[PATCH 0/3] sched: some fixes for vruntime calculation related to cgroup movement

[PATCH 0/3] sched: some fixes for vruntime calculation related to cgroup movement

[Daisuke Nishimura]

Hi,

These patches fix problems I could see in 3.2-rc2 when testing frequent cgroup movement
under very high load. Without these patches, some processes were not scheduled
(although they were queued into rq)for a very long time(minutes or hours!),
because vruntime of these processes were far bigger than min_vruntime.

Daisuke Nishimura (3):
  sched: fix cgroup movement of newly created process
  sched: fix cgroup movement of forking process
  sched: fix cgroup movement of waking process

 kernel/sched_fair.c |   11 +++++++----
 1 files changed, 7 insertions(+), 4 deletions(-)


Thanks,
Daisuke Nishimura.
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 1/3] sched: fix cgroup movement of newly created process

[Daisuke Nishimura]

There is a small race between do_fork() and sched_move_task(), which is trying
to move the child.

            do_fork()                 sched_move_task()
--------------------------------+---------------------------------
  copy_process()
    sched_fork()
      task_fork_fair()
        -> vruntime of the child is initialized
           based on that of the parent.
  -> we can see the child in "tasks" file now.
                                    task_rq_lock()
                                    task_move_group_fair()
                                      ->child.se.vruntime -= (old)cfs_rq->min_vruntime
                                      ->child.se.vruntime += (new)cfs_rq->min_vruntime
                                    task_rq_unlock()
  wake_up_new_task()
    ...
    enqueue_entity()
      child.se->vruntime += cfs_rq->min_vruntime

As a result, vruntime of the child becomes far bigger than min_vruntime,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by just ignoring such process in task_move_group_fair(),
because the vruntime has already been normalized in task_fork_fair().

Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
---
 kernel/sched_fair.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index 5c9e679..df145a9 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -4922,10 +4922,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
 	 * to another cgroup's rq. This does somewhat interfere with the
 	 * fair sleeper stuff for the first placement, but who cares.
 	 */
-	if (!on_rq)
+	if (!on_rq && p->state != TASK_RUNNING)
 		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
 	set_task_rq(p, task_cpu(p));
-	if (!on_rq)
+	if (!on_rq && p->state != TASK_RUNNING)
 		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
 }
 #endif
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] sched: fix cgroup movement of newly created process

[Paul Turner]

On 12/12/2011 10:57 PM, Daisuke Nishimura wrote:

> There is a small race between do_fork() and sched_move_task(), which is trying
> to move the child.
>
>             do_fork()                 sched_move_task()
> --------------------------------+---------------------------------
>   copy_process()
>     sched_fork()
>       task_fork_fair()
>         -> vruntime of the child is initialized
>            based on that of the parent.


Hmm, so here vruntime of child is actually initialized to vruntime - min_V

>   -> we can see the child in "tasks" file now.
>                                     task_rq_lock()
>                                     task_move_group_fair()


So since on a regular fork we just copy and don't actually go through
the attach muck I'm assuming this is an external actor who's seen the
child in the tasks file and is moving it?

>                                       ->child.se.vruntime -= (old)cfs_rq->min_vruntime
>                                       ->child.se.vruntime += (new)cfs_rq->min_vruntime


This would then add delta min_V between the two cfs_rqs

>                                     task_rq_unlock()
>   wake_up_new_task()
>     ...
>     enqueue_entity()
>       child.se->vruntime += cfs_rq->min_vruntime

>
> As a result, vruntime of the child becomes far bigger than min_vruntime,
> if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.
>
> This patch fixes this problem by just ignoring such process in task_move_group_fair(),
> because the vruntime has already been normalized in task_fork_fair().
>
> Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil-XMD5yJDbdMReXY1tMh2IBg@public.gmane.org>
> ---This would need an explanatory
>  kernel/sched_fair.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index 5c9e679..df145a9 100644
> --- a/kernel/sched_fair.c
> +++ b/kernel/sched_fair.c
> @@ -4922,10 +4922,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
>  	 * to another cgroup's rq. This does somewhat interfere with the
>  	 * fair sleeper stuff for the first placement, but who cares.
>  	 */
> -	if (!on_rq)
> +	if (!on_rq && p->state != TASK_RUNNING)
>  		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;


We also have the choice of keying off something like
!se.sum_exec_runtime here which is reset in sched_fork() which might
be less fragile/make the problem interaction more obvious to.  Either
way this is really a corner case deserving of an explanatory comment.
This is a little icky but I don't have any wildly better ideas.

