Re: [PATCH v2] sched/fair: sanitize vruntime of entity being migrated

[Dietmar Eggemann]

On 21/03/2023 11:49, Peter Zijlstra wrote:
> On Tue, Mar 21, 2023 at 11:29:13AM +0100, Dietmar Eggemann wrote:
>> On 21/03/2023 11:02, Peter Zijlstra wrote:
>>> On Fri, Mar 17, 2023 at 05:08:10PM +0100, Vincent Guittot wrote:
>>>> Commit 829c1651e9c4 ("sched/fair: sanitize vruntime of entity being placed")
>>>> fixes an overflowing bug, but ignore a case that se->exec_start is reset
>>>> after a migration.
>>>>
>>>> For fixing this case, we delay the reset of se->exec_start after
>>>> placing the entity which se->exec_start to detect long sleeping task.
>>>>
>>>> In order to take into account a possible divergence between the clock_task
>>>> of 2 rqs, we increase the threshold to around 104 days.
>>>>
>>>>
>>>> Fixes: 829c1651e9c4 ("sched/fair: sanitize vruntime of entity being placed")
>>>> Signed-off-by: Zhang Qiao <zhangqiao22@xxxxxxxxxx>
>>>> Signed-off-by: Vincent Guittot <vincent.guittot@xxxxxxxxxx>
>>>> ---
>>>
>>> Blergh, this just isn't going to be nice. I'll go queue this for
>>> sched/urgent and then we can forget about this for a little while.
>>>
>>> Thanks!
>>
>> Don't we miss setting `se->exec_start = 0` for fair task in
>> move_queued_task()? ( ... and __migrate_swap_task())
>>
>> https://lkml.kernel.org/r/df2cccda-1550-b06b-aa74-e0f054e9fb9d@xxxxxxx
> 
> Ah, I see what you mean now... When I read your and Vincent's replies
> earlier today I though you mean to avoid the extra ENQUEUE_MIGRATED use,
> but your actual goal was to capure more sites.
> 
> Hmm, we could of course go add more ENQUEUE_MIGRATED, but you're right
> in that TASK_ON_RQ_MIGRATING already captures that.

And in case of move_queued_task() this would have to be conditioned on
SCHED_NORMAL.

> An alternative is something like the below, that matches
> deactivate_task(), but still uses ENQUEUE_MIGRATED to pass it down into
> the class methods.
> 
> Hmm?
> 
> 
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -2084,6 +2084,9 @@ static inline void dequeue_task(struct r
>  
>  void activate_task(struct rq *rq, struct task_struct *p, int flags)
>  {
> +	if (task_on_rq_migrating(p))
> +		flags |= ENQUEUE_MIGRATED;
> +
>  	enqueue_task(rq, p, flags);
>  
>  	p->on_rq = TASK_ON_RQ_QUEUED;
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -8726,7 +8726,7 @@ static void attach_task(struct rq *rq, s
>  	lockdep_assert_rq_held(rq);
>  
>  	WARN_ON_ONCE(task_rq(p) != rq);
> -	activate_task(rq, p, ENQUEUE_NOCLOCK | ENQUEUE_MIGRATED);
> +	activate_task(rq, p, ENQUEUE_NOCLOCK);
>  	check_preempt_curr(rq, p, 0);
>  }

Would work too.

IMHO, setting `se->exec_start = 0` for task_on_rq_migrating(p) already
in migrate_task_rq_fair() would have the charm that
entity_is_long_sleeper() would bail out early for these tasks.