[PATCH 0/3] drm/amdgpu: Tweaks for high pressure on CPU visible VRAM

[Marek Olšák]

On Thu, May 25, 2017 at 5:31 AM, Michel Dänzer <michel at daenzer.net> wrote:
> On 24/05/17 08:27 PM, Christian König wrote:
>> Am 24.05.2017 um 13:03 schrieb Marek Olšák:
>>>>
>>> I think the final solution (done in fault_reserve_notify) should be:
>>> if (bo->num_cpu_page_faults++ > 20)
>>>     bo->preferred_domain = GTT_WC;
>
> I agree something like that will probably be part of the solution, but I
> doubt it's quite that simple or that it's the only thing that can be
> improved.
>
>
>> I more or less agree on that, but setting preferred_domain permanently
>> to GTT_WC is what worries me a bit.
>>
>> E.g. imagine you alt+tab from a game to your browser and back and the
>> game runs way slower now because BOs are never moved back to VRAM.
>
> Right, permanently moving a BO to GTT might itself cause performance to
> drop down a cliff in some cases. It's possible that this is irrelevant
> compared to excessive buffer migration for CPU access though.
>
>
>> What we need is a global limit of number of bytes transfered per second
>> for swap operations or something like that.
>>
>> Or maybe a timeout which says when a BO was moved (either by swapping it
>> out or by a CPU page fault) only move it back after +n jiffies or
>> something like that.
>
> I also feel like something like this will be more useful than the number
> of CPU page faults per se. But I'm curious what Marek comes up with. :)

I don't have any better idea at the moment. It looks like John Brooks
has already solved this issue based on his IRC comments.

Marek