[PATCH RFC] sched/fair: simplfy the work when reweighting entity

[PATCH RFC] sched/fair: simplfy the work when reweighting entity

[Jiang Biao]

If a se is on_rq when reweighting entity, all we need should be
updating the load of cfs_rq, other dequeue/enqueue works could be
redundant, such as,
* account_numa_dequeue/account_numa_enqueue
* list_del/list_add from/into cfs_tasks
* nr_running--/nr_running++

Just simplfy the work. Could be helpful for the hot path.

Signed-off-by: Jiang Biao <benbjiang@tencent.com>
---
 kernel/sched/fair.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 04fa8dbcfa4d..18a8fc7bd0de 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3086,7 +3086,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
 		/* commit outstanding execution time */
 		if (cfs_rq->curr == se)
 			update_curr(cfs_rq);
-		account_entity_dequeue(cfs_rq, se);
+		update_load_sub(&cfs_rq->load, se->load.weight);
 	}
 	dequeue_load_avg(cfs_rq, se);
 
@@ -3102,7 +3102,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
 
 	enqueue_load_avg(cfs_rq, se);
 	if (se->on_rq)
-		account_entity_enqueue(cfs_rq, se);
+		update_load_add(&cfs_rq->load, se->load.weight);
 
 }
 
-- 
2.21.0

Re: [PATCH RFC] sched/fair: simplfy the work when reweighting entity

[Dietmar Eggemann]

On 04/08/2020 09:12, Jiang Biao wrote:
> If a se is on_rq when reweighting entity, all we need should be
> updating the load of cfs_rq, other dequeue/enqueue works could be
> redundant, such as,
> * account_numa_dequeue/account_numa_enqueue
> * list_del/list_add from/into cfs_tasks
> * nr_running--/nr_running++

I think this could make sense. Have you spotted a code path where this
gives you a change?

I guess only for a task on the rq, so: entity_is_task(se) && se->on_rq

> Just simplfy the work. Could be helpful for the hot path.

IMHO hotpath is update_cfs_group() -> reweight_entity() but this is only
called for '!entity_is_task(se)'.

See

 3290         if (!gcfs_rq)
 3291                 return;

in update_cfs_group().

The 'entity_is_task(se)' case is

set_load_weight(struct task_struct *p, ...) -> reweight_task(p, ...) ->
reweight_entity(..., &p->se, ...)

but here !se->on_rq.

> Signed-off-by: Jiang Biao <benbjiang@tencent.com>
> ---
>  kernel/sched/fair.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 04fa8dbcfa4d..18a8fc7bd0de 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -3086,7 +3086,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>  		/* commit outstanding execution time */
>  		if (cfs_rq->curr == se)
>  			update_curr(cfs_rq);
> -		account_entity_dequeue(cfs_rq, se);
> +		update_load_sub(&cfs_rq->load, se->load.weight);
>  	}
>  	dequeue_load_avg(cfs_rq, se);
>  
> @@ -3102,7 +3102,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>  
>  	enqueue_load_avg(cfs_rq, se);
>  	if (se->on_rq)
> -		account_entity_enqueue(cfs_rq, se);
> +		update_load_add(&cfs_rq->load, se->load.weight);
>  
>  }

Re: [PATCH RFC] sched/fair: simplfy the work when reweighting entity

[benbjiang(蒋彪)]

> On Aug 6, 2020, at 12:21 AM, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
> 
> On 04/08/2020 09:12, Jiang Biao wrote:
>> If a se is on_rq when reweighting entity, all we need should be
>> updating the load of cfs_rq, other dequeue/enqueue works could be
>> redundant, such as,
>> * account_numa_dequeue/account_numa_enqueue
>> * list_del/list_add from/into cfs_tasks
>> * nr_running--/nr_running++
> 
> I think this could make sense. Have you spotted a code path where this
> gives you a change?
> 
> I guess only for a task on the rq, so: entity_is_task(se) && se->on_rq
Yes, you're right. No other code path I spotted except what you list below.

