Hi all,
I'm working for Linaro on enabling a zero copy path in GStreamer by using dmabuf. To make this possible I have patched gst wayland sink to use wayland drm protocol: https://bugzilla.gnome.org/show_bug.cgi?id=711155
Today wayland drm protocol is limited to Mesa so I have decided to move it into wayland-core. My hardware doesn't have gpu support yet so I have patched weston (pixman) to allow usage of wl_drm buffers. With this I able to share/use a buffer allocated by DRM in gstreamer pipeline even if I don't have gpu and EGL.
What do you think about make wayland drm protocol available like this ?
Please note those patches are for wayland/weston 1.1.0
Regards, Benjamin
On Nov 27, 2013 10:53 AM, "Benjamin Gaignard" benjamin.gaignard@linaro.org wrote:
Hi all,
I'm working for Linaro on enabling a zero copy path in GStreamer by using dmabuf. To make this possible I have patched gst wayland sink to use wayland drm protocol: https://bugzilla.gnome.org/show_bug.cgi?id=711155
Today wayland drm protocol is limited to Mesa so I have decided to move it into wayland-core. My hardware doesn't have gpu support yet so I have patched weston (pixman) to allow usage of wl_drm buffers. With this I able to share/use a buffer allocated by DRM in gstreamer pipeline even if I don't have gpu and EGL.
What do you think about make wayland drm protocol available like this ?
Benjamin,
The problem here is that wl_drm is really a mesa extension. Well, more of an open-source linux graphics stack extension. The point is that there are other graphics stacks: Rhaspberry Pi, libhybris, other proprietary stacks, etc. and each of these graphics stacks has its own extension for passing hardware buffers around. This means that if you want your GStreamer sink to work on these other stacks with zero-copy, you will have to talk their protocols. Because wl_drm isn't global to all of wayland, it probably shouldn't go into the wayland core.
This does not mean, however, that wl_drm can't be exposed in a more sensible way. As of 1.3, we are now exporting wayland.xml to /usr/share/wayland and we can put other extensions there as well. We could have, for instance, a /usr/share/mesa.xml file that provides the mesa extensions. Then projects wanting to talk directly to mesa can generate the header and C files from the system-installed xml file. This would solve the problem of keeping everything in sync.
One last note (this one's been bugging me for a while). If we are going to export wl_drm in some way, we should probably rename it to something like mesa_drm. We really need to get out of the habit of using the wl_ namespace for everything that talks the wayland protocol. If we don't alter the details of wl_drm in the rename, it shouldn't be too hard to move everyone over from wl_drm to mesa_drm.
Hope that's sensible/helpful. --Jason Ekstrand
Wasn't EGLStreams supposed to solve the use case of passing hardware buffers around in a standard way?
On Wed, Nov 27, 2013 at 1:22 PM, Jason Ekstrand jason@jlekstrand.netwrote:
On Nov 27, 2013 10:53 AM, "Benjamin Gaignard" < benjamin.gaignard@linaro.org> wrote:
Hi all,
I'm working for Linaro on enabling a zero copy path in GStreamer by using dmabuf. To make this possible I have patched gst wayland sink to use wayland drm protocol: https://bugzilla.gnome.org/show_bug.cgi?id=711155
Today wayland drm protocol is limited to Mesa so I have decided to move it into wayland-core. My hardware doesn't have gpu support yet so I have patched weston (pixman) to allow usage of wl_drm buffers. With this I able to share/use a buffer allocated by DRM in gstreamer pipeline even if I don't have gpu and EGL.
What do you think about make wayland drm protocol available like this ?
Benjamin,
The problem here is that wl_drm is really a mesa extension. Well, more of an open-source linux graphics stack extension. The point is that there are other graphics stacks: Rhaspberry Pi, libhybris, other proprietary stacks, etc. and each of these graphics stacks has its own extension for passing hardware buffers around. This means that if you want your GStreamer sink to work on these other stacks with zero-copy, you will have to talk their protocols. Because wl_drm isn't global to all of wayland, it probably shouldn't go into the wayland core.
This does not mean, however, that wl_drm can't be exposed in a more sensible way. As of 1.3, we are now exporting wayland.xml to /usr/share/wayland and we can put other extensions there as well. We could have, for instance, a /usr/share/mesa.xml file that provides the mesa extensions. Then projects wanting to talk directly to mesa can generate the header and C files from the system-installed xml file. This would solve the problem of keeping everything in sync.
