Re: [PATCH v4 3/5] sched/fair: Switch to task based throttle model

[Aaron Lu <ziqianlu@bytedance.com>]

Hi Peter,

On Thu, Sep 04, 2025 at 03:04:07PM +0800, Aaron Lu wrote:
> On Thu, Sep 04, 2025 at 11:14:31AM +0530, K Prateek Nayak wrote:
> > On 9/4/2025 1:57 AM, Peter Zijlstra wrote:
> > > So this is mostly tasks leaving/joining the class/cgroup. And its
> > > purpose seems to be to remove/add the blocked load component.
> > > 
> > > Previously throttle/unthrottle would {de,en}queue the whole subtree from
> > > PELT, see how {en,de}queue would also stop at throttle.
> > > 
> > > But now none of that is done; PELT is fully managed by the tasks
> > > {de,en}queueing.
> > > 
> > > So I'm thinking that when a task joins fair (deboost from RT or
> > > whatever), we add the blocking load and fully propagate it. If the task
> > > is subject to throttling, that will then happen 'naturally' and it will
> > > dequeue itself again.
> > 
> > That seems like the correct thing to do yes. Those throttled_cfs_rq()
> > checks in propagate_entity_cfs_rq() can be removed then.
> >
> 
> Not sure if I understand correctly, I've come to the below code
> according to your discussion:
>

Does the below diff look sane to you? If so, shall I send a separate
patch on top or fold it in patch3 and send an updated patch3?

Thanks.

> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 3e927b9b7eeb6..97ae561c60f5b 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5234,6 +5234,7 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
>  
>  static void check_enqueue_throttle(struct cfs_rq *cfs_rq);
>  static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq);
> +static inline int cfs_rq_pelt_clock_throttled(struct cfs_rq *cfs_rq);
>  
>  static void
>  requeue_delayed_entity(struct sched_entity *se);
> @@ -5729,6 +5730,11 @@ static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
>  	return cfs_bandwidth_used() && cfs_rq->throttled;
>  }
>  
> +static inline int cfs_rq_pelt_clock_throttled(struct cfs_rq *cfs_rq)
> +{
> +	return cfs_bandwidth_used() && cfs_rq->pelt_clock_throttled;
> +}
> +
>  /* check whether cfs_rq, or any parent, is throttled */
>  static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
>  {
> @@ -6721,6 +6727,11 @@ static inline int cfs_rq_throttled(struct cfs_rq *cfs_rq)
>  	return 0;
>  }
>  
> +static inline int cfs_rq_pelt_clock_throttled(struct cfs_rq *cfs_rq)
> +{
> +	return 0;
> +}
> +
>  static inline int throttled_hierarchy(struct cfs_rq *cfs_rq)
>  {
>  	return 0;
> @@ -13154,10 +13165,7 @@ static void propagate_entity_cfs_rq(struct sched_entity *se)
>  {
>  	struct cfs_rq *cfs_rq = cfs_rq_of(se);
>  
> -	if (cfs_rq_throttled(cfs_rq))
> -		return;
> -
> -	if (!throttled_hierarchy(cfs_rq))
> +	if (!cfs_rq_pelt_clock_throttled(cfs_rq))
>  		list_add_leaf_cfs_rq(cfs_rq);
>  
>  	/* Start to propagate at parent */
> @@ -13168,10 +13176,7 @@ static void propagate_entity_cfs_rq(struct sched_entity *se)
>  
>  		update_load_avg(cfs_rq, se, UPDATE_TG);
>  
> -		if (cfs_rq_throttled(cfs_rq))
> -			break;
> -
> -		if (!throttled_hierarchy(cfs_rq))
> +		if (!cfs_rq_pelt_clock_throttled(cfs_rq))
>  			list_add_leaf_cfs_rq(cfs_rq);
>  	}
>  }
> 
> So this means when a task left/joined a cfs_rq, we will do propagate
> immediately, no matter if the cfs_rq is throttled or has its pelt clock
> stopped or not; if cfs_rq still has pelt clock running, it will be added
> to leaf cfs_rq list to make sure its load can be decayed. If cfs_rq's
> pelt clock is stopped, it will be added to leaf cfs_rq list if necessary
> by enqueue_task_fair() or when it's unthrottled.