Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <fe40bae5-7ec4-d5d2-ffb8-1387300dbbaf@bytedance.com>
Date:   Thu, 14 Jul 2022 21:43:13 +0800
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.15; rv:102.0)
 Gecko/20100101 Thunderbird/102.0.1
Subject: Re: [External] Re: [PATCH v2 09/10] sched/fair: stop load tracking
 when task switched_from_fair()
Content-Language: en-US
To:     Dietmar Eggemann <dietmar.eggemann@arm.com>, mingo@redhat.com,
        peterz@infradead.org, vincent.guittot@linaro.org,
        rostedt@goodmis.org, bsegall@google.com, vschneid@redhat.com
Cc:     linux-kernel@vger.kernel.org
References: <20220713040430.25778-1-zhouchengming@bytedance.com>
 <20220713040430.25778-10-zhouchengming@bytedance.com>
 <27f0675e-8ac3-2200-749f-7290a256e3d9@arm.com>
From:   Chengming Zhou <zhouchengming@bytedance.com>
In-Reply-To: <27f0675e-8ac3-2200-749f-7290a256e3d9@arm.com>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 2022/7/14 20:33, Dietmar Eggemann wrote:
> On 13/07/2022 06:04, Chengming Zhou wrote:
>> The same reason as the previous commit, if we don't reset the
>> sched_avg last_update_time to 0, after a while in switched_to_fair():
>>
>> switched_to_fair
>>   attach_task_cfs_rq
>>     attach_entity_cfs_rq
>>       update_load_avg
>>         __update_load_avg_se(now, cfs_rq, se)
>>
>> The delta (now - sa->last_update_time) will wrongly contribute/decay
>> sched_avg depends on the task running/runnable status at that time.
> 
> IMHO, a queued !fair task when switching back to fair will already be
> enqueued (attached) as a fair task in __sched_setscheduler() prior to
> the check_class_changed() call.

Right, this is true for a queued !fair task, it will enqueued (attached) before
check_class_changed().

enqueue_task_fair()
  enqueue_entity()
    update_load_avg(cfs_rq, se, UPDATE_TG | DO_ATTACH)
      if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))  --> true
        __update_load_avg_se(now, cfs_rq, se)  --> (1)
check_class_changed()
  switched_to_fair()
    attach_task_cfs_rq()
      attach_entity_cfs_rq()
        update_load_avg(cfs_rq, se, sched_feat(ATTACH_AGE_LOAD) ? 0 : SKIP_AGE_LOAD)
          if (se->avg.last_update_time && !(flags & SKIP_AGE_LOAD))  --> true
            __update_load_avg_se(now, cfs_rq, se)  --> (2)


1. for queued !fair: (1) delta = (now - last_update_time), last_update_time is the time
   when switched_from_fair().

2. for !queued !fair: (2) delta = (now - last_update_time), last_update_time is the time
   when switched_from_fair().

The scenario in the commit message only cover !queued !fair case, I forget the queued !fair
case, their problem is the same.


> 
> I can't see how this will work with your proposed change in using
> last_update_time=0 for fair->!fair->fair class changes?
If we reset last_update_time=0 for !fair task, then:

1. for queued !fair: will not do (1) since the if condition is false.

2. for !queued !fair: will not do (2) since the if condition is false.

Thanks.

> 
>> This patch reset it's sched_avg last_update_time to 0, stop load
>> tracking for !fair task, later in switched_to_fair() ->
>> update_load_avg(), we can use its saved sched_avg.
>>
>> Signed-off-by: Chengming Zhou <zhouchengming@bytedance.com>
>> ---
>>  kernel/sched/fair.c | 5 +++++
>>  1 file changed, 5 insertions(+)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 50f65a2ede32..576028f5a09e 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -11552,6 +11552,11 @@ static void attach_task_cfs_rq(struct task_struct *p)
>>  static void switched_from_fair(struct rq *rq, struct task_struct *p)
>>  {
>>  	detach_task_cfs_rq(p);
>> +
>> +#ifdef CONFIG_SMP
>> +	/* Stop load tracking for !fair task */
>> +	p->se.avg.last_update_time = 0;
>> +#endif
>>  }
>>  
>>  static void switched_to_fair(struct rq *rq, struct task_struct *p)
>