[PATCH] drm/amdgpu: fix ib test hang with gfxoff enabled

[Christian König]

Am 04.06.2018 um 09:52 schrieb Huang Rui:
> On Sat, Jun 02, 2018 at 03:01:57AM +0800, Kuehling, Felix wrote:
>> On 2018-06-01 06:09 AM, Christian König wrote:
>>> Am 01.06.2018 um 11:29 schrieb Huang Rui:
>>>> On Fri, Jun 01, 2018 at 05:13:49PM +0800, Christian König wrote:
>>>>> Am 01.06.2018 um 08:41 schrieb Huang Rui:
>>>>>> After defer the execution of gfx/compute ib tests. However, at that
>>>>>> time, the
>>>>>> gfx already go into "mid state" of gfxoff.
>>>>>>
>>>>>> PWR_MISC_CNTL_STATUS: PWR_GFXOFF_STATUS field (2:1 bits)
>>>>>> 0 = GFXOFF.
>>>>>> 1 = Transition out of GFXOFF state.
>>>>>> 2 = Not in GFXOFF.
>>>>>> 3 = Transition into GFXOFF.
>>>>>>
>>>>>> If hit the mid state (1 or 3), the doorbell writing interrupt
>>>>>> cannot wake up the
>>>>>> gfx back successfully. And the field value is 1 when we issue the
>>>>>> ib test at
>>>>>> that, so we got the hang. This is the root cause that we
>>>>>> encountered the issue.
>>>>>>
>>>>>> Meanwhile, we cannot set clockgating of GFX after gfx is already in
>>>>>> "off" state.
>>>>>> So here we should move the gfx powergating and gfxoff enabling
>>>>>> behavior at the
>>>>>> end of initialization behind ib test and clockgating.
>>>>> Mhm, that still looks like a only halve backed solution:
>>>>>
>>>>> 1. What prevents this bug from happening during "normal" IB submission
>>>>> from userspace?
>>>>>
>>>>> 2. Shouldn't we poll the PWR_MISC_CNTL_STATUS register to make sure we
>>>>> are not in any transition phase instead?
>>>>>
>>>> Yes, right. How about also add polling of PWR_MISC_CNTL_STATUS in
>>>> amdgpu_ring_commit() behind set_wptr that confirm the status as "0"
>>>> or "2"?
>>> You could add an end_use() callback for that, but I think we rather
>>> need to do this in gfx_v9_0_ring_set_wptr_gfx() before we write the
>>> doorbell.
>> Isn't testing the status like this is a potential race condition.

Well it could when we use both GFX and compute at the same time.

>>
>> Having to do this at all is contrary to the documentation that I've
>> read. Writing a doorbell should wake up the GFX engine. Are we sure that
>> we understand the cause of the problem correctly? Does the IB test use
>> any MMIO? Maybe it's doing an HDP flush using MMIO for a ring that
>> doesn't support HDP flushing.
>>
> Felix, thanks to reminder. I supposed you mentioned MMIO using is to avoid
> runtime gfx register access, right? Our IB test uses WRITE_DATA packet to
> write specific pattern value into the gart memory. I don't use any gfx
> registers. And gfxoff is only supported on raven, we don't emit hdp flush
> on apu. Actually, I also doubted whether it is caused race condition. But
> the hang happens when only modprobe amdgpu module, and not startx at that
> time. It won't have any other commands from user space.

Felix is perfectly right that this doesn't sounds like a complete 
solution to the problem.

The IB test only brings the issue to the surface, working around it by 
delaying enabling gfxoff would only hide the real problem.

I'm pretty sure that the exact same race can happen with startx or other 
command submissions as well.

> + Morris, who works for raven SMC firmware.
> After discuessed with him, he suggested that we would better to confirm the
> GFXOFF_STATUS as 0 or 2, then write the doorbell. Because if GFXOFF_STATUS
> is 1 or 3 in mid state (in-progress of translation), SMC will drop the
> doorbell interrupt. When GFXOFF status is 0 or 2, already in the target
> state, SMC can repond interrupt at once. Morris, please correct me if I was
> wrong.

That will be rather hard to guarantee and would completely circumvent 
the idea behind gfxoff, e.g. the driver would need to assist the 
firmware again turning things on/off.

Regards,
Christian.

>
> Thanks,
> Ray