>  	set_task_rq(p, task_cpu(p));
> -	if (!on_rq)
> +	if (!on_rq && p->state != TASK_RUNNING)
>  		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
>  }
>  #endif


Acked-by: Paul Turner <pjt-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] sched: fix cgroup movement of newly created process

[Daisuke Nishimura]

On Tue, 13 Dec 2011 04:01:21 -0800
Paul Turner <pjt-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org> wrote:

> On 12/12/2011 10:57 PM, Daisuke Nishimura wrote:
> 
> > There is a small race between do_fork() and sched_move_task(), which is trying
> > to move the child.
> >
> >             do_fork()                 sched_move_task()
> > --------------------------------+---------------------------------
> >   copy_process()
> >     sched_fork()
> >       task_fork_fair()
> >         -> vruntime of the child is initialized
> >            based on that of the parent.
> 
> 
> Hmm, so here vruntime of child is actually initialized to vruntime - min_V
> 
> >   -> we can see the child in "tasks" file now.
> >                                     task_rq_lock()
> >                                     task_move_group_fair()
> 
> 
> So since on a regular fork we just copy and don't actually go through
> the attach muck I'm assuming this is an external actor who's seen the
> child in the tasks file and is moving it?
> 
Right.

> >                                       ->child.se.vruntime -= (old)cfs_rq->min_vruntime
> >                                       ->child.se.vruntime += (new)cfs_rq->min_vruntime
> 
> 
> This would then add delta min_V between the two cfs_rqs
> 
> >                                     task_rq_unlock()
> >   wake_up_new_task()
> >     ...
> >     enqueue_entity()
> >       child.se->vruntime += cfs_rq->min_vruntime
> 
> >
> > As a result, vruntime of the child becomes far bigger than min_vruntime,
> > if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.
> >
> > This patch fixes this problem by just ignoring such process in task_move_group_fair(),
> > because the vruntime has already been normalized in task_fork_fair().
> >
> > Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil-XMD5yJDbdMReXY1tMh2IBg@public.gmane.org>
> > ---This would need an explanatory
> >  kernel/sched_fair.c |    4 ++--
> >  1 files changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> > index 5c9e679..df145a9 100644
> > --- a/kernel/sched_fair.c
> > +++ b/kernel/sched_fair.c
> > @@ -4922,10 +4922,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
> >  	 * to another cgroup's rq. This does somewhat interfere with the
> >  	 * fair sleeper stuff for the first placement, but who cares.
> >  	 */
> > -	if (!on_rq)
> > +	if (!on_rq && p->state != TASK_RUNNING)
> >  		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
> 
> 
> We also have the choice of keying off something like
> !se.sum_exec_runtime here which is reset in sched_fork() which might
> be less fragile/make the problem interaction more obvious to.  Either

sum_exec_runtime can be used for a process that has just been created,
but IIUC it cannot be used for a process that is being woken up from
sleeping(in [3/3] case).
Or, do you mean using p->se.sum_exec_runtime for [1/3] and p->state for [3/3] ?
If so, I'll do it in the next post.

> way this is really a corner case deserving of an explanatory comment.
> This is a little icky but I don't have any wildly better ideas.
> 
I'll add comments.

> >  	set_task_rq(p, task_cpu(p));
> > -	if (!on_rq)
> > +	if (!on_rq && p->state != TASK_RUNNING)
> >  		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
> >  }
> >  #endif
> 
> 
> Acked-by: Paul Turner <pjt-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>

Thanks you for your reviews.

Daisuke Nishimura.
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] sched: fix cgroup movement of newly created process

[Peter Zijlstra]

On Tue, 2011-12-13 at 15:57 +0900, Daisuke Nishimura wrote:

>  kernel/sched_fair.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)

you blink you loose, that file doesn't exist anymore.


> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index 5c9e679..df145a9 100644
> --- a/kernel/sched_fair.c
> +++ b/kernel/sched_fair.c
> @@ -4922,10 +4922,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
>  	 * to another cgroup's rq. This does somewhat interfere with the
>  	 * fair sleeper stuff for the first placement, but who cares.
>  	 */
> -	if (!on_rq)
> +	if (!on_rq && p->state != TASK_RUNNING)
>  		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
>  	set_task_rq(p, task_cpu(p));
> -	if (!on_rq)
> +	if (!on_rq && p->state != TASK_RUNNING)
>  		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
>  }
>  #endif

The much saner way of writing that is something like:

 /*
  * Comment explaining stuff..
  */
 if (!on_rq && p->state == TASK_RUNNING)
	on_rq = 1;

 ...


--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 1/3] sched: fix cgroup movement of newly created process

[Daisuke Nishimura]

On Tue, 13 Dec 2011 13:41:09 +0100
Peter Zijlstra <a.p.zijlstra-/NLkJaSkS4VmR6Xm/wNWPw@public.gmane.org> wrote:

> On Tue, 2011-12-13 at 15:57 +0900, Daisuke Nishimura wrote:
> 
> >  kernel/sched_fair.c |    4 ++--
> >  1 files changed, 2 insertions(+), 2 deletions(-)
> 
> you blink you loose, that file doesn't exist anymore.
> 
hmm, indeed.
I'll rebase these patches onto the tip.

> 
> > diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> > index 5c9e679..df145a9 100644
> > --- a/kernel/sched_fair.c
> > +++ b/kernel/sched_fair.c
> > @@ -4922,10 +4922,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
> >  	 * to another cgroup's rq. This does somewhat interfere with the
> >  	 * fair sleeper stuff for the first placement, but who cares.
> >  	 */
> > -	if (!on_rq)
> > +	if (!on_rq && p->state != TASK_RUNNING)
> >  		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
> >  	set_task_rq(p, task_cpu(p));
> > -	if (!on_rq)
> > +	if (!on_rq && p->state != TASK_RUNNING)
> >  		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
> >  }
> >  #endif
> 
> The much saner way of writing that is something like:
> 
>  /*
>   * Comment explaining stuff..
>   */
>  if (!on_rq && p->state == TASK_RUNNING)
> 	on_rq = 1;
> 
>  ...
> 
will do in the next post.

Thanks,
Daisuke Nishimura.
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 2/3] sched: fix cgroup movement of forking process

[Daisuke Nishimura]

There is a small race between task_fork_fair() and sched_move_task(),
which is trying to move the parent.

        task_fork_fair()                 sched_move_task()
--------------------------------+---------------------------------
  cfs_rq = task_cfs_rq(current)
    -> cfs_rq is the "old" one.
  curr = cfs_rq->curr
    -> curr is set to the parent.
                                    task_rq_lock()
                                    dequeue_task()
                                      ->parent.se.vruntime -= (old)cfs_rq->min_vruntime
                                    enqueue_task()
                                      ->parent.se.vruntime += (new)cfs_rq->min_vruntime
                                    task_rq_unlock()
  raw_spin_lock_irqsave(rq->lock)
  se->vruntime = curr->vruntime
    -> vruntime of the child is set to that of the parent
       which has already been updated by sched_move_task().
  se->vruntime -= (old)cfs_rq->min_vruntime.
  raw_spin_unlock_irqrestore(rq->lock)

As a result, vruntime of the child becomes far bigger than expected,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by setting "cfs_rq" and "curr" after holding
the rq->lock.

Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
---
 kernel/sched_fair.c |    7 +++++--
 1 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index df145a9..bdaa4ab 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -4787,14 +4787,17 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
  */
 static void task_fork_fair(struct task_struct *p)
 {
-	struct cfs_rq *cfs_rq = task_cfs_rq(current);
-	struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
+	struct cfs_rq *cfs_rq;
+	struct sched_entity *se = &p->se, *curr;
 	int this_cpu = smp_processor_id();
 	struct rq *rq = this_rq();
 	unsigned long flags;
 
 	raw_spin_lock_irqsave(&rq->lock, flags);
 
+	cfs_rq = task_cfs_rq(current);
+	curr = cfs_rq->curr;
+
 	update_rq_clock(rq);
 
 	if (unlikely(task_cpu(p) != this_cpu)) {
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 2/3] sched: fix cgroup movement of forking process

[Paul Turner]

On Mon, Dec 12, 2011 at 10:58 PM, Daisuke Nishimura
<nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org> wrote:
> There is a small race between task_fork_fair() and sched_move_task(),
> which is trying to move the parent.
>
>        task_fork_fair()                 sched_move_task()
> --------------------------------+---------------------------------
>  cfs_rq = task_cfs_rq(current)
>    -> cfs_rq is the "old" one.
>  curr = cfs_rq->curr
>    -> curr is set to the parent.
>                                    task_rq_lock()
>                                    dequeue_task()
>                                      ->parent.se.vruntime -= (old)cfs_rq->min_vruntime
>                                    enqueue_task()
>                                      ->parent.se.vruntime += (new)cfs_rq->min_vruntime
>                                    task_rq_unlock()
>  raw_spin_lock_irqsave(rq->lock)
>  se->vruntime = curr->vruntime
>    -> vruntime of the child is set to that of the parent
>       which has already been updated by sched_move_task().
>  se->vruntime -= (old)cfs_rq->min_vruntime.
>  raw_spin_unlock_irqrestore(rq->lock)
>
> As a result, vruntime of the child becomes far bigger than expected,
> if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.
>
> This patch fixes this problem by setting "cfs_rq" and "curr" after holding
> the rq->lock.
>
> Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
> ---
>  kernel/sched_fair.c |    7 +++++--
>  1 files changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index df145a9..bdaa4ab 100644
> --- a/kernel/sched_fair.cthis
> +++ b/kernel/sched_fair.c
> @@ -4787,14 +4787,17 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
>  */this
>  static void task_fork_fair(struct task_struct *p)
>  {
> -       struct cfs_rq *cfs_rq = task_cfs_rq(current);
> -       struct sched_entity *se = &p->se, *curr = cfs_rq->curr;

Strictly speaking we're calling
current->sched_class-(*)->task_fork_fair() so we know current is in
sched_fair, which means it has to be cfs_rq->curr.

Because of that this could become be *curr = &current->se and then
cfs_rq_of(curr) below.

But the current healthy paranoia is ok too.

> +       struct cfs_rq *cfs_rq;
> +       struct sched_entity *se = &p->se, *curr;
>        int this_cpu = smp_processor_id();
>        struct rq *rq = this_rq();
>        unsigned long flags;
>
>        raw_spin_lock_irqsave(&rq->lock, flags);
>
> +       cfs_rq = task_cfs_rq(current);
> +       curr = cfs_rq->curr;
> +
>        update_rq_clock(rq);
>
>        if (unlikely(task_cpu(p) != this_cpu)) {
> --
> 1.7.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe cgroups" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Acked-by: Paul Turner <pjt-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 3/3] sched: fix cgroup movement of waking process

[Daisuke Nishimura]

There is a small race between try_to_wake_up() and sched_move_task(), which is trying
to move the process being woken up.

    try_to_wake_up() on CPU0       sched_move_task() on CPU1
--------------------------------+---------------------------------
  raw_spin_lock_irqsave(p->pi_lock)
  task_waking_fair()
    ->p.se.vruntime -= cfs_rq->min_vruntime
  ttwu_queue()
    ->send reschedule IPI to another CPU1
  raw_spin_unlock_irqsave(p->pi_lock)
                                   task_rq_lock()
                                     -> tring to aquire both p->pi_lock and rq->lock
                                        with IRQ disabled
                                   task_move_group_fair()
                                     ->p.se.vruntime -= (old)cfs_rq->min_vruntime
                                     ->p.se.vruntime += (new)cfs_rq->min_vruntime
                                   task_rq_unlock()

                                   (via IPI)
                                   sched_ttwu_pending()
                                     raw_spin_lock(rq->lock)
                                     ttwu_do_activate()
                                       ...
                                       enqueue_entity()
                                         child.se->vruntime += cfs_rq->min_vruntime
                                     raw_spin_unlock(rq->lock)

As a result, vruntime of the process becomes far bigger than min_vruntime,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by just ignoring such process in task_move_group_fair(),
because the vruntime has already been normalized in task_waking_fair().

Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
---
 kernel/sched_fair.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
index bdaa4ab..3feb3a2 100644
--- a/kernel/sched_fair.c
+++ b/kernel/sched_fair.c
@@ -4925,10 +4925,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
 	 * to another cgroup's rq. This does somewhat interfere with the
 	 * fair sleeper stuff for the first placement, but who cares.
 	 */
-	if (!on_rq && p->state != TASK_RUNNING)
+	if (!on_rq && p->state != TASK_RUNNING && p->state != TASK_WAKING)
 		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
 	set_task_rq(p, task_cpu(p));
-	if (!on_rq && p->state != TASK_RUNNING)
+	if (!on_rq && p->state != TASK_RUNNING && p->state != TASK_WAKING)
 		p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
 }
 #endif
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 3/3] sched: fix cgroup movement of waking process

[Paul Turner]

On Mon, Dec 12, 2011 at 10:59 PM, Daisuke Nishimura
<nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org> wrote:
> There is a small race between try_to_wake_up() and sched_move_task(), which is trying
> to move the process being woken up.
>
>    try_to_wake_up() on CPU0       sched_move_task() on CPU1
> --------------------------------+---------------------------------
>  raw_spin_lock_irqsave(p->pi_lock)
>  task_waking_fair()
>    ->p.se.vruntime -= cfs_rq->min_vruntime
>  ttwu_queue()
>    ->send reschedule IPI to another CPU1
>  raw_spin_unlock_irqsave(p->pi_lock)
>                                   task_rq_lock()
>                                     -> tring to aquire both p->pi_lock and rq->lock
>                                        with IRQ disabled
>                                   task_move_group_fair()
>                                     ->p.se.vruntime -= (old)cfs_rq->min_vruntime
>                                     ->p.se.vruntime += (new)cfs_rq->min_vruntime
>                                   task_rq_unlock()
>
>                                   (via IPI)
>                                   sched_ttwu_pending()
>                                     raw_spin_lock(rq->lock)
>                                     ttwu_do_activate()
>                                       ...
>                                       enqueue_entity()
>                                         child.se->vruntime += cfs_rq->min_vruntime
>                                     raw_spin_unlock(rq->lock)
>
> As a result, vruntime of the process becomes far bigger than min_vruntime,
> if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.
>
> This patch fixes this problem by just ignoring such process in task_move_group_fair(),
> because the vruntime has already been normalized in task_waking_fair().
>
> Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
> ---
>  kernel/sched_fair.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index bdaa4ab..3feb3a2 100644
> --- a/kernel/sched_fair.c
> +++ b/kernel/sched_fair.c
> @@ -4925,10 +4925,10 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
>         * to another cgroup's rq. This does somewhat interfere with the
>         * fair sleeper stuff for the first placement, but who cares.
>         */
> -       if (!on_rq && p->state != TASK_RUNNING)
> +       if (!on_rq && p->state != TASK_RUNNING && p->state != TASK_WAKING)

!p->se.sum_exec_runtime is starting to look more attractive here...

>                p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
>        set_task_rq(p, task_cpu(p));
> -       if (!on_rq && p->state != TASK_RUNNING)
> +       if (!on_rq && p->state != TASK_RUNNING && p->state != TASK_WAKING)
>                p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
>  }
>  #endif
> --
> 1.7.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe cgroups" in
> the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH -tip 0/3] sched: some fixes for vruntime calculation related to cgroup movement(v2)

[Daisuke Nishimura]

Hi, all.

This is an updated version based on comments and rebased onto the -tip tree.

