[Intel-gfx] ✗ Fi.CI.BAT: failure for series starting with [01/25] dma-buf: add dma_resv_for_each_fence_unlocked v5

[Tvrtko Ursulin]

On 24/09/2021 10:13, Christian König wrote:
> 
> 
> Am 24.09.21 um 11:11 schrieb Tvrtko Ursulin:
>>
>> On 23/09/2021 14:05, Patchwork wrote:
>>
>> [snip]
>>
>>>   *
>>>
>>>     igt at gem_busy@busy at all:
>>>
>>>       o fi-apl-guc: PASS
>>> <https://intel-gfx-ci.01.org/tree/drm-tip/CI_DRM_10630/fi-apl-guc/igt@gem_busy@busy@all.html> 
>>>
>>>         -> FAIL
>>> <https://intel-gfx-ci.01.org/tree/drm-tip/Patchwork_21144/fi-apl-guc/igt@gem_busy@busy@all.html> 
>>>
>>>         +2 similar issues
>>
>> All seem to be the same failure:
>>
>> (gem_busy:874) igt_dummyload-CRITICAL: Test assertion failure function 
>> igt_spin_factory, file ../lib/igt_dummyload.c:490:
>> (gem_busy:874) igt_dummyload-CRITICAL: Failed assertion: 
>> gem_bo_busy(fd, spin->handle)
>>
>> Which is saying spinner which was just submitted is not immediately 
>> reported as busy. And that sounds impossible. Must be a pretty basic 
>> bug somewhere which I don't immediately see. Like unlocked iterator 
>> failing to walk the fences or something.
> 
> I was just to write a mail to you about this since I'm currently 
> scratching my head what exactly goes wrong here.
> 
> Is there an igt test which uses only vgem and currently fails which I 
> could also run on AMD hardware?

Not that I am aware of but you could add debug and use trybot?

However, is the bug perhaps in "all fences" mode is there is no 
exclusive fence? First iteration returns NULL and for loop aborts before 
looking at the shared fences.

Regards,

Tvrtko