Re: [PATCH] psi: Optimize task switch inside shared cgroups

[Johannes Weiner <hannes@cmpxchg.org>]

Hello Chengming,

This patch looks useful to me. A couple of comments below:

On Tue, Feb 09, 2021 at 03:14:13PM +0800, Chengming Zhou wrote:
> The commit 36b238d57172 ("psi: Optimize switching tasks inside shared
> cgroups") only update cgroups whose state actually changes during a
> task switch only in task preempt case, not in task sleep case.
> 
> We actually don't need to clear and set TSK_ONCPU state for common cgroups
> of next and prev task in sleep case, that can save many psi_group_change
> especially when most activity comes from one leaf cgroup.

Can you please make this a bit more concrete? Maybe include this:

sleep before:
psi_dequeue()
  while ((group = iterate_groups(prev)))			# all ancestors
    psi_group_change(prev, .clear=TSK_RUNNING|TSK_ONCPU)
psi_task_switch()
  while ((group = iterate_groups(next)))			# all ancestors
    psi_group_change(next, .set=TSK_ONCPU)

sleep after:
psi_dequeue()
  nop
psi_task_switch()
  while ((group = iterate_groups(next)))			# until (prev & next)
    psi_group_change(next, .set=TSK_ONCPU)
  while ((group = iterate_groups(prev)))			# all ancestors
    psi_group_change(prev, .clear = common ? TSK_RUNNING : TSK_RUNNING|TSK_ONCPU)

When a voluntary sleep switches to another task, we remove one call of
psi_group_change() for every common cgroup ancestor of the two tasks.

> Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
> ---
>  kernel/sched/psi.c   | 27 +++++++++++++++++----------
>  kernel/sched/stats.h | 17 +++--------------
>  2 files changed, 20 insertions(+), 24 deletions(-)
> 
> diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> index 6e46d9eb279b..6061e87089ac 100644
> --- a/kernel/sched/psi.c
> +++ b/kernel/sched/psi.c
> @@ -836,20 +836,27 @@ void psi_task_switch(struct task_struct *prev, struct task_struct *next,
>  		}
>  	}
>  
> -	/*
> -	 * If this is a voluntary sleep, dequeue will have taken care
> -	 * of the outgoing TSK_ONCPU alongside TSK_RUNNING already. We
> -	 * only need to deal with it during preemption.
> -	 */
> -	if (sleep)
> -		return;
> -
>  	if (prev->pid) {
> -		psi_flags_change(prev, TSK_ONCPU, 0);
> +		int clear = 0, set = 0;
> +
> +		if (sleep) {
> +			clear |= TSK_RUNNING;
> +			if (prev->in_iowait)
> +				set |= TSK_IOWAIT;
> +		}

This needs a comment why it's doing psi_dequeue()'s job. How about this?

		/*
		 * When we're going to sleep, psi_dequeue() lets us handle
		 * TSK_RUNNING and TSK_IOWAIT here, where we can combine it
		 * with TSK_ONCPU and save walking common ancestors twice.
		 */
		if (sleep) {
			...

> +		psi_flags_change(prev, clear | TSK_ONCPU, set);
>  
>  		iter = NULL;
>  		while ((group = iterate_groups(prev, &iter)) && group != common)
> -			psi_group_change(group, cpu, TSK_ONCPU, 0, true);
> +			psi_group_change(group, cpu, clear | TSK_ONCPU, set, true);
> +
> +		if (sleep) {
> +			while (group) {
> +				psi_group_change(group, cpu, clear, set, true);
> +				group = iterate_groups(prev, &iter);
> +			}
> +		}

This function is *primarily* about handling TSK_ONCPU and secondarily
optimizes the dequeue. It would be a bit clearer to do:

	int clear = TSK_ONCPU, set = 0;

	...

	/*
	 * TSK_ONCPU is handled up to the common ancestor. If we're tasked
	 * with dequeuing too, finish that for the rest of the hierarchy.
	 */
	if (sleep) {
		clear &= TSK_ONCPU;
		for (; group; group = iterate_groups(prev, &iter))
			psi_group_change(group, cpu, clear, set, true);			
	}


> diff --git a/kernel/sched/stats.h b/kernel/sched/stats.h
> index 9e4e67a94731..2d92c8467678 100644
> --- a/kernel/sched/stats.h
> +++ b/kernel/sched/stats.h
> @@ -84,28 +84,17 @@ static inline void psi_enqueue(struct task_struct *p, bool wakeup)
>  
>  static inline void psi_dequeue(struct task_struct *p, bool sleep)
>  {
> -	int clear = TSK_RUNNING, set = 0;
> -
>  	if (static_branch_likely(&psi_disabled))
>  		return;
>  
>  	if (!sleep) {
> +		int clear = TSK_RUNNING;
> +
>  		if (p->in_memstall)
>  			clear |= TSK_MEMSTALL;
> -	} else {
> -		/*
> -		 * When a task sleeps, schedule() dequeues it before
> -		 * switching to the next one. Merge the clearing of
> -		 * TSK_RUNNING and TSK_ONCPU to save an unnecessary
> -		 * psi_task_change() call in psi_sched_switch().
> -		 */
> -		clear |= TSK_ONCPU;
>  
> -		if (p->in_iowait)
> -			set |= TSK_IOWAIT;
> +		psi_task_change(p, clear, 0);
>  	}

Likewise, this really should have a comment for why it's not handling
TSK_RUNNING to match psi_enqueue()!

	int clear = TSK_RUNNING;

	/*
	 * A voluntary sleep is a dequeue followed by a task switch. To
	 * avoid walking all ancestors twice, psi_task_switch() handles
	 * TSK_RUNNING and TSK_IOWAIT for us when it moves TSK_ONCPU.
	 * Do nothing here.
	 */
	 if (sleep)
		return;

	if (p->in_memstall)
		clear |= TSK_MEMSTALL;

	psi_task_change(p, clear, 0);
	
Thanks