These patches fix problems I could see in 3.2-rc2 when testing frequent cgroup movement
under very high load. Without these patches, some processes were not scheduled
(although they were queued into rq)for a very long time(minutes or hours!),
because vruntime of these processes were far bigger than min_vruntime.

Daisuke Nishimura (3):
  sched: fix cgroup movement of newly created process
  sched: fix cgroup movement of forking process
  sched: fix cgroup movement of waking process

 kernel/sched/fair.c |   22 ++++++++++++++++++++--
 1 files changed, 20 insertions(+), 2 deletions(-)

Thanks,
Daisuke Nishimura.

[PATCH 1/3] sched: fix cgroup movement of newly created process

[Daisuke Nishimura]

There is a small race between do_fork() and sched_move_task(), which is
trying to move the child.

            do_fork()                 sched_move_task()
--------------------------------+---------------------------------
  copy_process()
    sched_fork()
      task_fork_fair()
        -> vruntime of the child is initialized
           based on that of the parent.
  -> we can see the child in "tasks" file now.
                                    task_rq_lock()
                                    task_move_group_fair()
                                      -> child.se.vruntime
                                           -= (old)cfs_rq->min_vruntime
                                           += (new)cfs_rq->min_vruntime
                                    task_rq_unlock()
  wake_up_new_task()
    ...
    enqueue_entity()
      child.se.vruntime += cfs_rq->min_vruntime

As a result, vruntime of the child becomes far bigger than min_vruntime,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by just ignoring such process in
task_move_group_fair(), because the vruntime has already been normalized in
task_fork_fair().

Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
---
 kernel/sched/fair.c |   13 +++++++++++++
 1 files changed, 13 insertions(+), 0 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index a4d2b7a..43cadd0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5336,6 +5336,19 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
 	 * to another cgroup's rq. This does somewhat interfere with the
 	 * fair sleeper stuff for the first placement, but who cares.
 	 */
+	/*
+	 * When !on_rq, vruntime of the task has usually NOT been normalized.
+	 * But there are some cases where it has already been normalized:
+	 *
+	 * - Moving a forked child which is waiting for being woken up by
+	 *   wake_up_new_task().
+	 *
+	 * To prevent boost or penalty in the new cfs_rq caused by delta
+	 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
+	 */
+	if (!on_rq && !p->se.sum_exec_runtime)
+		on_rq = 1;
+
 	if (!on_rq)
 		p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
 	set_task_rq(p, task_cpu(p));
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 2/3] sched: fix cgroup movement of forking process

[Daisuke Nishimura]

There is a small race between task_fork_fair() and sched_move_task(),
which is trying to move the parent.

        task_fork_fair()                 sched_move_task()
--------------------------------+---------------------------------
  cfs_rq = task_cfs_rq(current)
    -> cfs_rq is the "old" one.
  curr = cfs_rq->curr
    -> curr is set to the parent.
                                    task_rq_lock()
                                    dequeue_task()
                                      ->parent.se.vruntime -= (old)cfs_rq->min_vruntime
                                    enqueue_task()
                                      ->parent.se.vruntime += (new)cfs_rq->min_vruntime
                                    task_rq_unlock()
  raw_spin_lock_irqsave(rq->lock)
  se->vruntime = curr->vruntime
    -> vruntime of the child is set to that of the parent
       which has already been updated by sched_move_task().
  se->vruntime -= (old)cfs_rq->min_vruntime.
  raw_spin_unlock_irqrestore(rq->lock)

As a result, vruntime of the child becomes far bigger than expected,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by setting "cfs_rq" and "curr" after
holding the rq->lock.

