Wayland generic dmabuf protocol

[Pekka Paalanen]

On Mon, 9 Jun 2014 14:06:33 +0300
Pekka Paalanen <pekka.paalanen at collabora.co.uk> wrote:

> On Mon, 9 Jun 2014 11:00:04 +0200
> Benjamin Gaignard <benjamin.gaignard at linaro.org> wrote:
> 
> > On my hardware the patches you have (+ this one on gstwaylandsink
> > https://bugzilla.gnome.org/show_bug.cgi?id=711155) allow me to do zero
> > copy between the hardware video decoder and the display engine. I
> > don't have implemented GPU yet because my hardware is able to do
> > compose few video overlays planes and it was enough for my tests.
> 
> Right.
> 
> What I have been thinking is, that the compositor must be able to use
> the new wl_buffer and we need to guarantee that before-hand. If the
> compositor fails to use a wl_buffer when the client has already
> attached it to a wl_surface and it is time to repaint, it is too late
> and the user will see a glitch. Recovering from that requires asking
> the client to provide a new wl_buffer of a different kind, which might
> take time. Or a very rude compositor would just send a protocol error,
> and then we'd get bug reports like "the video player just disappears
> when I try to play (and ps. I have an old kernel that doesn't support
> importing whatever)".
> 
> I believe we must allow the compositor to test the wl_buffer before it
> is usable for the client. That is the reason for the roundtrippy design
> of the below proposal.
> 
> Because we do not even try to communicate all the possible restrictions
> to the client for it to match, we can leave the validation strictly as
> a kernel internal issue. Buffer migration inside the kernel might even
> magically solve some of the mismatches. It does leave the problem of
> what can the client do, if it doesn't fill all the requirements for the
> compositor to be able to import the dmabufs. But what restrictions
> other than color format we can or should communicate, and where does
> user space get them in the first place... *hand-waving*
> 
> But, this also leaves it up to the compositor to choose how/where it
> wants to import the dmabufs. If a compositor is usually compositing
> with GL, it will try to import with EGL on whatever GPU it is using. If
> the compositor uses a software renderer, it can try to mmap the dmabufs
> (or try this as a fallback, if the EGL import fails). If the compositor
> is absolutely sure it can rely on the hardware display engine to
> composite these buffers (note, buffers! You don't know to which
> surfaces these buffers will be attached to), it can import directly
> with DRM as FB objects, or V4L, or whatever. A compositor with the
> fullscreen shell extension but without the sub-surface extension comes
> to mind.
> 
> In summary, the compositor must be able to use the wl_buffer in its
> default/fallback compositing path. If the wl_buffer is also suitable
> for direct scanout, e.g. on an overlay, that is "just" a bonus.
> 
> With the round-trippy design, I am assuming that you can
> export-pass-import a set of dmabufs once, and then reuse them as long
> as you don't need to e.g. resize them. Is this a reasonable assumption?
> Are there any, for instance, hardware video decoders that just insist on
> exporting a new buffer for every frame?
> 
> I am tracking the proposal in
> http://cgit.collabora.com/git/user/pq/weston.git/log/?h=linux_dmabuf
> 
> So far I added back the event to advertise the supported
> drm_fourcc formats, since that is probably quite crucial.

Yeah, about that...

https://www.khronos.org/registry/egl/extensions/EXT/EGL_EXT_image_dma_buf_import.txt
provides no way for the compositor to query, what formats the EGL
implementation might support for importing dmabufs. I'm not sure GBM
has that yet either.

So there is no way a compositor could advertise the set of supported
formats, since it has no way of knowing, has it?

Any suggested solutions for this? Or would probing (export dmabuf, send
to compositor, wait for compositor to ack/reject) for suitable formats
be enough?


Thanks,
pq

> > 2014-06-06 17:30 GMT+02:00 Pekka Paalanen <pekka.paalanen at collabora.co.uk>:
> > > Hi,
> > >
> > > the previous attempt at introducing a generic wl_dmabuf protocol to
> > > Wayland didn't end too well:
> > > http://lists.freedesktop.org/archives/wayland-devel/2013-December/012390.html
> > > http://lists.freedesktop.org/archives/wayland-devel/2013-December/012455.html
> > > http://lists.freedesktop.org/archives/wayland-devel/2013-December/012566.html
> > > http://lists.freedesktop.org/archives/wayland-devel/2014-January/012727.html
> > >
> > > We are again interested in this, and I did a quick Friday evening draft
> > > to open the discussion again. The base of the draft was a quick look at
> > > https://www.khronos.org/registry/egl/extensions/EXT/EGL_EXT_image_dma_buf_import.txt
> > >
> > > The basic idea is, that a client has one or more dmabufs it wants to
> > > share with the compositor, making up a single logical buffer (a single
> > > image). The client chooses where and how to export those dmabufs. The
> > > dmabuf fds and metadata are sent to the compositor, the compositor
> > > assembles and tries to import them. If the import succeeds, a wl_buffer
> > > object is created. If the import fails, the client is notified that the
> > > compositor can't use these, it would be better to try something else.
> > >
> > > I assume that if the "import" succeeds, the compositor is able to use
> > > the buffers, e.g. at least turn them into a GL-texture or mmap them, if
> > > not also able to scan out or put on a hw overlay. This could be any kind
> > > of checking to verify that the buffers are usable. Finding out that it
> > > won't work after the client is already using the wl_buffer must not
> > > happen, as we have no way to recover from it: the client will get
> > > disconnected. So the point is knowing in advance, that the buffers are
> > > usable on both sides, preferrably before the client has filled them with
> > > data, but I suppose in the usual case the buffer is already filled.
> > >
> > > As creating a dmabuf-based wl_buffer requires a roundtrip in this
> > > scheme, I assume it only needs to be done rarely, and the same buffer
> > > can be re-used many times with proper synchronization.
> > >
> > > The crude draft is below. Some questions:
> > > - Does this sound sane to you?
> > > - What other metadata would we need? Thierry had some issues with
> > >   tiling formats I think.
> > > - This "check if the dmabuf is really usable" is needed, right? We
> > >   can't just assume that any dmabuf will work?
> > > - Do we need anything for fences here, or is the dmabuf fd enough?
> > > - Does someone already have something similar running?