2023-11-16 10:55:07

by Yiwei Lin

[permalink] [raw]
Subject: [PATCH 0/1] sched/fair: Update min_vruntime in more relaxed way

Because we only use min_vruntime as an offset to avoid
overflow, and we will always update_curr() before making
change to csf_rq. Unlike the requirement on CFS, I
think it is possible to just update_min_vruntime() when
update_curr(), which reduces the cost to evaluate the statistic.

It is worth metioning that maybe we can update_min_vruntime() in
a more relaxed form according to this fact, but we may need more
works to optimize in the future.

The following link shows some discussion which relates to this patch:
https://lore.kernel.org/lkml/[email protected]/T/#mc277f045367f4186b34c83e636907ce2f5eb481e

Yiwei Lin (1):
sched/fair: Update min_vruntime in more relaxed way

kernel/sched/fair.c | 20 +-------------------
1 file changed, 1 insertion(+), 19 deletions(-)

--
2.34.1


2023-11-16 10:56:01

by Yiwei Lin

[permalink] [raw]
Subject: [PATCH 1/1] sched/fair: Update min_vruntime in more relaxed way

As EEVDF adopts lag-based solution which is irrespective of
min_vruntime like CFS before, min_vruntime is only used as
an offset to avoid overflow on evaluation of avg_vruntime now.
Rely on the fact we will always update_curr() before change
to cfs_rq, it seems to make sense if we just
update_min_vruntime() with update_curr() to reduce the cost.

Signed-off-by: Yiwei Lin <[email protected]>
---
kernel/sched/fair.c | 20 +-------------------
1 file changed, 1 insertion(+), 19 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 07f555857..5c40adfae 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -3815,17 +3815,8 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
enqueue_load_avg(cfs_rq, se);
if (se->on_rq) {
update_load_add(&cfs_rq->load, se->load.weight);
- if (!curr) {
- /*
- * The entity's vruntime has been adjusted, so let's check
- * whether the rq-wide min_vruntime needs updated too. Since
- * the calculations above require stable min_vruntime rather
- * than up-to-date one, we do the update at the end of the
- * reweight process.
- */
+ if (!curr)
__enqueue_entity(cfs_rq, se);
- update_min_vruntime(cfs_rq);
- }
}
}

@@ -5347,15 +5338,6 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)

update_cfs_group(se);

- /*
- * Now advance min_vruntime if @se was the entity holding it back,
- * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
- * put back on, and if we advance min_vruntime, we'll be placed back
- * further than we started -- ie. we'll be penalized.
- */
- if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
- update_min_vruntime(cfs_rq);
-
if (cfs_rq->nr_running == 0)
update_idle_cfs_rq_clock_pelt(cfs_rq);
}
--
2.34.1

2023-11-16 11:17:17

by Abel Wu

[permalink] [raw]
Subject: Re: [PATCH 1/1] sched/fair: Update min_vruntime in more relaxed way

On 11/16/23 6:54 PM, Yiwei Lin Wrote:
> As EEVDF adopts lag-based solution which is irrespective of
> min_vruntime like CFS before, min_vruntime is only used as
> an offset to avoid overflow on evaluation of avg_vruntime now.
> Rely on the fact we will always update_curr() before change
> to cfs_rq, it seems to make sense if we just
> update_min_vruntime() with update_curr() to reduce the cost.
>
> Signed-off-by: Yiwei Lin <[email protected]>
> ---
> kernel/sched/fair.c | 20 +-------------------
> 1 file changed, 1 insertion(+), 19 deletions(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 07f555857..5c40adfae 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -3815,17 +3815,8 @@ static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
> enqueue_load_avg(cfs_rq, se);
> if (se->on_rq) {
> update_load_add(&cfs_rq->load, se->load.weight);
> - if (!curr) {
> - /*
> - * The entity's vruntime has been adjusted, so let's check
> - * whether the rq-wide min_vruntime needs updated too. Since
> - * the calculations above require stable min_vruntime rather
> - * than up-to-date one, we do the update at the end of the
> - * reweight process.
> - */
> + if (!curr)
> __enqueue_entity(cfs_rq, se);
> - update_min_vruntime(cfs_rq);
> - }
> }
> }
>
> @@ -5347,15 +5338,6 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
>
> update_cfs_group(se);
>
> - /*
> - * Now advance min_vruntime if @se was the entity holding it back,
> - * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case we'll be
> - * put back on, and if we advance min_vruntime, we'll be placed back
> - * further than we started -- ie. we'll be penalized.
> - */
> - if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
> - update_min_vruntime(cfs_rq);
> -
> if (cfs_rq->nr_running == 0)
> update_idle_cfs_rq_clock_pelt(cfs_rq);
> }

