[PATCH] drm/ttm: make sure pool pages are cleared

Christian König ckoenig.leichtzumerken at gmail.com
Thu Feb 11 16:04:44 UTC 2021



Am 11.02.21 um 16:58 schrieb Daniel Vetter:
> On Wed, Feb 10, 2021 at 09:23:52PM +0100, Christian König wrote:
>>
>> Am 10.02.21 um 19:15 schrieb Daniel Vetter:
>>> On Wed, Feb 10, 2021 at 5:05 PM Christian König
>>> <ckoenig.leichtzumerken at gmail.com> wrote:
>>>> The old implementation wasn't consistend on this.
>>>>
>>>> But it looks like we depend on this so better bring it back.
>>>>
>>>> Signed-off-by: Christian König <christian.koenig at amd.com>
>>>> Reported-and-tested-by: Mike Galbraith <efault at gmx.de>
>>>> Fixes: d099fc8f540a ("drm/ttm: new TT backend allocation pool v3")
>>> Well I think in general there's a bit an issue in ttm with not
>>> clearing stuff everywhere. So definitely in favour of clearing stuff.
>>> Looking at the code this only fixes the clearing, at alloc time we're
>>> still very optional with clearing. I think we should just set
>>> __GFP_ZERO unconditionally there too.
>> No, the alloc handling is actually correct.
>>
>> We are clearing only when we allocate pages for userspace. Not for the
>> kernel nor for eviction when pages are overwritten anyway.
>>
>> The key point is that the old page pool was ignoring the flag for this in
>> some code paths and I wasn't sure if that's still necessary or not.
>>
>> Turned out it was necessary after all.
> Somehow my git grep went wrong and I didn't find the users. You're right,
> and I learned a few things more :-)
>
> I'm kinda wondering, should we perhaps move the clearing to the use side,
> and then only do when required? Might allow us to save it quite a few
> times when we're thrashing around buffers in/out of vram, at the cost of
> moving it to the alloc side for other cases.

I was playing with that idea in my mind as well.

The key argument against it is that the pool can optimize by clearing on 
free instead of during allocation.

This way we also implement a bit heuristic in the pool and have a list 
of cleared and not cleared pages.

Regards,
Christian.

>
> Just an idea.
> -Daniel



More information about the dri-devel mailing list