We can schedule an update from the cursor manager, but that doesn't mean
there will be an actual plane assignment changed at the time of the
update processing, since for example we might have "touched" a CRTC, but
already left it before the processing started, meaning we have nothing
to change after all.
Add a test case that checks that this works properly.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This removes the old hardware cursor management code and outsources it
to MetaKmsCursorManager. What the native cursor renderer still does,
however, is the preprocessing i.e. rotating/scaling cursor that wouldn't
otherwise be fit for a cursor plane.
The cursor DRM buffers are instead of being per cursor sprite now per
CRTC, meaning we don't need to stop doing hardware cursors if part of
the cursor is on an output that doesn't support it. This is why the
whole scale/transform code changed from being per GPU to per CRTC.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
It can be quite slow to set up the test environment inside the VM, as
well as outside, leaving very little time for the test itself. While
it'd be nice to not run the mock env etc outside the VM, let's just bump
the timeout for now, to avoid unnecessary timeout failures.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
If we turn of a CRTC, we might have invalidated the cursor manager for
the same CRTC, but that should not mean a cursor plane is assigned when
turning off the CRTC.
Add a test case for this.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This new manager object intends to take over management of the cursor
plane from the native cursor renderer. It's API is intended to be used
from the main thread, except for the _in_input() function, but mainly
operates in the KMS context, i.e. the KMS thread.
It makes use of an "update filter" that is called before each
MetaKmsUpdate is turned into a atomic KMS commit or a set of legacy
drmMode*() API calls. When the cursor position has been invalidated,
it'll assign the cursor plane in the filter callback, using an as up to
date as possible pointer position as the source for the cursor plane
position.
Cursor updates from the input thread schedules updates for the affected
CRTCs which will cause the filter to be run, potentially for cursor-only
commits.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This adds some plumbing to get the "default" paint flags for regular
stage painting, where one either wants to paint the overlay, or not.
If inhibited, the 'no-cursors' paint flag is used, otherwise the 'none'
flag. This will be used to allow having a per stage view hw cursor
state.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This makes it possible to post KMS updates that will always defer until
just before the scanout deadline. This is useful to allow queuing cursor
updates where we don't want to post them to KMS immediately, but rather
wait until as late as possible to get lower latency.
We cannot delay primary plane compositions however, and this is due to
how the kernel may prioritize GPU work - not until a pipeline gets
attached to a atomic commit will it in some drivers get bumped to high
priority. This means we still need to post any update that depends on
OpenGL pipelines as soon as possible.
To avoid working on compositing, then getting stomped on the feet by the
deadline scheduler, the deadline timer is disarmed whenever there is a
frame currently being painted. This will still allow new cursor updates
to arrive during composition, but will delay the actual KMS commit until
the primary plane update has been posted.
Still, even for cursor-only we still need higher than default timing
capabilities, thus the deadline scheduler depends on the KMS thread
getting real-time scheduling priority. When the thread isn't realtime
scheduled, the KMS thread instead asks the main thread to "flush" the
commit as part of the regular frame update. A flushing update means one
that isn't set to always defer and has a latching CRTC.
The verbose KMS debug logging makes the processing take too long, making
us more likely to miss the deadline. Avoid this by increasing the
evasion length when debug logging is enabled. Not the best, but better
than changing the behavior completely.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This is helpful when we add callbacks that should be dispatched in the
KMS impl thread.
This invalidates an assumption about callbacks not being in the impl
context, so some asserts for that are also removed.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This signal is emitted before terminating the thread, but also when
resetting the thread type. This is to allow thread implementations to
make sure they have no stale pending callbacks to any old main contexts.
This commit "terminates" the impl thread even if there is no actual
thread; this is to trigger the "reset" signal, also when switching from
a user thread to a kernel thread.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This means we can add COGL_TRACE*() instrumentation that is grouped
correctly in sysprof. If kernel threading is enabled, they will end up
in a "Compositor (KMS thread)" group (ignoring translations).
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
Real time scheduling is needed for better control of when we commit
updates to the kernel, so add a property to MetaThread that, if the
thread implementation uses a kernel thread and not a user thread, RTKit
is asked to make the thread real time scheduled using the maximum
priority allowed.
Currently RTKit doesn't support the GetAll() D-Bus properties method, so
some fall back code is added, as GDBusProxy depends on GetAll() working
to make the cached properties up to date. Once
https://github.com/heftig/rtkit/pull/30 lands and becomes widely
available in distributions, the work around can be dropped.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
Also add an API to inhibit the kernel thread from being used, and make
MetaRenderDeviceEglStream inhibit the kernel thread from being used if
it's active.
