[Mesa-dev] [PATCH 09/10] st/vdpau: implement the new DMA-buf based interop

[Ilia Mirkin]

On Thu, Sep 15, 2016 at 4:54 PM, Marek Olšák <maraeo at gmail.com> wrote:
> On Thu, Sep 15, 2016 at 5:14 AM, Dave Airlie <airlied at gmail.com> wrote:
>> On 15 September 2016 at 13:03, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>>> On Wed, Sep 14, 2016 at 10:15 PM, Michel Dänzer <michel at daenzer.net> wrote:
>>>>> No, the current impl is pretty radeon-specific (note - it doesn't work
>>>>> on nouveau, and no other drivers support the interfaces, so ... it's
>>>>> radeon-specific).
>>>>
>>>> We're getting into semantics here, but since the reason it doesn't work
>>>> well with nouveau is a fundamental issue in nouveau (which should also
>>>> affect at least DRI3 in general), while you may call it "de facto radeon
>>>> specific" if you're so inclined, that doesn't make the implementation
>>>> actually radeon specific.
>>>
>>> No one's reported any issues with DRI3, I use it on my home desktop
>>> every day. And VDPAU used to work great until these changes to
>>> st/vdpau went in. Prior to those changes in st/vdpau, saying that
>>> "shared == gart" was a perfectly reasonable thing to say, since no one
>>> tried blending/readback on those surfaces (or at least not enough for
>>> it to matter). But now ... poof ... it doesn't work [actually, worse -
>>> it works - but can't come close to keeping up with 24fps video].
>>>
>>> Anyways, I realize this is a losing argument. Interfaces and usages
>>> move forward and change over time. This happens to be a change that
>>> leaves nouveau behind. As a spare-time contributor, I can't keep up
>>> with multiple full timers. I had hoped that there'd be some way to
>>> make it all still work, but that doesn't seem to be the case.
>>> Unfortunately end users are going to lose out on functionality as a
>>> result.
>>
>> So (a) this is a regression, regressions aren't allowed, so it would
>> be good to back the change out until it can be fixed.
>>
>> The problem is the combo of LINEAR and SHARED means that
>> GART placement is most likely, radeon should be doing the same
>> in most circumstances.
>>
>> We should possible introduced SHARED_OTHER_GPU maybe
>> and use that throughout the stack where it matters.
>
> The main problem is that nouveau lacks a proper memory management and
> buffers are pinned after allocation forever.
>
> The workaround is to add PIPE_BIND_something, which would do what you
> need it to do, and use it where you need to use it. I don't care about
> the name as long as it works for nouveau. Does that sound reasonable?

I don't really care how it's resolved, but the current situation is
unacceptable.

 - Reverting definitely works.
 - Dropping libvdpau_nouveau definitely works.
 - Adding a new flag may work, depends on details, would require a
bunch of testing.

Dave sent a patch for the first, I've sent a patch for the second. I
don't think a patch for the third has materialized.

  -ilia