> 
>> Just simplfy the work. Could be helpful for the hot path.
> 
> IMHO hotpath is update_cfs_group() -> reweight_entity() but this is only
> called for '!entity_is_task(se)'.
> 
> See
> 
> 3290         if (!gcfs_rq)
> 3291                 return;
> 
> in update_cfs_group().
Yes, It is.
But *nr_running--/nr_running++* works are still redundant for
‘!entity_is_task(se)' hot path. :)
Besides, I guess we could simplify the logic and make it cleaner and
more readable with this patch.
If it could make sense to you, would you mind if I resend the patch
with the commit log amended?

> 
> The 'entity_is_task(se)' case is
> 
> set_load_weight(struct task_struct *p, ...) -> reweight_task(p, ...) ->
> reweight_entity(..., &p->se, ...)
> 
> but here !se->on_rq.
Yes, indeed.

Thanks a lot for your comments.
Regards,
Jiang

> 
>> Signed-off-by: Jiang Biao <benbjiang@tencent.com>
>> ---
>> kernel/sched/fair.c | 4 ++--
>> 1 file changed, 2 insertions(+), 2 deletions(-)
>> 
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 04fa8dbcfa4d..18a8fc7bd0de 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -3086,7 +3086,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>> 		/* commit outstanding execution time */
>> 		if (cfs_rq->curr == se)
>> 			update_curr(cfs_rq);
>> -		account_entity_dequeue(cfs_rq, se);
>> +		update_load_sub(&cfs_rq->load, se->load.weight);
>> 	}
>> 	dequeue_load_avg(cfs_rq, se);
>> 
>> @@ -3102,7 +3102,7 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>> 
>> 	enqueue_load_avg(cfs_rq, se);
>> 	if (se->on_rq)
>> -		account_entity_enqueue(cfs_rq, se);
>> +		update_load_add(&cfs_rq->load, se->load.weight);
>> 
>> }
> 

Re: [PATCH RFC] sched/fair: simplfy the work when reweighting entity

[Dietmar Eggemann]

On 06/08/2020 04:42, benbjiang(蒋彪) wrote:
> 
> 
>> On Aug 6, 2020, at 12:21 AM, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
>>
>> On 04/08/2020 09:12, Jiang Biao wrote:
>>> If a se is on_rq when reweighting entity, all we need should be
>>> updating the load of cfs_rq, other dequeue/enqueue works could be
>>> redundant, such as,
>>> * account_numa_dequeue/account_numa_enqueue
>>> * list_del/list_add from/into cfs_tasks
>>> * nr_running--/nr_running++
>>
>> I think this could make sense. Have you spotted a code path where this
>> gives you a change?
>>
>> I guess only for a task on the rq, so: entity_is_task(se) && se->on_rq
> Yes, you're right. No other code path I spotted except what you list below.
> 
>>
>>> Just simplfy the work. Could be helpful for the hot path.
>>
>> IMHO hotpath is update_cfs_group() -> reweight_entity() but this is only
>> called for '!entity_is_task(se)'.
>>
>> See
>>
>> 3290         if (!gcfs_rq)
>> 3291                 return;
>>
>> in update_cfs_group().
> Yes, It is.
> But *nr_running--/nr_running++* works are still redundant for
> ‘!entity_is_task(se)' hot path. :)

True.

> Besides, I guess we could simplify the logic and make it cleaner and
> more readable with this patch.

Yes.

> If it could make sense to you, would you mind if I resend the patch
> with the commit log amended?

LGTM so why not?

>> The 'entity_is_task(se)' case is
>>
>> set_load_weight(struct task_struct *p, ...) -> reweight_task(p, ...) ->
>> reweight_entity(..., &p->se, ...)
>>
>> but here !se->on_rq.
> Yes, indeed.

[...]