Re: Crash in fair scheduler

[Dietmar Eggemann]

On 03/12/2019 12:09, Valentin Schneider wrote:
> On 03/12/2019 10:40, Dietmar Eggemann wrote:
>> On 03/12/2019 11:30, Valentin Schneider wrote:
>>> On 03/12/2019 09:11, Schmid, Carsten wrote:
>>
>> [...]
>>
>>> That looks a lot like a recent issue we've had, see
>>>
>>>   https://lore.kernel.org/lkml/20191108131909.428842459@xxxxxxxxxxxxx/
>>>
>>> The issue is caused by
>>>   
>>>   67692435c411 ("sched: Rework pick_next_task() slow-path")
>>>
>>> which 5.4-rc2 has (without the fix which landed in -rc7) but 4.14 really
>>> shouldn't, unless the kernel you're using has had core scheduling somehow
>>> backported to it?
>>>
>>> I've only scraped the surface but I'd like to first ask: can you reproduce
>>> the issue on v5.4 final ?
>>
>> Can't be. 4.14.86 does not have ("sched: Rework pick_next_task()
>> slow-path").
>>
> 
> Right, which is why I wondered if the kernel under test had had that 
> backported to it for some reason (e.g. core scheduling). Peter pointed out
> that this is a slightly different issue (nr_running matches the rbt), so
> this is probably unrelated.

I can't reproduce it on Arm64 Juno running 4.14.86. I suppose that there
is no extra reproducer testcase since the issue happened with
prev->sched_class eq. &idle_sched_class [prev eq. swapper/X 0] in the
simple path of pick_next_task_fair().

I'm running with CONFIG_SCHED_AUTOGROUP=y and CONFIG_FAIR_GROUP_SCHED=y
some taskgroup related tests for hours now. So the sched_entity (se) can
be a task, an autogroup or a taskgroup in the simple path. pref is
either swapper/X or migration/X.