[PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[Chengming Zhou]

We use update_load_avg(cfs_rq, se, 0) in throttle_cfs_rq(), so the
cfs_rq->tg_load_avg_contrib and task_group->load_avg won't be updated
even when the cfs_rq's load_avg has changed.

And we also don't call update_cfs_group(se), so the se->load won't
be updated too.

Change to use update_load_avg(cfs_rq, se, UPDATE_TG) and add
update_cfs_group(se) in throttle_cfs_rq(), like we do in
dequeue_task_fair().

Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
---
 kernel/sched/fair.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index d4bd299d67ab..b37dc1db7be7 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4936,8 +4936,9 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
 		if (!se->on_rq)
 			goto done;
 
-		update_load_avg(qcfs_rq, se, 0);
+		update_load_avg(qcfs_rq, se, UPDATE_TG);
 		se_update_runnable(se);
+		update_cfs_group(se);
 
 		if (cfs_rq_is_idle(group_cfs_rq(se)))
 			idle_task_delta = cfs_rq->h_nr_running;
-- 
2.35.1

Re: [PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[Benjamin Segall]

Chengming Zhou <zhouchengming@bytedance.com> writes:

> We use update_load_avg(cfs_rq, se, 0) in throttle_cfs_rq(), so the
> cfs_rq->tg_load_avg_contrib and task_group->load_avg won't be updated
> even when the cfs_rq's load_avg has changed.
>
> And we also don't call update_cfs_group(se), so the se->load won't
> be updated too.
>
> Change to use update_load_avg(cfs_rq, se, UPDATE_TG) and add
> update_cfs_group(se) in throttle_cfs_rq(), like we do in
> dequeue_task_fair().

Hmm, this does look more correct; Vincent, was having this not do
UPDATE_TG deliberate, or an accident that we all missed when checking?

It looks like the unthrottle_cfs_rq side got UPDATE_TG added later in
the two-loops pass, but not the throttle_cfs_rq side.

Also unthrottle_cfs_rq I'm guessing could still use update_cfs_group(se)


>
> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
> ---
>  kernel/sched/fair.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index d4bd299d67ab..b37dc1db7be7 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -4936,8 +4936,9 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
>  		if (!se->on_rq)
>  			goto done;
>  
> -		update_load_avg(qcfs_rq, se, 0);
> +		update_load_avg(qcfs_rq, se, UPDATE_TG);
>  		se_update_runnable(se);
> +		update_cfs_group(se);
>  
>  		if (cfs_rq_is_idle(group_cfs_rq(se)))
>  			idle_task_delta = cfs_rq->h_nr_running;

Re: [External] Re: [PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[Chengming Zhou]

On 2022/4/14 01:30, Benjamin Segall wrote:
> Chengming Zhou <zhouchengming@bytedance.com> writes:
> 
>> We use update_load_avg(cfs_rq, se, 0) in throttle_cfs_rq(), so the
>> cfs_rq->tg_load_avg_contrib and task_group->load_avg won't be updated
>> even when the cfs_rq's load_avg has changed.
>>
>> And we also don't call update_cfs_group(se), so the se->load won't
>> be updated too.
>>
>> Change to use update_load_avg(cfs_rq, se, UPDATE_TG) and add
>> update_cfs_group(se) in throttle_cfs_rq(), like we do in
>> dequeue_task_fair().
> 
> Hmm, this does look more correct; Vincent, was having this not do
> UPDATE_TG deliberate, or an accident that we all missed when checking?
> 
> It looks like the unthrottle_cfs_rq side got UPDATE_TG added later in
> the two-loops pass, but not the throttle_cfs_rq side.

Yes, UPDATE_TG was added in unthrottle_cfs_rq() in commit 39f23ce07b93
("sched/fair: Fix unthrottle_cfs_rq() for leaf_cfs_rq list").

> 
> Also unthrottle_cfs_rq I'm guessing could still use update_cfs_group(se)

It looks like we should also add update_cfs_group(se) in unthrottle_cfs_rq().

Thanks.

> 
> 
>>
>> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
>> ---
>>  kernel/sched/fair.c | 3 ++-
>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index d4bd299d67ab..b37dc1db7be7 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -4936,8 +4936,9 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
>>  		if (!se->on_rq)
>>  			goto done;
>>  
>> -		update_load_avg(qcfs_rq, se, 0);
>> +		update_load_avg(qcfs_rq, se, UPDATE_TG);
>>  		se_update_runnable(se);
>> +		update_cfs_group(se);
>>  
>>  		if (cfs_rq_is_idle(group_cfs_rq(se)))
>>  			idle_task_delta = cfs_rq->h_nr_running;

Re: [External] Re: [PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[Vincent Guittot]

On Fri, 15 Apr 2022 at 07:42, Chengming Zhou
<zhouchengming@bytedance.com> wrote:
>
> On 2022/4/14 01:30, Benjamin Segall wrote:
> > Chengming Zhou <zhouchengming@bytedance.com> writes:
> >
> >> We use update_load_avg(cfs_rq, se, 0) in throttle_cfs_rq(), so the
> >> cfs_rq->tg_load_avg_contrib and task_group->load_avg won't be updated
> >> even when the cfs_rq's load_avg has changed.
> >>
> >> And we also don't call update_cfs_group(se), so the se->load won't
> >> be updated too.
> >>
> >> Change to use update_load_avg(cfs_rq, se, UPDATE_TG) and add
> >> update_cfs_group(se) in throttle_cfs_rq(), like we do in
> >> dequeue_task_fair().
> >
> > Hmm, this does look more correct; Vincent, was having this not do
> > UPDATE_TG deliberate, or an accident that we all missed when checking?