One last note (this one's been bugging me for a while). If we are going to export wl_drm in some way, we should probably rename it to something like mesa_drm. We really need to get out of the habit of using the wl_ namespace for everything that talks the wayland protocol. If we don't alter the details of wl_drm in the rename, it shouldn't be too hard to move everyone over from wl_drm to mesa_drm.
Hope that's sensible/helpful. --Jason Ekstrand
wayland-devel mailing list wayland-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/wayland-devel
Jason,
From my point of view wl_drm isn't link to Mesa, it is only about
exchange buffers by using a file descriptor and, for example, doesn't rely on EGL.
I understand that other graphic stacks could have defined their own way to for zero-copy (and so other protocols). I don't aim to make gstreamer wayland sink works for all of them but at least with wl_drm protocol which is quite generic.
Since dmabuf has been adopted in kernel we have the opportunity to rationalize also some code in userland and for that we need a common wayland protocol. Move wl_drm into wayland-core make it more easily accessible for all software even for those who don't use Mesa.
Benjamin
PS: I have an updated version of my patch for latest wayland
2013/11/27 Jasper St. Pierre jstpierre@mecheye.net:
Wasn't EGLStreams supposed to solve the use case of passing hardware buffers around in a standard way?
On Wed, Nov 27, 2013 at 1:22 PM, Jason Ekstrand jason@jlekstrand.net wrote:
On Nov 27, 2013 10:53 AM, "Benjamin Gaignard" benjamin.gaignard@linaro.org wrote:
Hi all,
I'm working for Linaro on enabling a zero copy path in GStreamer by using dmabuf. To make this possible I have patched gst wayland sink to use wayland drm protocol: https://bugzilla.gnome.org/show_bug.cgi?id=711155
Today wayland drm protocol is limited to Mesa so I have decided to move it into wayland-core. My hardware doesn't have gpu support yet so I have patched weston (pixman) to allow usage of wl_drm buffers. With this I able to share/use a buffer allocated by DRM in gstreamer pipeline even if I don't have gpu and EGL.
What do you think about make wayland drm protocol available like this ?
Benjamin,
The problem here is that wl_drm is really a mesa extension. Well, more of an open-source linux graphics stack extension. The point is that there are other graphics stacks: Rhaspberry Pi, libhybris, other proprietary stacks, etc. and each of these graphics stacks has its own extension for passing hardware buffers around. This means that if you want your GStreamer sink to work on these other stacks with zero-copy, you will have to talk their protocols. Because wl_drm isn't global to all of wayland, it probably shouldn't go into the wayland core.
This does not mean, however, that wl_drm can't be exposed in a more sensible way. As of 1.3, we are now exporting wayland.xml to /usr/share/wayland and we can put other extensions there as well. We could have, for instance, a /usr/share/mesa.xml file that provides the mesa extensions. Then projects wanting to talk directly to mesa can generate the header and C files from the system-installed xml file. This would solve the problem of keeping everything in sync.
One last note (this one's been bugging me for a while). If we are going to export wl_drm in some way, we should probably rename it to something like mesa_drm. We really need to get out of the habit of using the wl_ namespace for everything that talks the wayland protocol. If we don't alter the details of wl_drm in the rename, it shouldn't be too hard to move everyone over from wl_drm to mesa_drm.
Hope that's sensible/helpful. --Jason Ekstrand
wayland-devel mailing list wayland-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/wayland-devel
-- Jasper
On Thu, 28 Nov 2013 10:24:33 +0100 Benjamin Gaignard benjamin.gaignard@linaro.org wrote:
Jason,
From my point of view wl_drm isn't link to Mesa, it is only about exchange buffers by using a file descriptor and, for example, doesn't rely on EGL.
I understand that other graphic stacks could have defined their own way to for zero-copy (and so other protocols). I don't aim to make gstreamer wayland sink works for all of them but at least with wl_drm protocol which is quite generic.
Since dmabuf has been adopted in kernel we have the opportunity to rationalize also some code in userland and for that we need a common wayland protocol. Move wl_drm into wayland-core make it more easily accessible for all software even for those who don't use Mesa.
Hi,
if you aim particularly for dma-buf, then I would propose making a new Wayland protocol extension mesa_dma_buf or gbm_dma_buf or whatever name is appropriate, and drop all the unneeded parts compared to wl_drm.
I think dma-buf with render nodes will not need authentication, and it certainly does not use flink names. Would that work?
Note that still it would not be about a generic fd-based buffer passing but specifically dma-buf. If you want it really generic, then you would need something like Android's native_buffer_t (IIRC) which contains an arbitrary number of fds and integers, and then you are going to need platform-specific code to handle them right and whooops, the protocol itself is not meaningful anymore.