The reason for this is that the MetaRenderDeviceEGlStream is used when
using EGLStreams instead of KMS for page flipping. This means the actual
page flipping happens as a side effect of using EGL/OpenGL, which can't
easily be done off thread.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This will be necessary in order to default to 'kernel' and then switch
to 'user' if the thread instance can no longer be properly multi
threaded.
To avoid having the same thread impl creating and destroying
GMainContext's, this also means always creating a GMainContext for the
thread-impl. When running in user-thread mode, the GMainContext is
wrapped in a wrapper source and dispatched as part of the real main
thread GMainContext, and when in kernel-thread mode, it runs
independently in the dedicated thread.
This has the consequence that the wrapper source will always have the
priority of the highest impl context GSource, but only after it has
dispatched once. Would we need it earlier than that, we either need a
way to introspect existing sources in a GMainContext and their
priorities, or manually track known sources in MetaThreadImpl.
The wrapper source will never be below 0, as that'd mean it could reach
INT_MAX priority if it had no more sources attached to it, meaning it'd
never be dispatched again.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
While doing this, rename the old synchronous functions to more clearly
communicate that they expect to actually process the update during the
call, not just post it.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
While the default when passing NULL will be the main context of the main
thread, make it possible to specify another main context, so that
result handlers can be invoked on the right thread.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
Callbacks could be queued to be invoked either on the impl side or the
main thread side of the thread; change this to take a GMainContext,
which effectively means a callback can be queued to be invoked on any
thread that has a GMainLoop running on its own GMainContext.
Flushing is made to handle flushing callbacks synchronously on all
threads. This works by keeping a hash table of queued callbacks per
thread (GMainContext); when flushing (from the main thread), callbacks
on the main thread context is flushed, followed by synchronization with
all the other threads.
meta_thread_flush_callbacks() is changed to no longer return the number
of dispatched callbacks; it becomes much harder when there are N queues
spread across multiple threads. Since it wasn't used for anything, just
drop the counting, making life slightly easier.
Feedback to thread tasks are however always queued on the callers
thread.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This means each test is run 4 times:
* with atomic mode setting using a kernel thread,
* with atomic mode setting using a user thread,
* with legacy mode setting using a kernel thread, and
* with legacy mode setting using a user thread.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This isn't a problem for user space threads, as there are no race
conditions, but when kernel thread support is introduced, we must make
sure that e.g. the main loop is actually running before quitting it.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
This uses the queue that was introduced when migrating impl task
management from MetaThread to MetaThreadImpl, with the exception that
it's now fully used as an actual queue. It now has a GSource that sits
on the right GMainContext that is dispatched whenever there are tasks to
execute.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
It's the impl side that wants to add impl side idle sources, or fd
sources, etc, so make it part of MetaThreadImpl.
This changes things to be GAsyncQueue based. While things are still
technically single threaded, the GAsyncQueue type is used as later we'll
introduce queuing tasks asynchronously, then eventually queuing across
thread barriers.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
It currently does exactly what MetaKms and MetaKmsImpl did regarding the
context separation, which is to isolate what may eventually run on a KMS
thread into a separate unit. It works somewhat like a "user thread",
i.e. not a real thread, but will eventually learn how to spawn a
"kernel thread", but provide the same API from the outside.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2777>
Remote desktop version 2 added a new method ConnectToEIS .
ConnectToEIS allows clients to requests a file descriptor from the
compositor which can then be used directly from libei.
Once established, the communication between compositor and application
is direct, without the need to go through the portal process(es).
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
Accessibility should be handled on the receiving end, if needed. Make
sure this is the case by listening on some signals, verifying they are
only triggered if we're not capturing input.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
This adds the actual input capturing rerouting that takes events and
first hands them to the input capture session, would it be active.
Events are right now not actually processed in any way, but will
eventually be passed to a libei client using libeis.
A key binding for allowing cancelling the capture session is added
(defaults to <Super><Shift>Escape) to avoid getting stuck in case the client
doesn't even terminate the session.
The added test case makes sure that the pointer moves again after
pressing the keybinding.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
When a relative pointer motion gets constrained (e.g. a monitor edge or
barrier), save the constrained relative motion delta too.
