Bad rss-counter state from drm/ttm, drm/vmwgfx: Support huge TTM pagefaults

[Thomas Hellström (VMware)]

On 4/7/20 5:36 PM, Alex Xu (Hello71) wrote:
> Excerpts from Thomas Hellström (VMware)'s message of April 7, 2020 7:26 am:
>> On 4/7/20 2:38 AM, Alex Xu (Hello71) wrote:
>>> Excerpts from Thomas Hellström (VMware)'s message of April 6, 2020 5:04 pm:
>>>> Hi,
>>>>
>>>> On 4/6/20 9:51 PM, Alex Xu (Hello71) wrote:
>>>>> Using 314b658 with amdgpu, starting sway and firefox causes "BUG: Bad
>>>>> rss-counter state" and "BUG: non-zero pgtables_bytes on freeing mm" to
>>>>> start filling dmesg, and then closing programs causes more BUGs and
>>>>> hangs, and then everything grinds to a halt (can't start more programs,
>>>>> can't even reboot through systemd).
>>>>>
>>>>> Using master and reverting that branch up to that point fixes the
>>>>> problem.
>>>>>
>>>>> I'm using a Ryzen 1600 and AMD Radeon RX 480 on an ASRock B450 Pro4
>>>>> board with IOMMU enabled.
>>>> If you could try the attached patch, that'd be great!
>>>>
>>>> Thanks,
>>>>
>>>> Thomas
>>>>
>>> Yeah, that works too. Kernel config sent off-list.
>>>
>>> Regards,
>>> Alex.
>> Thanks. Do you want me to add your
>>
>> Reported-by: and Tested-by: To this patch?
>>
>> /Thomas
>>
>>
> Sure. Shouldn't we fix it properly though?

It's still enabled for vmwgfx for which it is reasonably well tested and 
where I can't see any such errors.

The code we remove with this patch enables huge page-table entries in 
some circumstances for other drivers, but given the problems you're 
seeing for amdgpu, it's better to enable this on a per-driver basis 
after thorough testing. Since I don't have amdgpu hardware I'm not sure 
what it's doing differently, and can't debug the issue properly.

/Thomas