[2/2] drm/ttm: make TT creation purely optional v3

[Christian König]

Am 06.08.20 um 11:05 schrieb Michel Dänzer:
> On 2020-08-06 10:43 a.m., Christian König wrote:
>> Am 06.08.20 um 10:01 schrieb Michel Dänzer:
>>> On 2020-08-06 9:13 a.m., Christian König wrote:
>>>> Am 05.08.20 um 18:16 schrieb Michel Dänzer:
>>>>> On 2020-06-29 5:19 p.m., Christian König wrote:
>>>>>> We only need the page array when the BO is about to be accessed.
>>>>>>
>>>>>> So not only populate, but also create it on demand.
>>>>>>
>>>>>> v2: move NULL check into ttm_tt_create()
>>>>>> v3: fix the occurrence in ttm_bo_kmap_ttm as well
>>>>> This broke amdgpu userptr functionality for me, in particular all
>>>>> GL_AMD_pinned_memory piglit tests, e.g.
>>>>>
>>>>> .../piglit/bin/amd_pinned_memory decrement-offset -auto -fbo
>>>>> Offset is decremented, no wait.
>>>>> Mesa: User error: GL_INVALID_OPERATION in glBufferData
>>>>> [...]
>>>>> Unexpected GL error: GL_INVALID_OPERATION 0x502
>>>>> (Error at tests/spec/amd_pinned_memory/test.c:214)
>>>>> PIGLIT: {"result": "fail" }
>>>> But no crashes? I would have expected a NULL pointer deref if we have
>>>> missed call path.
>>> No crashes. Since a GL error is raised, my guess would be that an ioctl
>>> which was previously succeeding is now returning an error.
>> Mhm, I can reproduce the problem that userptrs doesn't work any more.
>> But even with the mentioned patch reverted that still doesn't work on
>> the tip of drm-misc-next.
> Maybe there are other changes affecting it as well, e.g.
>
> e04be2310b5e "drm/ttm: further cleanup ttm_mem_reg handling"
> 1e691e244487 "drm/amdgpu: stop allocating dummy GTT nodes"
>
> come to mind.
>
>
>> Are you sure your bisecting is right?
> Fairly sure. If that commit works or its parent doesn't work for you, I
> can retry.

I've figured out what's wrong here, and yes that needs to extra handling.

Going to have a patch in a minute.

Thanks for the notice,
Christian.