The cost of UPDATE_TG/update_tg_load_avg() is not free and the parent
cfs->load_avg should not change because of the throttling but only the
cfs->weight so I don't see a real benefit of UPDATE_TG.

Chengming,
have you faced an issue or this change is based on code review ?

> >
> > It looks like the unthrottle_cfs_rq side got UPDATE_TG added later in
> > the two-loops pass, but not the throttle_cfs_rq side.
>
> Yes, UPDATE_TG was added in unthrottle_cfs_rq() in commit 39f23ce07b93
> ("sched/fair: Fix unthrottle_cfs_rq() for leaf_cfs_rq list").
>
> >
> > Also unthrottle_cfs_rq I'm guessing could still use update_cfs_group(se)
>
> It looks like we should also add update_cfs_group(se) in unthrottle_cfs_rq().
>
> Thanks.
>
> >
> >
> >>
> >> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
> >> ---
> >>  kernel/sched/fair.c | 3 ++-
> >>  1 file changed, 2 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >> index d4bd299d67ab..b37dc1db7be7 100644
> >> --- a/kernel/sched/fair.c
> >> +++ b/kernel/sched/fair.c
> >> @@ -4936,8 +4936,9 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
> >>              if (!se->on_rq)
> >>                      goto done;
> >>
> >> -            update_load_avg(qcfs_rq, se, 0);
> >> +            update_load_avg(qcfs_rq, se, UPDATE_TG);
> >>              se_update_runnable(se);
> >> +            update_cfs_group(se);
> >>
> >>              if (cfs_rq_is_idle(group_cfs_rq(se)))
> >>                      idle_task_delta = cfs_rq->h_nr_running;

Re: [External] Re: [PATCH] sched/fair: update tg->load_avg and se->load in throttle_cfs_rq()

[Chengming Zhou]

On 2022/4/15 15:51, Vincent Guittot wrote:
> On Fri, 15 Apr 2022 at 07:42, Chengming Zhou
> <zhouchengming@bytedance.com> wrote:
>>
>> On 2022/4/14 01:30, Benjamin Segall wrote:
>>> Chengming Zhou <zhouchengming@bytedance.com> writes:
>>>
>>>> We use update_load_avg(cfs_rq, se, 0) in throttle_cfs_rq(), so the
>>>> cfs_rq->tg_load_avg_contrib and task_group->load_avg won't be updated
>>>> even when the cfs_rq's load_avg has changed.
>>>>
>>>> And we also don't call update_cfs_group(se), so the se->load won't
>>>> be updated too.
>>>>
>>>> Change to use update_load_avg(cfs_rq, se, UPDATE_TG) and add
>>>> update_cfs_group(se) in throttle_cfs_rq(), like we do in
>>>> dequeue_task_fair().
>>>
>>> Hmm, this does look more correct; Vincent, was having this not do
>>> UPDATE_TG deliberate, or an accident that we all missed when checking?
> 
> The cost of UPDATE_TG/update_tg_load_avg() is not free and the parent
> cfs->load_avg should not change because of the throttling but only the
> cfs->weight so I don't see a real benefit of UPDATE_TG.

Hi Vincent,

If the current task has dequeued before throttle_cfs_rq() when pick_next_task_fair,
the parent cfs_rq will wait to update_tg_load_avg() until the throttle_cfs_rq()
when enqueue_entity(), cause delay update of parent cfs_rq->load_avg and the
load.weight of that group_se, so the fairness of task_groups may be delayed.

update_tg_load_avg() won't touch tg->load_avg if (delta <= cfs_rq->tg_load_avg_contrib / 64).
So the cost may have been avoided if the load_avg is really unchanged ?

> 
> Chengming,
> have you faced an issue or this change is based on code review ?

Yes, this change is based on code review and git log history.

Thanks.

> 
>>>
>>> It looks like the unthrottle_cfs_rq side got UPDATE_TG added later in
>>> the two-loops pass, but not the throttle_cfs_rq side.
>>
>> Yes, UPDATE_TG was added in unthrottle_cfs_rq() in commit 39f23ce07b93
>> ("sched/fair: Fix unthrottle_cfs_rq() for leaf_cfs_rq list").
>>
>>>
>>> Also unthrottle_cfs_rq I'm guessing could still use update_cfs_group(se)
>>
>> It looks like we should also add update_cfs_group(se) in unthrottle_cfs_rq().
>>
>> Thanks.
>>
>>>
>>>
>>>>
>>>> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
>>>> ---
>>>>  kernel/sched/fair.c | 3 ++-
>>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index d4bd299d67ab..b37dc1db7be7 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -4936,8 +4936,9 @@ static bool throttle_cfs_rq(struct cfs_rq *cfs_rq)
>>>>              if (!se->on_rq)
>>>>                      goto done;
>>>>
>>>> -            update_load_avg(qcfs_rq, se, 0);
>>>> +            update_load_avg(qcfs_rq, se, UPDATE_TG);
>>>>              se_update_runnable(se);
>>>> +            update_cfs_group(se);
>>>>
>>>>              if (cfs_rq_is_idle(group_cfs_rq(se)))
>>>>                      idle_task_delta = cfs_rq->h_nr_running;