You might be interested in http://lists.freedesktop.org/archives/wayland-devel/2013-October/011593.html which is David Herrmann's proof of concept for passing mmappable dmabufs over wl_shm, and not needing any changes to Weston or Mesa AFAIU.
Also note that not all drivers support mmap'ing dmabufs, and mmap'ing in general is usually not a good thing for gfx buffers. I believe you would want to avoid mmap when possible.
If you wanted to be truely portable, I guess you might create a library pair similar to libEGL, where the actual protocol extension would again be just an internal detail of the library. But the question there is how would that library bind to rendering in client side and compositing in the server side.
Where do you want to draw the line?
Thanks, pq
2013/11/27 Jasper St. Pierre jstpierre@mecheye.net:
Wasn't EGLStreams supposed to solve the use case of passing hardware buffers around in a standard way?
On Wed, Nov 27, 2013 at 1:22 PM, Jason Ekstrand jason@jlekstrand.net wrote:
On Nov 27, 2013 10:53 AM, "Benjamin Gaignard" benjamin.gaignard@linaro.org wrote:
Hi all,
I'm working for Linaro on enabling a zero copy path in GStreamer by using dmabuf. To make this possible I have patched gst wayland sink to use wayland drm protocol: https://bugzilla.gnome.org/show_bug.cgi?id=711155
Today wayland drm protocol is limited to Mesa so I have decided to move it into wayland-core. My hardware doesn't have gpu support yet so I have patched weston (pixman) to allow usage of wl_drm buffers. With this I able to share/use a buffer allocated by DRM in gstreamer pipeline even if I don't have gpu and EGL.
What do you think about make wayland drm protocol available like this ?
Benjamin,
The problem here is that wl_drm is really a mesa extension. Well, more of an open-source linux graphics stack extension. The point is that there are other graphics stacks: Rhaspberry Pi, libhybris, other proprietary stacks, etc. and each of these graphics stacks has its own extension for passing hardware buffers around. This means that if you want your GStreamer sink to work on these other stacks with zero-copy, you will have to talk their protocols. Because wl_drm isn't global to all of wayland, it probably shouldn't go into the wayland core.
This does not mean, however, that wl_drm can't be exposed in a more sensible way. As of 1.3, we are now exporting wayland.xml to /usr/share/wayland and we can put other extensions there as well. We could have, for instance, a /usr/share/mesa.xml file that provides the mesa extensions. Then projects wanting to talk directly to mesa can generate the header and C files from the system-installed xml file. This would solve the problem of keeping everything in sync.
One last note (this one's been bugging me for a while). If we are going to export wl_drm in some way, we should probably rename it to something like mesa_drm. We really need to get out of the habit of using the wl_ namespace for everything that talks the wayland protocol. If we don't alter the details of wl_drm in the rename, it shouldn't be too hard to move everyone over from wl_drm to mesa_drm.
Hope that's sensible/helpful. --Jason Ekstrand
wayland-devel mailing list wayland-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/wayland-devel
-- Jasper
Hi,
On 28 November 2013 10:04, Pekka Paalanen ppaalanen@gmail.com wrote:
On Thu, 28 Nov 2013 10:24:33 +0100 Benjamin Gaignard benjamin.gaignard@linaro.org wrote:
From my point of view wl_drm isn't link to Mesa, it is only about exchange buffers by using a file descriptor and, for example, doesn't rely on EGL.
I understand that other graphic stacks could have defined their own way to for zero-copy (and so other protocols). I don't aim to make gstreamer wayland sink works for all of them but at least with wl_drm protocol which is quite generic.
Since dmabuf has been adopted in kernel we have the opportunity to rationalize also some code in userland and for that we need a common wayland protocol. Move wl_drm into wayland-core make it more easily accessible for all software even for those who don't use Mesa.
if you aim particularly for dma-buf, then I would propose making a new Wayland protocol extension mesa_dma_buf or gbm_dma_buf or whatever name is appropriate, and drop all the unneeded parts compared to wl_drm.
Given that dmabuf doesn't specify format codes or anything like that, you can't make it properly generic. My original thought was to make a generic wl_dmabuf interface, but given the format code issue, it seemed prudent to leave it as a part of wl_drm for now.
Cheers, Daniel
linaro-mm-sig@lists.linaro.org
-
Benjamin Gaignard -
Daniel Stone -
Jason Ekstrand -
Jasper St. Pierre -
Pekka Paalanen