This will later be used to send the remaining motion delta to input
capture clients.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
An input only grab is a ClutterGrab on the stage that doesn't have an
explicit actor associated with it. This is useful for cases where event
should be captured as if focus was stolen to some mysterious place that
doesn't have anything in the scene graph that represents it.
Internally, it's implemented using a 0x0 sized actor attached directly
to the stage, and a clutter action that consumes the events. An
input-only grab takes a handler, user data and a destroy function for
the user data. These are handed to the ClutterAction, which handles the
actual event handling.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
Adding a barrier and later enabling the input capture session will
create MetaBarrier instances for each added input capture barrier.
The barriers are created as "sticky" which means that when a pointer
hits the barrier, it'll stick to the point of entry, until it's
released.
The input capture session is also turned into a state machine with
explicit state, to more easily track things.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
This allows for a sticky barrier to hold the pointer until it is
released, but the owner of the barrier doesn't need a barrier event to
release it. It will be used to implement input capturing.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
A sticky barrier means that a pointer in motion intersecting a barrier
doesn't move once having hit it. The intention with this is to allow an
input capture clients to continue a motion once a barrier is hit.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
This API aims to provide a way for users to capture input devices under
certain conditions, for example when a pointer crosses a specified
barrier.
So far only part of the API is implemented, specifially the session
management as well as zone advertisement, where a zone refers to a
region in the compositor which edges will eventually be made available
for barrier placement.
So far the remote access handle is created while the session is enable,
despite the input capturing isn't actually active yet. This will change
in the future once it can actually become active.
v2: Remove absolute/relative pointer, keep only pointer (ofourdan)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2628>
This was a temporary fix until there was a better crossing event
delivery mechanism that accounted for actor changes beneath the pointer.
We nowadays have that, and don't seem to need this extra kick to get
crossing events triggered (and cursor changes, etc) when windows appear
or disappear under the pointer.
This commit is effectively a revert of commit
a64dba4d7a.
Closes: https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/6808
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3104>
With window_is_terminal gone, "strict" and "smart" focus mode have no
behavioural difference. Let's broaden the scope of strict focus mode,
such that windows never automatically focus unless they are an ancestor
to the transient.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3063>
As noted in the comments of window_is_terminal, this is a hack. This
code has not been touched for the better part of a decade. App res_class
tends to differ between Wayland and X11, so it is likely that none of
these apps have been recognised as terminals under Wayland ever. Also,
there are reports that strict focus mode also does not work under X11,
likely due to changes in these terminal apps over the years resulting
in different res_class than those manually specified in here. Let's remove
this hack and change strict focus mode accordingly.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3063>
Instead of using `clutter_actor_get_resource_scale()`, we now deduce the
intended buffer scale from the window by dividing the unscaled size by
the final actor size. This is more correct as while the return value of
`clutter_actor_get_resource_scale()` depends only on the monitor where
the surface resides, the actual scale of the surface is determined
solely by the application itself. `get_resource_scale` will differ from
the actual buffer scale if the application only supports 100% scaling
(Xwayland), or is performing scaling with wp_viewporter (clients using
fractional_scale_v1).
This also fixes a mismatch between the calculated buffer sizes between
`meta_window_actor_get_buffer_bounds` and
`meta_window_actor_blit_to_framebuffer` which causes broken
screencasting for Chromium 114 and later when using the native Ozone
Wayland backend.
Additionally, this commit also changes
`meta_window_actor_blit_to_framebuffer` from using a simple translation
to using an inverted matrix transformation of the transformation matrix
between the parent of the window actor and the surface actor to ensure
maximum sharpness for fractionally scaled windows.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3053>
Previously, restarting mutter in an X11 session resulted in
the previously set color temperature not being applied.
Fix that by applying the color temperature right after
the org.gnome.SettingsDaemon.Color proxy has been created.
