In reweight_entity(), if !se->on_rq (e.g. when we are reweighting a
sleeping task), we should sync the load_avg of se to cfs_rq before calling
dequeue_load_avg(). Otherwise, the load_avg of this se can be inaccurate.

Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
---
 kernel/sched/fair.c | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 9057584ec06d..2807f6e72ad1 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3782,6 +3782,8 @@ static void reweight_eevdf(struct sched_entity *se, u64 avruntime,
 	se->deadline = avruntime + vslice;
 }
 
+static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags);
+
 static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
 			    unsigned long weight)
 {
@@ -3795,7 +3797,11 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
 		if (!curr)
 			__dequeue_entity(cfs_rq, se);
 		update_load_sub(&cfs_rq->load, se->load.weight);
+	} else {
+		/* Sync with the cfs_rq before removing our load_avg */
+		update_load_avg(cfs_rq, se, 0);
 	}
+
 	dequeue_load_avg(cfs_rq, se);
 
 	if (se->on_rq) {
-- 
2.20.1

On 2024/7/16 23:08, Chuyi Zhou wrote:
> In reweight_entity(), if !se->on_rq (e.g. when we are reweighting a
> sleeping task), we should sync the load_avg of se to cfs_rq before calling
> dequeue_load_avg(). Otherwise, the load_avg of this se can be inaccurate.

Good catch!

> 
> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
> ---
>   kernel/sched/fair.c | 6 ++++++
>   1 file changed, 6 insertions(+)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 9057584ec06d..2807f6e72ad1 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -3782,6 +3782,8 @@ static void reweight_eevdf(struct sched_entity *se, u64 avruntime,
>   	se->deadline = avruntime + vslice;
>   }
>   
> +static inline void update_load_avg(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags);
> +
>   static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>   			    unsigned long weight)
>   {
> @@ -3795,7 +3797,11 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
>   		if (!curr)
>   			__dequeue_entity(cfs_rq, se);
>   		update_load_sub(&cfs_rq->load, se->load.weight);
> +	} else {
> +		/* Sync with the cfs_rq before removing our load_avg */
> +		update_load_avg(cfs_rq, se, 0);

I think it's suboptimal to update_load_avg() here unconditionally.

Because reweight_entity() has two types of usages:

1. group se, which uses reweight_entity() in update_cfs_group(), which
should already update_load_avg(), so should have no problem.

2. task se, which uses reweight_entity() in reweight_task(), which 
should be fixed for sleep task entity as you described above.

So IMHO, we should only update_load_avg() or sync_entity_load_avg() in
reweight_task(), right?

>   	}
> +
>   	dequeue_load_avg(cfs_rq, se);
>   
>   	if (se->on_rq) {

在 2024/7/17 19:18, Chengming Zhou 写道:
> On 2024/7/16 23:08, Chuyi Zhou wrote:
>> In reweight_entity(), if !se->on_rq (e.g. when we are reweighting a
>> sleeping task), we should sync the load_avg of se to cfs_rq before 
>> calling
>> dequeue_load_avg(). Otherwise, the load_avg of this se can be inaccurate.
> 
> Good catch!
> 
>>
>> Signed-off-by: Chuyi Zhou <zhouchuyi@bytedance.com>
>> ---
>>   kernel/sched/fair.c | 6 ++++++
>>   1 file changed, 6 insertions(+)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 9057584ec06d..2807f6e72ad1 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -3782,6 +3782,8 @@ static void reweight_eevdf(struct sched_entity 
>> *se, u64 avruntime,
>>       se->deadline = avruntime + vslice;
>>   }
>> +static inline void update_load_avg(struct cfs_rq *cfs_rq, struct 
>> sched_entity *se, int flags);
>> +
>>   static void reweight_entity(struct cfs_rq *cfs_rq, struct 
>> sched_entity *se,
>>                   unsigned long weight)
>>   {
>> @@ -3795,7 +3797,11 @@ static void reweight_entity(struct cfs_rq 
>> *cfs_rq, struct sched_entity *se,
>>           if (!curr)
>>               __dequeue_entity(cfs_rq, se);
>>           update_load_sub(&cfs_rq->load, se->load.weight);
>> +    } else {
>> +        /* Sync with the cfs_rq before removing our load_avg */
>> +        update_load_avg(cfs_rq, se, 0);
> 
> I think it's suboptimal to update_load_avg() here unconditionally.
> 
> Because reweight_entity() has two types of usages:
> 
> 1. group se, which uses reweight_entity() in update_cfs_group(), which
> should already update_load_avg(), so should have no problem.
> 
> 2. task se, which uses reweight_entity() in reweight_task(), which 
> should be fixed for sleep task entity as you described above.
> 
> So IMHO, we should only update_load_avg() or sync_entity_load_avg() in
> reweight_task(), right?
> 

I will send another patch later.

Thanks for the review!