For now, core pick of core scheduling relies on min_vruntime to be fresh,
so please just fix commit eab03c23c2a1 to preserve its original behavior
by moving update_min_vruntime() to proper position. And behavior change
can be posted separated.

BTW it seems unnecessary to include a cover-letter for a single patch.

Thanks,
Abel

2023-11-16 11:33:47

by Yiwei Lin

[permalink] [raw]
Subject: Re: [PATCH 1/1] sched/fair: Update min_vruntime in more relaxed way


On 11/16/23 19:16, Abel Wu wrote:
> On 11/16/23 6:54 PM, Yiwei Lin Wrote:
>> As EEVDF adopts lag-based solution which is irrespective of
>> min_vruntime like CFS before, min_vruntime is only used as
>> an offset to avoid overflow on evaluation of avg_vruntime now.
>> Rely on the fact we will always update_curr() before change
>> to cfs_rq, it seems to make sense if we just
>> update_min_vruntime() with update_curr() to reduce the cost.
>>
>> Signed-off-by: Yiwei Lin <[email protected]>
>> ---
>>   kernel/sched/fair.c | 20 +-------------------
>>   1 file changed, 1 insertion(+), 19 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 07f555857..5c40adfae 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -3815,17 +3815,8 @@ static void reweight_entity(struct cfs_rq
>> *cfs_rq, struct sched_entity *se,
>>       enqueue_load_avg(cfs_rq, se);
>>       if (se->on_rq) {
>>           update_load_add(&cfs_rq->load, se->load.weight);
>> -        if (!curr) {
>> -            /*
>> -             * The entity's vruntime has been adjusted, so let's check
>> -             * whether the rq-wide min_vruntime needs updated too.
>> Since
>> -             * the calculations above require stable min_vruntime
>> rather
>> -             * than up-to-date one, we do the update at the end of the
>> -             * reweight process.
>> -             */
>> +        if (!curr)
>>               __enqueue_entity(cfs_rq, se);
>> -            update_min_vruntime(cfs_rq);
>> -        }
>>       }
>>   }
>>   @@ -5347,15 +5338,6 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct
>> sched_entity *se, int flags)
>>         update_cfs_group(se);
>>   -    /*
>> -     * Now advance min_vruntime if @se was the entity holding it back,
>> -     * except when: DEQUEUE_SAVE && !DEQUEUE_MOVE, in this case
>> we'll be
>> -     * put back on, and if we advance min_vruntime, we'll be placed
>> back
>> -     * further than we started -- ie. we'll be penalized.
>> -     */
>> -    if ((flags & (DEQUEUE_SAVE | DEQUEUE_MOVE)) != DEQUEUE_SAVE)
>> -        update_min_vruntime(cfs_rq);
>> -
>>       if (cfs_rq->nr_running == 0)
>>           update_idle_cfs_rq_clock_pelt(cfs_rq);
>>   }
>
> For now, core pick of core scheduling relies on min_vruntime to be fresh,
> so please just fix commit eab03c23c2a1 to preserve its original behavior
> by moving update_min_vruntime() to proper position. And behavior change
> can be posted separated.

Sorry for not noticing the requirement on core scheduling and applying
bad solution. I should take a closer look for the influence when
changing the approach to update_min_vruntime().

I'll send another patch which just move update_min_vruntime() to the
right place later on.

>
> BTW it seems unnecessary to include a cover-letter for a single patch.
>
Got it! Still learning how to work with the kernel mailing list. Thanks
for the kind suggestion!
> Thanks,
>     Abel