_task(rq), that is
> in general <= rq_clock(rq), that we use to handle deadlines. So, if we
> do like you suggest, in some cases we could end up stealing some
> bandwidth from the system. Indeed, we prefer some pessimism here.
>
> Thanks,
>
> - Juri
>
>> So the following line is not needed in the case of both overrun and missing
>> deadline:
>>
>>     dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
>>
>> Or did I miss anything?
>>
>> thanks,
>>
>>
>> On Tue, Jul 1, 2014 at 9:59 AM, Juri Lelli <juri.lelli@gmail.com> wrote:
>>
>> > On Tue, 1 Jul 2014 15:08:16 +0200
>> > Peter Zijlstra <peterz@infradead.org> wrote:
>> >
>> > > On Sun, Jun 29, 2014 at 09:26:10PM -0400, Zhihui Zhang wrote:
>> > > > Unless we want to double-penalize an overrun task, the time after the
>> > deadline
>> > > > and before the current time is already accounted in the negative
>> > dl_se->runtime
>> > > > value. So we can leave it as is in the case of dmiss && rorun.
>> > >
>> > > Juri?
>> > >
>> > > > Signed-off-by: Zhihui Zhang <zzhsuny@gmail.com>
>> > > > ---
>> > > >  kernel/sched/deadline.c | 6 ++----
>> > > >  1 file changed, 2 insertions(+), 4 deletions(-)
>> > > >
>> > > > diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
>> > > > index fc4f98b1..67df0d6 100644
>> > > > --- a/kernel/sched/deadline.c
>> > > > +++ b/kernel/sched/deadline.c
>> > > > @@ -579,10 +579,8 @@ int dl_runtime_exceeded(struct rq *rq, struct
>> > sched_dl_entity *dl_se)
>> > > >      * the next instance. Thus, if we do not account that, we are
>> > > >      * stealing bandwidth from the system at each deadline miss!
>> > > >      */
>> > > > -   if (dmiss) {
>> > > > -           dl_se->runtime = rorun ? dl_se->runtime : 0;
>> >
>> > If we didn't return 0 before, we are going to throttle (or replenish)
>> > the entity, and you want runtime to be <=0. So, this is needed.
>> >
>> > > > -           dl_se->runtime -= rq_clock(rq) - dl_se->deadline;
>> > > > -   }
>> >
>> > A little pessimism in some cases, due to the fact that we use both
>> > rq_clock and rq_clock_task (for the budget).
>> >
>> > Thanks,
>> >
>> > - Juri
>> >
>> > > > +   if (dmiss && !rorun)
>> > > > +           dl_se->runtime = dl_se->deadline - rq_clock(rq);
>> > > >
>> > > >     return 1;
>> > > >  }
>> > > > --
>> > > > 1.8.1.2
>> > > >
>> >