Re: [PATCH v6 1/4] sched/fair: Fix attaching task sched avgs twice when switching to fair or changing task group

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jun 16, 2016 at 06:30:13PM +0200, Vincent Guittot wrote:
> Le Wednesday 15 Jun 2016 à 17:22:17 (+0200), Peter Zijlstra a écrit :
> > On Wed, Jun 15, 2016 at 09:46:53AM +0200, Vincent Guittot wrote:
> > > I still have concerned with this change of the behavior that  attaches
> > > the task only when it is enqueued. The load avg of the task will not
> > > be decayed between the time we move it into its new group until its
> > > enqueue. With this change, a task's load can stay high whereas it has
> > > slept for the last couple of seconds. Then, its load and utilization
> > > is no more accounted anywhere in the mean time just because we have
> > > moved the task which will be enqueued on the same rq.
> > > A task should always be attached to a cfs_rq and its load/utilization
> > > should always be accounted on a cfs_rq and decayed for its sleep
> > > period
> > 
> > OK; so I think I agree with that. Does the below (completely untested,
> > hasn't even been near a compiler) look reasonable?
> > 
> > The general idea is to always attach to a cfs_rq -- through
> > post_init_entity_util_avg(). This covers both the new task isn't
> > attached yet and was never in the fair class to begin with issues.
> 
> Your patch ensures that a task will be attached to a cfs_rq and fix
> the issue raised by Yuyang because of se->avg.last_update_time = 0 at
> init.

> During the test the following message has raised  "BUG: using
> smp_processor_id() in preemptible [00000000] code: systemd/1" because
> of cfs_rq_util_change that is called in attach_entity_load_avg

But per commit:

  b7fa30c9cc48 ("sched/fair: Fix post_init_entity_util_avg() serialization")

that should be with rq->lock held !?

> With patch [1] for the init of cfs_rq side, all use cases will be
> covered regarding the issue linked to a last_update_time set to 0 at
> init [1] https://lkml.org/lkml/2016/5/30/508

Yes, I saw that patch. However, it felt strange to me to set
last_update_time to 1, would that not result in a massive aging because
now - last_update_time ends up as a giant delta?

By doing attach_entity_load_avg() we actually add the initial values to
the cfs_rq avg and set last_update_time to the current time. Ensuring
'regular' contribution and progress.

> > That only leaves a tiny hole in fork() where the task is hashed but
> > hasn't yet passed through wake_up_new_task() in which someone can do
> > cgroup move on it. That is closed with TASK_NEW and can_attach()
> > refusing those tasks.
> > 
> 
> But a new fair task is still detached and attached from/to task_group
> with

> cgroup_post_fork()-->
    ss->fork(child)-->
      cpu_cgroup_fork()-->
        sched_move_task()-->
	  task_move_group_fair().

Blergh, I knew I missed something in there..

> cpu_cgroup_can_attach is not used in this path and sched_move_task do
> the move unconditionally for fair task.
> 
> With your patch, we still have the sequence
> 
> sched_fork()
>     set_task_cpu()
> cgroup_post_fork()--> ... --> task_move_group_fair()
>     detach_task_cfs_rq()
>     set_task_rq()
>     attach_task_cfs_rq()
> wake_up_new_task()
>     select_task_rq() can select a new cpu
>     set_task_cpu()
>         migrate_task_rq_fair if the new_cpu != task_cpu
>              remove_load()
>         __set_task_cpu
>     post_init_entity_util_avg
>         attach_task_cfs_rq()
>     activate_task
>         enqueue_task
> 
> In fact, cpu_cgroup_fork needs a small part of sched_move_task so we
> can just call this small part directly instead sched_move_task. And
> the task doesn't really migrate because it is not yet attached so we
> need the sequence:

> sched_fork()
>     __set_task_cpu()
> cgroup_post_fork()--> ... --> task_move_group_fair()
>     set_task_rq() to set task group and runqueue
> wake_up_new_task()
>     select_task_rq() can select a new cpu
>     __set_task_cpu
>     post_init_entity_util_avg
>         attach_task_cfs_rq()
>     activate_task
>         enqueue_task
> 
> The patch below on top of your patch, ensures that we follow the right sequence :
> 
> ---
>  kernel/sched/core.c | 60 +++++++++++++++++++++++++++++++++++------------------
>  1 file changed, 40 insertions(+), 20 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 7895689a..a21e3dc 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2373,7 +2373,7 @@ int sched_fork(unsigned long clone_flags, struct task_struct *p)
>  	 * Silence PROVE_RCU.
>  	 */
>  	raw_spin_lock_irqsave(&p->pi_lock, flags);
> -	set_task_cpu(p, cpu);
> +	__set_task_cpu(p, cpu);

Indeed, this should not be calling migrate, we're setting the CPU for
the first time.

>  	if (p->sched_class->task_fork)
>  		p->sched_class->task_fork(p);
>  	raw_spin_unlock_irqrestore(&p->pi_lock, flags);
> @@ -2515,7 +2515,7 @@ void wake_up_new_task(struct task_struct *p)
>  	 *  - cpus_allowed can change in the fork path
>  	 *  - any previously selected cpu might disappear through hotplug
>  	 */
> -	set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
> +	__set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));

Similarly here I suppose, since we're not yet properly attached as such.

>  #endif
>  	/* Post initialize new task's util average when its cfs_rq is set */
>  	post_init_entity_util_avg(&p->se);

You were indeed running an 'old' kernel, as this
post_init_entity_util_avg() call should be _after_ the __task_rq_lock().

> @@ -7715,6 +7715,35 @@ void sched_offline_group(struct task_group *tg)
>  	spin_unlock_irqrestore(&task_group_lock, flags);
>  }
>  
> +/* Set task's runqueue and group
> + *     In case of a move between group, we update src and dst group
> + *     thanks to sched_class->task_move_group. Otherwise, we just need to set
> + *     runqueue and group pointers. The task will be attached to the runqueue
> + *     during its wake up.

Broken comment style.

> + */
> +static void sched_set_group(struct task_struct *tsk, bool move)
> +{
> +	struct task_group *tg;
> +
> +	/*
> +	 * All callers are synchronized by task_rq_lock(); we do not use RCU
> +	 * which is pointless here. Thus, we pass "true" to task_css_check()
> +	 * to prevent lockdep warnings.
> +	 */
> +	tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
> +			  struct task_group, css);
> +	tg = autogroup_task_group(tsk, tg);
> +	tsk->sched_task_group = tg;
> +
> +#ifdef CONFIG_FAIR_GROUP_SCHED
> +	if (move && tsk->sched_class->task_move_group)
> +		tsk->sched_class->task_move_group(tsk);
> +	else
> +#endif
> +		set_task_rq(tsk, task_cpu(tsk));
> +
> +}
> +
>  /* change task's runqueue when it moves between groups.
>   *	The caller of this function should have put the task in its new group
>   *	by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
> @@ -7722,7 +7751,6 @@ void sched_offline_group(struct task_group *tg)
>   */
>  void sched_move_task(struct task_struct *tsk)
>  {
> -	struct task_group *tg;
>  	int queued, running;
>  	struct rq_flags rf;
>  	struct rq *rq;
> @@ -7737,22 +7765,7 @@ void sched_move_task(struct task_struct *tsk)
>  	if (unlikely(running))
>  		put_prev_task(rq, tsk);
>  
> +	sched_set_group(tsk, true);
>  
>  	if (unlikely(running))
>  		tsk->sched_class->set_curr_task(rq);
> @@ -8182,7 +8195,14 @@ static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
>  
>  static void cpu_cgroup_fork(struct task_struct *task)
>  {
> -	sched_move_task(task);
> +	struct rq_flags rf;
> +	struct rq *rq;
> +
> +	rq = task_rq_lock(task, &rf);
> +
> +	sched_set_group(task, false);
> +
> +	task_rq_unlock(rq, task, &rf);
>  }

Hmm, yeah, I think you're right.

Let me fold that.