Re: [Intel-gfx] [PATCH v9 2/8] drm/i915/ttm: add tt shmem backend

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On 12/8/21 09:30, Tvrtko Ursulin wrote:


...


>>> Apart from the code organisation questions, on the practical level - 
>>> do you need writeback from the TTM backend or while I am proposing 
>>> to remove it from the "legacy" paths, I can propose removing it from 
>>> the TTM flow as well?
>>
>> Yeah, if that is somehow busted then we should remove from TTM 
>> backend also.
>
> Okay thanks, I wanted to check in case there was an extra need in TTM. 
> I will float a patch soon hopefully but testing will be a problem 
> since it seems very hard to repro at the moment.

Do we have some information about what's causing the deadlock or a 
signature? I'm asking because if some sort of shrinker was added to TTM 
itself, for the TTM page vectors, it would need to allocate shmem pages 
at shrink time rather than to unpin them at shrink time as we do here. 
And for that to have any chance of working sort of reliably, I think 
writeback is needed.

But I agree for this implementation, the need for writeback isn't 
different than for the non-TTM shmem objects

Thanks,

Thomas


>
> Regards,
>
> Tvrtko