There are two cases covered here:
1) The GstWlDisplay forces the release of the last buffer and the pool
gets destroyed in this context, which means it unregisters all the
other buffers from the GstWlDisplay as well and the display->buffers
hash table gets corrupted because it is iterating.
2) The pool and its buffers get destroyed concurrently from another
thread while GstWlDisplay is finalizing and many things get corrupted.
The main reason behind this is that when the video caps change and the video
subsurface needs to resize and change position, the wl_subsurface.set_position
call needs a commit in its parent in order to take effect. Previously,
the parent was the application's surface, over which there is no control.
Now, the parent is inside the sink, so we can commit it and change size smoothly.
As a side effect, this also allows the sink to draw its black borders on
its own, without the need for the application to do that. And another side
effect is that this can now allow resizing the sink when it is in top-level
mode and have it respect the aspect ratio.
Because we no longer have a custom buffer pool that holds a reference
to the display, there is no way for a cyclic reference to happen like
before, so we no longer need to explicitly call a function from the
display to release the wl_buffers.
However, the general mechanism of registering buffers to the display
and forcibly releasing them when the display is destroyed is still
needed to avoid potential memory leaks. The comment in wlbuffer.c
is updated to reflect the current situation.
This reduces the complexity of having a custom buffer pool, as
we don't really need it. We only need the custom allocation part.
And since the wl_buffer is no longer saved in a GstMeta, we can
create it and add it on the buffers in the sink's render()
function, which removes the reference cycle caused by the pool
holding a reference to the display and also allows more generic
scenarios (the allocator being used in another pool, or buffers
being allocated without a pool [if anything stupid does that]).
This commit also simplifies the propose_allocation() function,
which doesn't really need to do all these complicated checks,
since there is always a correct buffer pool available, created
in set_caps().
The other side effect of this commit is that a new wl_shm_pool
is now created for every GstMemory, which means that we use
as much shm memory as we actually need and no more. Previously,
the created wl_shm_pool would allocate space for 15 buffers, no
matter if they were being used or not.
This also removes the GstWlMeta and adds a wrapper class for wl_buffer
which is saved in the GstBuffer qdata instead of being a GstMeta.
The motivation behind this is mainly to allow attaching wl_buffers on
GstBuffers that have not been allocated inside the GstWaylandBufferPool,
so that if for example an upstream element is sending us a buffer
from a different pool, which however does not need to be copied
to a buffer from our pool because it may be a hardware buffer
(hello dmabuf!), we can create a wl_buffer directly from it and first,
attach it on it so that we don't have to re-create a wl_buffer every
time the same GstBuffer arrives and second, force the whole mechanism
for keeping the buffer out of the pool until there is a wl_buffer::release
on that foreign GstBuffer.
gstwaylandsink.c:480:14: error: comparison of constant -1 with expression of
type 'enum wl_shm_format' is always false
[-Werror,-Wtautological-constant-out-of-range-compare]
if (format == -1)
~~~~~~ ^ ~~
This allows waylandsink to fail gracefully before going to READY
in case one of the required interfaces does not exist. Not all
interfaces are necessary for all modes of operation, but it is
better imho to fail before going to READY if at least one feature
is not supported, than to fail and/or crash at some later point.
In the future we may want to relax this restriction and allow certain
interfaces not to be present under certain circumstances, for example
if there is an alternative similar interface available (for instance,
xdg_shell instead of wl_shell), but for now let's require them all.
Weston supports them all, which is enough for us now. Other compositors
should really implement them if they don't already. I don't like the
idea of supporting many different compositors with different sets of
interfaces implemented. wl_subcompositor, wl_shm and wl_scaler are
really essential for having a nice video sink. Enough said.
This essentially hides the video and allows the application to
potentially draw a black background or whatever else it wants.
This allows to differentiate the "paused" and "stopped" modes
from the user's point of view.
Also reworded a comment there to make my thinking more clear,
since the "reason for keeping the display around" is not really
the exposed() calls, as there is no buffer shown in READY/NULL
anymore.
1) We know that gst_wayland_sink_render() will commit the surface
in the same thread a little later, as gst_wl_window_set_video_info()
is always called from there, so we can save the compositor from
some extra calculations.
2) We should not commit a resize with the new video info while we are still
showing the buffer of the previous video, with the old caps, as that
would probably be a visible resize glitch.
Previously, in order to change the surface size we had to let the pipeline
redraw it, which at first also involved re-negotiating caps, etc, so a
synchronization with the pipeline was absolutely necessary.
At the moment, we are using wl_viewport, which separates the surface size
from the buffer size and it also allows us to commit a surface resize without
attaching a new buffer, so it is enough to just do:
gst_wayland_video_pause_rendering():
wl_subsurface_set_sync()
gst_video_overlay_set_render_rectangle():
wl_subsurface_set_position()
wl_viewport_set_destination()
wl_surface_damage()
wl_surface_commit()
... commit the parent surface ...
gst_wayland_video_resume_rendering():
wl_subsurface_set_desync()
This is enough to synchronize a surface resize and the pipeline can continue
drawing independently. Now of course, the names pause/resume_rendering are
bad. I will rename them in another commit.
