Random short freezes due to TTM buffer migrations

[Felix Kuehling]

I think the scatter-gather tables only support system memory. As I
understand it, a buffer in VRAM has be migrated to system memory before
it can be shared with another driver.

I'm more concerned about sharing with the same driver. There is a
special code path for that, where we simply add another reference to the
same BO, instead of looking at a scatter gather table. We use that for
OpenGL-OpenCL interop, and also planning to use it for IPC buffer
sharing in HSA. As long as a split VRAM buffer is still a single
amdgpu_bo, and becomes a single dmabuf when exporting it, I think that
should work.

Regards,
  Felix


On 16-08-17 02:58 AM, Christian König wrote:
>> One question: Will it be possible to share these split BOs as dmabufs?
> In theory yes, in practice I'm not sure.
>
> DMA-bufs are designed around scatter gather tables, those fortunately
> support buffers split over the whole address space.
>
> The problem is the importing device needs to be able to handle that as
> well.
>
> Regards,
> Christian.
>
> Am 16.08.2016 um 20:33 schrieb Felix Kuehling:
>> Very nice. I'm looking forward to this for KFD as well.
>>
>> One question: Will it be possible to share these split BOs as dmabufs?
>>
>> Regards,
>>    Felix
>>
>>
>> On 16-08-16 11:27 AM, Christian König wrote:
>>> Hi Marek,
>>>
>>> I'm already working on this.
>>>
>>> My current approach is to use a custom BO manager for VRAM with TTM
>>> and so split allocations into chunks of 4MB.
>>>
>>> Large BOs are still swapped out as one, but it makes it much more
>>> likely to that you can allocate 1/2 of VRAM as one buffer.
>>>
>>> Give me till the end of the week to finish this and then we can test
>>> if that's sufficient or if we need to do more.
>>>
>>> Regards,
>>> Christian.
>>>
>>> Am 16.08.2016 um 16:33 schrieb Marek Olšák:
>>>> Hi,
>>>>
>>>> I'm seeing random temporary freezes (up to 2 seconds) under memory
>>>> pressure. Before I describe the exact circumstances, I'd like to say
>>>> that this is a serious issue affecting playability of certain AAA
>>>> Linux games.
>>>>
>>>> In order to reproduce this, an application should:
>>>> - allocate a few very large buffers (256-512 MB per buffer)
>>>> - allocate more memory than there is available VRAM. The issue also
>>>> occurs (but at a lower frequency) if the app needs only 80% of VRAM.
>>>>
>>>> Example: ttm_bo_validate needs to migrate a 512 MB buffer. The total
>>>> size of moved memory for that call can be as high as 1.5 GB. This is
>>>> always followed by a big temporary drop in VRAM usage.
>>>>
>>>> The game I'm testing needs 3.4 GB of VRAM.
>>>>
>>>> Setups:
>>>> Tonga - 2 GB: It's nearly unplayable, because freezes occur too often.
>>>> Fiji - 4 GB: There is one freeze at the beginning (which is annoying
>>>> too), after that it's smooth.
>>>>
>>>> So even 4 GB is not enough.
>>>>
>>>> Workarounds:
>>>> - Split buffers into smaller pieces in the kernel. It's not necessary
>>>> to manage memory at page granularity (64KB). Splitting buffers into
>>>> 16MB-large pieces might not be optimal but it would be a significant
>>>> improvement.
>>>> - Or do the same in Mesa. This would prevent inter-process and
>>>> inter-API buffer sharing for split buffers (DRI, OpenCL), but we would
>>>> at least verify how much the situation improves.
>>>>
>>>> Other issues sharing the same cause:
>>>> - Allocations requesting 1/3 or more VRAM have a high chance of
>>>> failing. It's generally not possible to allocate 1/2 or more VRAM as
>>>> one buffer.
>>>>
>>>> Comments welcome,
>>>>
>>>> Marek
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>>
>>> _______________________________________________
>>> amd-gfx mailing list
>>> amd-gfx at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>