Furthermore, only call `update_all_gamma()` from `on_gsd_color_ready()`
when the temperature has actually changed. Otherwise there is no need
since the current temperature has already been (or will soon be) applied
to all ready color devices.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3008>
We'd get a re-entry like scenario when destroying the PipeWire stream
object, where PipeWire would call the stream process vfunc. When this
happened, we had already destroyed the stream, so don't try to dequeue
or anything, just do an early exit. Fixes the following crash in the
test case client:
#0 pw_stream_dequeue_buffer() in /usr/lib64/libpipewire-0.3.so.0.367.0
#1 on_stream_process() at ../src/tests/screen-cast-client.c:348
#2 do_call_process() in /usr/lib64/libpipewire-0.3.so.0.367.0
#3 flush_items() in /usr/lib64/spa-0.2/support/libspa-support.so
#4 loop_invoke() in /usr/lib64/spa-0.2/support/libspa-support.so
#5 impl_send_command.lto_priv.0() in /usr/lib64/libpipewire-0.3.so.0.367.0
#6 suspend_node.lto_priv.0() in /usr/lib64/libpipewire-0.3.so.0.367.0
#7 pw_impl_node_set_state() in /usr/lib64/libpipewire-0.3.so.0.367.0
#8 client_node_removed() in /usr/lib64/pipewire-0.3/libpipewire-module-client-node.so
#9 pw_proxy_destroy() in /usr/lib64/libpipewire-0.3.so.0.367.0
#10 pw_stream_disconnect() in /usr/lib64/libpipewire-0.3.so.0.367.0
#11 pw_stream_destroy() in /usr/lib64/libpipewire-0.3.so.0.367.0
#12 stream_free() at ../src/tests/screen-cast-client.c:530
#13 main() at ../src/tests/screen-cast-client.c:803
#14 __libc_start_call_main() at ../sysdeps/nptl/libc_start_call_main.h:58
#15 __libc_start_main() at ../csu/libc-start.c:360
#16 _start() in /home/jonas/Dev/gnome/mutter/build/src/tests/mutter-screen-cast-client
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3095>
If the timelines don't get destroyed they keep references to frame
clocks. Later tests check for the destruction of those frame clocks and
then can fail if the frame clock is implemented slightly differenty.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3084>
In remote desktop sessions, streams can be created and destroyed
on-the-fly.
If a stream is gone, it is not necessarily an error.
So, don't treat that situation like an erroneous one.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/2911>
The previous logic didn't work correctly at least when priority-based
preeption wasn't supported by the DRM driver, such as in the case
of amdgpu. The call to glGetQueryObjecti64v would block on client
work which is already in progress (most likely for the next frame)
and delay notifying the ClutterFrameClock about presentation.
Conveniently, the Wayland transactions mechanism guarantees that all
fences of a dma-buf buffer are signalled before the buffer is
included in a frame, which means that dma-buf buffers are ready for
presentation when being directly scanned-out.
Direct scanout is only supported for dma-buf buffers too, which means
that all buffers going through direct scanout are effectively ready
and require no GPU rendering before presentation.
Assuming zero rendering time for dma-buf buffers going through direct
scanout simplifies the code and removes the need for
glGetQueryObjecti64v, thus avoiding the aforementioned issue where it
could block for longer than expected.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/2766
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3080>
Depending on the ordering of the surface-associated resources
being destroyed, we may fall into the following situation:
- wl_surface is destroyed
- destruction notifications for the surface runs
- The MetaWaylandKeyboard attempts to synchronize the window
focus
- The MetaWindow is not destroyed yet, so the focused window
remains the same, and the MetaWaylandKeyboard keeps the same
focus MetaWaylandSurface.
- wl_surface finalizes destruction, MetaWaylandSurface now has
a NULL resource
- xdg_toplevel destructor kicks in, it unmanages the window
- The current focus window is again looked up, forced to look
a different window
- The MetaWaylandKeyboard focus now changes, tries to leave the
old surface, but it has a NULL resource already, and raises
a protocol error.
If the order is inverted, the window being unmanaged triggers a
focus change into a different window, the MetaWaylandKeyboard
triggers a focus change while the MetaWaylandSurface is still
intact, it succeeds, and the window gets properly destroyed.
In order to make this independent of the order, it makes sense
to make MetaWaylandKeyboard do like the other objects tracking
focus surfaces, and have it care of its own little parcel. The
surface destructor changed to simply unsetting the keyboard focus
to NULL (guaranteeing that the old focus is left while the surface
resource is still up), and leaving potential focus changes to
the xdg_toplevel_destructor->unmanage->update_focus paths.
Doing that alone is basically a revert of commit 228d681b, thus
is still subject to keyboard focus being lost after a popup is
destroyed. Change the approach to trigger the focus sync (and
new focus surface lookup) so it happens from xdg_popup_destructor
specifically to popups and alike xdg_toplevel.
Fixes: 228d681b ("wayland: Trigger full focus sync after keyboard focus surface is destroyed")
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/2853
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3077>
