Date:	Mon, 6 Jun 2016 14:11:03 +0200
From:	Vincent Guittot <vincent.guittot@...aro.org>
To:	Dietmar Eggemann <dietmar.eggemann@....com>
Cc:	Peter Zijlstra <peterz@...radead.org>,
	linux-kernel <linux-kernel@...r.kernel.org>,
	Ben Segall <bsegall@...gle.com>,
	Morten Rasmussen <morten.rasmussen@....com>,
	Yuyang Du <yuyang.du@...el.com>
Subject: Re: [RFC PATCH 2/3] sched/fair: Sync se with root cfs_rq

Hi Dietmar,

On 1 June 2016 at 21:39, Dietmar Eggemann <dietmar.eggemann@....com> wrote:
> Since task utilization is accrued only on the root cfs_rq, there are a
> couple of places where the se has to be synced with the root cfs_rq:
>
> (1) The root cfs_rq has to be updated in attach_entity_load_avg() for
>     an se representing a task in a tg other than the root tg before
>     the se utilization can be added to it.
>
> (2) The last_update_time value of the root cfs_rq can be higher
>     than the one of the cfs_rq the se is enqueued in. Call
>     __update_load_avg() on the se with the last_update_time value of
>     the root cfs_rq before removing se's utilization from the root
>     cfs_rq in [remove|detach]_entity_load_avg().
>
> In case the difference between the last_update_time value of the cfs_rq
> and the root cfs_rq is smaller than 1024ns, the additional calls to
> __update_load_avg() will bail early.
>
> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@....com>
> ---
>  kernel/sched/fair.c | 21 +++++++++++++++++++--
>  1 file changed, 19 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 212becd3708f..3ae8e79fb687 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2970,6 +2970,8 @@ static inline void update_load_avg(struct sched_entity *se, int update_tg)
>
>  static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se)
>  {
> +       struct cfs_rq* root_cfs_rq;
> +
>         if (!sched_feat(ATTACH_AGE_LOAD))
>                 goto skip_aging;
>
> @@ -2995,8 +2997,16 @@ static void attach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
>         if (!entity_is_task(se))
>                 return;
>
> -       rq_of(cfs_rq)->cfs.avg.util_avg += se->avg.util_avg;
> -       rq_of(cfs_rq)->cfs.avg.util_sum += se->avg.util_sum;
> +       root_cfs_rq = &rq_of(cfs_rq)->cfs;
> +
> +       if (parent_entity(se))
> +               __update_load_avg(cfs_rq_clock_task(root_cfs_rq),
> +                                 cpu_of(rq_of(root_cfs_rq)), &root_cfs_rq->avg,
> +                                 scale_load_down(root_cfs_rq->load.weight),
> +                                 upd_util_cfs_rq(root_cfs_rq), root_cfs_rq);
> +
> +       root_cfs_rq->avg.util_avg += se->avg.util_avg;
> +       root_cfs_rq->avg.util_sum += se->avg.util_sum;

The main issue with flat utilization is that we can't keep the
sched_avg on an sched_entity synced (from a last_update_time pov) with
both the cfs_rq on which load is attached and the root_cfs rq on which
the utilization is attached.

With this additional sync to root cfs_rq in
attach/detach_entity_load_avg and in remove_entity_load_avg, the load
of a sched_entity is no more synced to the time stamp of cfs_rq onto
which it is attached. This  can generate several wrong update of the
load of the latter.
As an example, lets take a task TA that sleeps and move it on TGB
which has not run recently so TGB.avg.last_update_time << root
cfs_rq.avg.last_update_time (a decay of 20ms remove 35% of the load)
When we attach TA to TGB, TA is sync with TGB for attaching it and
then decayed to be synced with root cfs_rq.
If TA is then moved to another task group, we try to sync TA to TGB
but TA is in the future so TA.avg.last_update_time is set to TGB one.
Then, TA load is removed to TGB but TA load has been decayed so only a
part will be effectively subtracted. Then, TA load is synced with root
cfs_rq which means decayed one more time for the same time slot
because TA.avg.last_update_time has been reset to
TGB.avg.last_update_time so we will substract less utilization than
what we should in root cfs_rq.

I think that similar behavior can apply with the removed load.


>
>         cfs_rq_util_change(cfs_rq);
>  }
> @@ -3013,6 +3023,10 @@ static void detach_entity_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *s
>         if (!entity_is_task(se))
>                 return;
>
> +       __update_load_avg(rq_of(cfs_rq)->cfs.avg.last_update_time, cpu_of(rq_of(cfs_rq)),
> +                         &se->avg, se->on_rq * scale_load_down(se->load.weight),
> +                         cfs_rq->curr == se, NULL);
> +
>         rq_of(cfs_rq)->cfs.avg.util_avg =
>             max_t(long, rq_of(cfs_rq)->cfs.avg.util_avg - se->avg.util_avg, 0);
>         rq_of(cfs_rq)->cfs.avg.util_sum =
> @@ -3105,6 +3119,9 @@ void remove_entity_load_avg(struct sched_entity *se)
>         if (!entity_is_task(se))
>                 return;
>
> +       last_update_time = cfs_rq_last_update_time(&rq_of(cfs_rq)->cfs);
> +
> +       __update_load_avg(last_update_time, cpu_of(rq_of(cfs_rq)), &se->avg, 0, 0, NULL);
>         atomic_long_add(se->avg.util_avg, &rq_of(cfs_rq)->cfs.removed_util_avg);
>  }
>
> --
> 1.9.1
>

Powered by blists - more mailing lists