Re: [PATCH 1/3] sched/fair: Disable runtime_enabled on dying rq

[Kirill Tkhai <tkhai@yandex.ru>]

On 24.06.2014 01:05, bsegall@google.com wrote:
> Kirill Tkhai <tkhai@yandex.ru> writes:
> 
>> On 23.06.2014 21:29, bsegall@google.com wrote:
>>> Peter Zijlstra <peterz@infradead.org> writes:
>>>
>>>> On Tue, Jun 17, 2014 at 05:24:10PM +0400, Kirill Tkhai wrote:
>>>>> @@ -3790,6 +3803,12 @@ static void __maybe_unused unthrottle_offline_cfs_rqs(struct rq *rq)
>>>>>  		cfs_rq->runtime_remaining = 1;
>>>>>  		if (cfs_rq_throttled(cfs_rq))
>>>>>  			unthrottle_cfs_rq(cfs_rq);
>>>>> +
>>>>> +		/*
>>>>> +		 * Offline rq is schedulable till cpu is completely disabled
>>>>> +		 * in take_cpu_down(), so we prevent new cfs throttling here.
>>>>> +		 */
>>>>> +		cfs_rq->runtime_enabled = 0;
>>>>
>>>> Does it make sense to clear this before calling unthrottle_cfs_rq()?
>>>> Just to make sure they're in the right order..
>>>
>>> I believe that order is irrelevant here - I do not believe that
>>> unthrottle_cfs_rq(a) can cause a throttle_cfs_rq(a). In fact, I don't
>>> see any code that will check it at all from unthrottle, although I might
>>> be missing something. It _can_ cause a throttle_cfs_rq(parent_cfs_rq(a)),
>>> but that should be fine as long as for_each_leaf_cfs_rq is sorted
>>> correctly.
>>
>> I think this is correct. We may change the order just for the hope, that
>> anybody will work on it in some way in the future, and this person could
>> skip this opaque place. Ok, I don't know how is better :)
>>
>>> That said, migrate_tasks drops rq->lock, and I /think/ another cpu could
>>> wake another task onto this cpu, which could then enqueue_throttle its
>>> cfs_rq (which previously had no tasks and thus wasn't on
>>> leaf_cfs_rq_list). You certainly could have tg_set_bandwidth come in and
>>> turn runtime_enabled on.
>>
>> We mask cpu inactive on CPU_DOWN_PREPARE stage (in sched_cpu_inactive).
>> Other cpu is not able to wake a task there after that.
>>
>> rq is masked offline in cpuset_cpu_inactive() (during the same stage).
>> But priority of sched_cpu_inactive() is higher than priority of
>> cpuset_cpu_inactive().
>>
>>         CPU_PRI_SCHED_INACTIVE  = INT_MIN + 1,
>>         CPU_PRI_CPUSET_INACTIVE = INT_MIN,
>>
>> This guarantees that nobody could use dying_cpu even before we start
>> unthrottling. Another cpu will see dying_cpu is inactive.
>>
>> So, it looks like we are free of this problem.
> 
> Ah, ok, I haven't looked that hard at hotplug, and wasn't sure of the
> ordering there. We still have the tg_set_cfs_bandwidth issue, because
> that uses for_each_possible_cpu. However, with the addition of
> rq_online_fair, that can probably be changed to for_each_active_cpu, and
> then I think we would be fine.

Ok, now I see. Thanks, Ben.