Lockdep spalt on killing a processes

[Christian König]

Am 28.10.21 um 19:26 schrieb Andrey Grodzovsky:
>
> On 2021-10-27 3:58 p.m., Andrey Grodzovsky wrote:
>>
>> On 2021-10-27 10:50 a.m., Christian König wrote:
>>> Am 27.10.21 um 16:47 schrieb Andrey Grodzovsky:
>>>>
>>>> On 2021-10-27 10:34 a.m., Christian König wrote:
>>>>> Am 27.10.21 um 16:27 schrieb Andrey Grodzovsky:
>>>>>> [SNIP]
>>>>>>>
>>>>>>>> Let me please know if I am still missing some point of yours.
>>>>>>>
>>>>>>> Well, I mean we need to be able to handle this for all drivers.
>>>>>>
>>>>>>
>>>>>> For sure, but as i said above in my opinion we need to change 
>>>>>> only for those drivers that don't use the _locked version.
>>>>>
>>>>> And that absolutely won't work.
>>>>>
>>>>> See the dma_fence is a contract between drivers, so you need the 
>>>>> same calling convention between all drivers.
>>>>>
>>>>> Either we always call the callback with the lock held or we always 
>>>>> call it without the lock, but sometimes like that and sometimes 
>>>>> otherwise won't work.
>>>>>
>>>>> Christian.
>>>>
>>>>
>>>> I am not sure I fully understand what problems this will cause but 
>>>> anyway, then we are back to irq_work. We cannot embed irq_work as 
>>>> union within dma_fenc's cb_list
>>>> because it's already reused as timestamp and as rcu head after the 
>>>> fence is signaled. So I will do it within drm_scheduler with single 
>>>> irq_work per drm_sched_entity
>>>> as we discussed before.
>>>
>>> That won't work either. We free up the entity after the cleanup 
>>> function. That's the reason we use the callback on the job in the 
>>> first place.
>>
>>
>> Yep, missed it.
>>
>>
>>>
>>> We could overlead the cb structure in the job though.
>>
>>
>> I guess, since no one else is using this member it after the cb 
>> executed.
>>
>> Andrey
>
>
> Attached a patch. Give it a try please, I tested it on my side and 
> tried to generate the right conditions to trigger this code path by 
> repeatedly submitting commands while issuing GPU reset to stop the 
> scheduler and then killing command submissions process in the middle. 
> But for some reason looks like the job_queue was always empty already 
> at the time of entity kill.

It was trivial to trigger with the stress utility I've hacked together:

amdgpu_stress -b v 1g -b g 1g -c 1 2 1g 1k

Then while it is copying just cntrl+c to kill it.

The patch itself is:

Tested-by: Christian König <christian.koenig at amd.com>
Reviewed-by: Christian König <christian.koenig at amd.com>

Thanks,
Christian.

>
> Andrey
>
>
>>
>>
>>>
>>> Christian.
>>>
>>>>
>>>> Andrey
>>>>
>>>>
>>>>>
>>>>>>
>>>>>> Andrey
>>>>>
>>>