Access is protected only for setting/creating/destroying the display
handle. set_caps() for example is not protected because it cannot be
called before changing state to READY, at which point there will be
a display handle available and which cannot change by any thread at
that point
This is because:
* GST_ELEMENT_WARNING/ERROR do lock the OBJECT_LOCK and we deadlock instantly
* In future commits I want to make use of GstBaseSink functions that also
lock the OBJECT_LOCK inside this code
* own_surface is not needed anymore
* gst_wl_window_from_surface is not used externally anymore
* many initializations to 0 are not needed (GObject does them)
This means that the given surface in set_window_handle can now be
the window's top-level surface on top of which waylandsink creates
its own subsurface for rendering the video.
This has many advantages:
* We can maintain aspect ratio by overlaying the subsurface in
the center of the given area and fill the parent surface's area
black in case we need to draw borders (instead of adding another
subsurface inside the subsurface given from the application,
so, less subsurfaces)
* We can more easily support toolkits without subsurfaces (see gtk)
* We can get properly use gst_video_overlay_set_render_rectangle
as our api to set the video area size from the application and
therefore remove gst_wayland_video_set_surface_size.
This drops the ugly GstWaylandWindowHandle structure and is much
more elegant because we can now request the display separately
from the window handle. Therefore the window handle can be requested
in render(), i.e. when it is really needed and we can still open
the correct display for getting caps and creating the pool earlier.
This change also separates setting the wl_surface from setting its size.
Applications should do that by calling two functions in sequence:
gst_video_overlay_set_window_handle (overlay, surface);
gst_wayland_video_set_surface_size (overlay, w, h);
This is the only way to get the negotiation working with the dynamic
detection of formats from the display, because the pipeline needs
to know the supported formats in the READY state and the supported
formats can only be known if we open the display.
Unfortunately,in wayland we cannot have a separate connection to
the display from the rest of the application, so we need to ask for a
window handle when going to READY in order to get the display from it.
And since it's too early to create a top level window from the state
change to READY, create it in render() when there is no other window.
This also changes set_window_handle() to not support window handle
changes in PAUSED/PLAYING (because it's complex to handle and useless
in practice) and make sure that there is always a valid display pointer
around in the READY state.
This fixes weird freezes because of frame_redraw_callback() not being
called from the main thread when it should with weston's toy toolkit.
It's also safer to know that frame_redraw_callback() will always be
called from our display thread... Otherwise it could be called after
the sink has been destroyed for example.
We are not supposed to redraw until we receive a frame callback and this
is especially useful to avoid allocating too many buffers while the
window is not visible, because the compositor may not call wl_buffer.release
until the window becomes visible (ok, this is a wayland bug, but...).
This is achieved by adding an extra reference on the buffers, which does
not allow them to return to the pool. When they are released, this reference
is dropped.
The rest complexity of this patch (hash table, mutex, flag, explicit release calls)
merely exists to allow a safe, guaranteed and deadlock-free destruction sequence.
See the added comment on gstwaylandsink.c for details.
start() makes sure that the minimum ammount of buffers requested is allocated.
stop() makes sure that buffers are actually destroyed and prevents
filling the file system when resizing the surface a lot, because the
wayland-shm-* files will stay on the file system as long as the wl_buffers
created out of them are alive.
This is the initial implementation, without the GstVideoOverlay.expose()
method. It only implements using an external (sub)surface and resizing
it with GstWaylandVideo.
The reference to the sink is not really needed anyway in waylandpool,
what matters basically is that the display is active as long as the
pool is active, so we really want to reference the display object
instead of the sink.
* make use of GstBufferPool::start/stop functions to allocate/deallocate memory
* get rid of struct shm_pool and do all operations cleanly inside WaylandBufferPool
* store a GstVideoInfo during configuration instead of the width & height
and use the stride from the video info instead of hardcoding its value
Wayland interface could offer two buffers pixels formats: WL_SHM_FORMAT_XRGB8888 and WL_SHM_FORMAT_ARGB8888.
Update waylandsink to support them and check if the format is really available.
https://bugzilla.gnome.org/show_bug.cgi?id=702112
Defer certain canvas particulars to the player (or the Wayland
compositor). Before this change, a fullscreen canvas was always
requested. This may not be desirable in all cases.
https://bugzilla.gnome.org/show_bug.cgi?id=690442
- bump wayland version to 0.95.0 which will lead to stable 1.0 release
- avoid memcopy and use propose_allocation for GstBufferPool allocation
- using WaylandBufferPool
- shm: Allocate shm buffers through new wl_shm_pool interface
(the shm buffer allocation is a two step process now: first
allocate a wl_shm_pool, then allocate a buffer from the pool)
https://bugzilla.gnome.org/show_bug.cgi?id=681453
