Re: [PATCH v2] sched/pelt: Keep UTIL_AVG_UNCHANGED flag in sync when calculating last_enqueued_diff

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 07/05/2021 14:35, Vincent Donnefort wrote:
> On Fri, May 07, 2021 at 07:20:31PM +0800, Xuewen Yan wrote:
>> From: Xuewen Yan <xuewen.yan@unisoc.com>
>>
>> Last_enqueued_diff's meaning: "diff = util_est.enqueued(p) - task_util(p)".
>> When calculating last_enqueued_diff, we add UTIL_AVG_UNCHANGED flag, which
>> is the LSB, for task_util, but don't add the flag for util_est.enqueued.
>> However the enqueued's flag had been cleared when the task util changed.
>> As a result, we add +1 to the diff, this is therefore reducing slightly
>> UTIL_EST_MARGIN.
> 
> Unless I miss something it actually depends on the situation, doesn't it?
> 
> if ue.enqueued > task_util(), we remove 1 from the diff => UTIL_EST_MARGIN + 1
> 
> if ue.enqueued < task_util(), we add 1 to the diff => UTIL_EST_MARGIN -1

I get:

ue.enqueued & UTIL_AVG_UNCHANGED = 0

last_enqueued_diff = ue.enqueued_old                    -  ue.enqueued_new

last_enqueued_diff = (ue.enqueued | UTIL_AVG_UNCHANGED) - (task_util(p) | UTIL_AVG_UNCHANGED)

                                   ^^^^^^^^^^^^^^^^^^^^
                                   added by patch


ue.enqueued_old didn't have the flag, otherwise would return earlier

task_util(p) could have the LSB set but we just set it to make sure it's set

So last_enqueued_diff is 1 larger.

But UTIL_EST_MARGIN stays `SCHED_CAPACITY_SCALE / 100` ?

Did I miss something here?

>>
>> Add the flag for util_est.enqueued to have a accurate computation.
>>
>> Fixes: b89997aa88f0b sched/pelt: Fix task util_est update filtering
>>
>> Signed-off-by: Xuewen Yan <xuewen.yan@unisoc.com>
>> ---
>>  kernel/sched/fair.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index e5e457fa9dc8..94d77b4fa601 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -3996,7 +3996,7 @@ static inline void util_est_update(struct cfs_rq *cfs_rq,
>>  	if (ue.enqueued & UTIL_AVG_UNCHANGED)
>>  		return;
>>  
>> -	last_enqueued_diff = ue.enqueued;
>> +	last_enqueued_diff = (ue.enqueued | UTIL_AVG_UNCHANGED);
>>  
>>  	/*
>>  	 * Reset EWMA on utilization increases, the moving average is used only
>> -- 
>> 2.29.0
>>