Re: NULL pointer dereference in pick_next_task_fair

[Kirill Tkhai <ktkhai@virtuozzo.com>]

On 07.11.2019 18:42, Peter Zijlstra wrote:
> On Thu, Nov 07, 2019 at 06:12:07PM +0300, Kirill Tkhai wrote:
>> On 07.11.2019 16:26, Peter Zijlstra wrote:
> 
>>> Urgh... throttling.
> 
>> One more thing about current code in git. After rq->lock became able to
>> be unlocked after put_prev_task() is commited, we got a new corner case.
>> We usually had the same order for running task:
>>
>>   dequeue_task()
>>   put_prev_task()
>>
>> Now the order may be reversed (this is also in case of throttling):
>>
>>   put_prev_task() (called from pick_next_task())
>>   dequeue_task()  (called from another cpu)
>>
>> This is more theoretically, since I don't see a problem here. But there are
>> too many statistics and counters in sched_class methods, that it is impossible
>> to be sure all of them work as expected.
> 
> Hmm,.. where does throttling happen on a remote CPU? I through both
> cfs-bandwidth and dl throttle locally.
> 
> Or are you talking about NO_HZ_FULL's sched_remote_tick() ?

I mean ordinary path: local throttling -> resched_curr -> schedule().
Then rq->nr_running == 0, but task is on rq. We call put_prev_task()
and newidle_balance().

On another cpu someone calls set_user_nice() and it makes dequeue_task()
in the middle of local cpu's newidle_balance().

Thus, we first made put_prev_task() and second dequeue_task().