Signed-off-by: Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org>
Acked-by: Paul Turner <pjt-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>
---
 kernel/sched/fair.c |    7 +++++--
 1 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 43cadd0..b1db01b 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5191,8 +5191,8 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
  */
 static void task_fork_fair(struct task_struct *p)
 {
-	struct cfs_rq *cfs_rq = task_cfs_rq(current);
-	struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
+	struct cfs_rq *cfs_rq;
+	struct sched_entity *se = &p->se, *curr;
 	int this_cpu = smp_processor_id();
 	struct rq *rq = this_rq();
 	unsigned long flags;
@@ -5201,6 +5201,9 @@ static void task_fork_fair(struct task_struct *p)
 
 	update_rq_clock(rq);
 
+	cfs_rq = task_cfs_rq(current);
+	curr = cfs_rq->curr;
+
 	if (unlikely(task_cpu(p) != this_cpu)) {
 		rcu_read_lock();
 		__set_task_cpu(p, this_cpu);
-- 
1.7.1

--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH -tip 0/3] sched: some fixes for vruntime calculation related to cgroup movement(v2)

[Daisuke Nishimura]

Someone who have comments ?
If there are some problems that prevent these patches from being accepted,
I'll fix them.

Thanks,
Daisuke Nishimura

On Thu, 15 Dec 2011 14:35:06 +0900
Daisuke Nishimura <nishimura-YQH0OdQVrdy45+QrQBaojngSJqDPrsil@public.gmane.org> wrote:

> Hi, all.
> 
> This is an updated version based on comments and rebased onto the -tip tree.
> 
> These patches fix problems I could see in 3.2-rc2 when testing frequent cgroup movement
> under very high load. Without these patches, some processes were not scheduled
> (although they were queued into rq)for a very long time(minutes or hours!),
> because vruntime of these processes were far bigger than min_vruntime.
> 
> Daisuke Nishimura (3):
>   sched: fix cgroup movement of newly created process
>   sched: fix cgroup movement of forking process
>   sched: fix cgroup movement of waking process
> 
>  kernel/sched/fair.c |   22 ++++++++++++++++++++--
>  1 files changed, 20 insertions(+), 2 deletions(-)
> 
> Thanks,
> Daisuke Nishimura.

[PATCH 3/3] sched: fix cgroup movement of waking process

[Daisuke Nishimura]

There is a small race between try_to_wake_up() and sched_move_task(),
which is trying to move the process being woken up.

    try_to_wake_up() on CPU0       sched_move_task() on CPU1
--------------------------------+---------------------------------
  raw_spin_lock_irqsave(p->pi_lock)
  task_waking_fair()
    ->p.se.vruntime -= cfs_rq->min_vruntime
  ttwu_queue()
    ->send reschedule IPI to CPU1
  raw_spin_unlock_irqsave(p->pi_lock)
                                   task_rq_lock()
                                     -> tring to aquire both p->pi_lock and
                                        rq->lock with IRQ disabled
                                   task_move_group_fair()
                                     -> p.se.vruntime
                                          -= (old)cfs_rq->min_vruntime
                                          += (new)cfs_rq->min_vruntime
                                   task_rq_unlock()

                                   (via IPI)
                                   sched_ttwu_pending()
                                     raw_spin_lock(rq->lock)
                                     ttwu_do_activate()
                                       ...
                                       enqueue_entity()
                                         child.se->vruntime += cfs_rq->min_vruntime
                                     raw_spin_unlock(rq->lock)

As a result, vruntime of the process becomes far bigger than min_vruntime,
if (new)cfs_rq->min_vruntime >> (old)cfs_rq->min_vruntime.

This patch fixes this problem by just ignoring such process in
task_move_group_fair(), because the vruntime has already been normalized in
task_waking_fair().

Signed-off-by: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp>
---
 kernel/sched/fair.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index b1db01b..afd13a5 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -5345,11 +5345,13 @@ static void task_move_group_fair(struct task_struct *p, int on_rq)
 	 *
 	 * - Moving a forked child which is waiting for being woken up by
 	 *   wake_up_new_task().
+	 * - Moving a task which has been woken up by try_to_wake_up() and
+	 *   waiting for actually being woken up by sched_ttwu_pending().
 	 *
 	 * To prevent boost or penalty in the new cfs_rq caused by delta
 	 * min_vruntime between the two cfs_rqs, we skip vruntime adjustment.
 	 */
-	if (!on_rq && !p->se.sum_exec_runtime)
+	if (!on_rq && (!p->se.sum_exec_runtime || p->state == TASK_WAKING))
 		on_rq = 1;
 
 	if (!on_rq)
-- 
1.7.1