Re: [PATCH] sched/fair: do not announce throttled next buddy in dequeue_task_fair

From: bsegall
Date: Tue Jul 12 2016 - 13:25:30 EST

Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx> writes:

> On 11.07.2016 15:12, Xunlei Pang wrote:
>> On 2016/07/11 at 17:54, Wanpeng Li wrote:
>>> Hi Konstantin, Xunlei,
>>> 2016-07-11 16:42 GMT+08:00 Xunlei Pang <xpang@xxxxxxxxxx>:
>>>> On 2016/07/11 at 16:22, Xunlei Pang wrote:
>>>>> On 2016/07/11 at 15:25, Wanpeng Li wrote:
>>>>>> 2016-06-16 20:57 GMT+08:00 Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>:
>>>>>>> Hierarchy could be already throttled at this point. Throttled next
>>>>>>> buddy could trigger null pointer dereference in pick_next_task_fair().
>>>>>> There is cfs_rq->next check in pick_next_entity(), so how can null
>>>>>> pointer dereference happen?
>>>>> I guess it's the following code leading to a NULL se returned:
>>>> s/NULL/empty-entity cfs_rq se/
>>>>> pick_next_entity():
>>>>> if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
>>> ^^^^^^^^^^^^^
>>> I think this will return false.
>> With the wrong throttled_hierarchy(), I think this can happen. But after we have the
>> corrected throttled_hierarchy() patch, I can't see how it is possible.
>> dequeue_task_fair():
>> if (task_sleep && parent_entity(se))
>> set_next_buddy(parent_entity(se));
>> How does dequeue_task_fair() with DEQUEUE_SLEEP set(true task_sleep) happen to a throttled hierarchy?
>> IOW, a task belongs to a throttled hierarchy is running?
>> Maybe Konstantin knows the reason.
> This function (dequeue_task_fair) check throttling but at point it could skip several
> levels and announce as next buddy actually throttled entry.
> Probably this bug hadn't happened but this's really hard to prove that this is impossible.
> ->set_curr_task(), PI-boost or some tricky migration in balancer could break this easily.

sched_setscheduler can call put_prev_task, which then can cause a
throttle outside of __schedule(), then the task blocks normally and
deactivate_task(DEQUEUE_SLEEP) happens and you lose.

The obvious way to avoid these would be to somehow change put_prev so
that it only does throttles in the schedule() path (which is what we
/want/), which would probably involve adding a parameter to put_prev for
just this.