Re: sched: observed instability under stress in 6.12 and mainline

[Hao Jia <jiahao.kernel@gmail.com>]

On 2025/9/8 09:51, Hao Jia wrote:
> 
> 
> On 2025/9/5 00:33, Chris Friesen wrote:
>> Hi,
>>
>> I'd like to draw the attention of the scheduler maintainers to a 
>> number of kernel bugzilla reports submitted by a colleague a couple of 
>> weeks ago:
>>
>> 6.12.18:
>> https://bugzilla.kernel.org/show_bug.cgi?id=220447
>> https://bugzilla.kernel.org/show_bug.cgi?id=220448
>>
>> v6.16-rt3
>> https://bugzilla.kernel.org/show_bug.cgi?id=220450
>> https://bugzilla.kernel.org/show_bug.cgi?id=220449
>>
>> There seems to be something wrong with either the logic or the 
>> locking. In one case this resulted in a NULL pointer dereference in 
>> pick_next_entity().  In another case it resulted in 
>> BUG_ON(!rq->nr_running) in dequeue_top_rt_rq() and 
>> SCHED_WARN_ON(!se->on_rq) in update_entity_lag().
>>
>> My colleague suggests that the NULL pointer dereference may be due to 
>> pick_eevdf() returning NULL in pick_next_entity().
>>
>> I did some digging and found that 
>> https://gitlab.com/linux-kernel/stable/-/commit/86b37810 would not 
>> have been included in 6.12.18, but the equivalent fix should have been 
>> in the 6.16 load.
>>
>> We haven't yet bottomed out the root cause.
>>
>> Any suggestions or assistance would be appreciated.
>>
>> Thanks,
>> Chris
>>
>>
> 
> Maybe this patch can be useful for your problem.
> https://lore.kernel.org/all/tencent_3177343A3163451463643E434C61911B4208@qq.com/
> 
> If I understand correctly, we may dequeue_entity twice in 
> rt_mutex_setprio()/__sched_setscheduler(). cfs_bandwidth may break the 
> state of p->on_rq and se->on_rq.
> 

Perhaps the "Defer throttle when task exits to user" patch set from the 
sched/core branch can also fix this bug.

https://lore.kernel.org/all/20250829081120.806-4-ziqianlu@bytedance.com/

Thanks,
Hao