Issues adding driver support for present extension

[Frank Binns]

On 15/04/14 19:55, Keith Packard wrote:
> Frank Binns <frank.binns at imgtec.com> writes:
>
>> Hi,
>> I've been working on adding support for the present extension in the PVR
>> video driver and I've hit a couple of issues related to flipping:
> Awesome. As you may know, this has only been implemented in the Intel
> driver and hasn't been accepted upstream yet. The server API closely
> matches the internals of that driver, and while I spent a week trying to
> rewrite that this spring, I didn't succeed. I'd love to clean it up and
> make it actually usable.
>
>> 1) When the display driver supports async flipping this mode of
>> operation is always used. This means that, when the swap interval is
>> none 0, the present extension relies on vblank events to get
>> synchronised flips. However, by the time we get the event, wait for any
>> outstanding rendering and do the flip we have probably missed the vblank
>> period meaning we get screen artefacts.
> Right, this is fast enough on my laptop GPU to work. I wonder why you're
> waiting for the GPU to flush all of the rendering though? You should
> only need to flush rendering targeting the new back buffer and then flip
> the scanout engine to that.
The driver is only waiting for any rendering to the back buffer before 
doing the flip.
>> Is there a good reason (other
>> than simplifying the code) why present isn't relying on the default
>> behaviour of drmModePageFlip() to get synchronised flips?
> Yes, there's a really good reason -- the drm interface doesn't provide
> for any way to replace the pending buffer. Once you've queued it, you're
> stuck until that buffer becomes active, tying down *two* buffers in the
> kernel. And, you don't get the benefit of providing a new buffer after
> the swap is first queued.
This happens in the non-async case in that once the the flip is 
initiated, i.e. we now have a pending flip, if another swap comes in for 
the same MSC then we can't abort the pending flip.
>
>> 2) When the display driver doesn't support async flipping and the swap
>> interval is 0 we see swaps alternating between flipping and blitting.
>> This results in a back buffer continually being created and destroyed on
>> the client side, which is obviously highly undesirable.
> That's a bug. I know I've fixed similar things several times; which X
> server version are you running?
We're currently using 1.15.0.
>
>> My understanding is that, in the async case, a vblank event gets
>> requested for the next vblank. Any swaps that come in before this
>> event's received never get presented to the screen. In addition to this,
>> it's fairly unlikely that we catch things in the middle of a flip, i.e.
>> we've called drmModePageFlip() but not got back the flip event, meaning
>> that any queued vblank events, i.e. queued swaps, are going to be
>> completed via a flip. However, in the non-async case we're very likely
>> to catch things in the middle of a flip as a flip can only be completed
>> during a vblank period. The result is this ping-ponging between flipping
>> and blitting.
>> My current thoughts are that two things probably need to be done to fix
>> this:
>> a) If present gets a swap for the same MSC as a pending flip then it
>> should never try to do the presentation and send back
>> PresentCompleteModeSkip in the reply.
> No, this is wrong -- you want to *replace* the old presentation with the
> new one. Leaving the old one in place is not what the application wants
> at all.
>
>> b) If present gets a swap for an MSC in the future then it shouldn't be
>> prevented from completing the swap via a flip just because there is
>> currently a pending flip.
> Agreed. I think this is fixed in the current code.
I've looked at the git history and I can only see one present related 
commit since 1.15.0:
60014a4 Replace 'pointer' type with 'void *'

I couldn't see anything in your git repo either.
>> Does this sound reasonable, is there a better way of fixing this or am I
>> missing something?
> As I said, the current driver API was designed with the Intel driver in
> mind, and using the current DRM kernel interface's limitations. It may
> be that the current DRM kernel API just isn't suitable for async
> updates.

I'm probably missing something so correct me if I'm wrong but it seems 
that maybe the scheme being used needs reversing. Currently if we get 
multiple swaps for the same MSC then we only execute the most recent 
swap when the MSC is met (skipping the preceding swaps). However, this 
doesn't work in the non-async swap interval 0 case and it only works for 
none 0 swap intervals because there are only 2 back buffers.

By reversing things, so that if we get multiple swaps for the same MSC 
then we execute the oldest swap (skipping the later swaps), we're able 
to properly support the non-async swap interval 0 case. It also means 
that we can take advantage of the default drmModePageFlip() behaviour in 
the async none 0 swap interval case. Of course, this would mean that 
we'd always need 3 back buffers when flipping.

Thanks
Frank