`clutter_actor_destroy()` queues a stage update. Under certain
circumstances - i.e. when run in a very slow container - this can race
with the stage update triggered by the following
`clutter_virtual_input_device_notify_button()`, occasionally resulting in
`wait_stage_updated()` to return before the
`on_event_return_propagate()` callbacks ran, making the test fail.
This notably became more common since
8f27ebf87e (clutter/frame-clock: Start next update ASAP after idle period)
landed.
Thus wait for a stage update to happen after `clutter_actor_destroy()`,
preventing the race.
Fixes: f6da583d06 (tests/clutter/event-delivery: Add tests for implicit grabbing)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3332>
For frame updates in response to sporadic user interaction, this results
in input → output latency somewhere between the minimum possible and the
minimum plus the length of one display refresh cycle (assuming the frame
update can complete within a refresh cycle).
Applying a max_render_time based deadline which corresponds to higher
than the minimum possible latency would result in higher effective
minimum latency for sporadic user interaction.
This was discovered by Ivan Molodetskikh, based on measurements
described in https://mastodon.online/@YaLTeR/110848066454900941 .
v2:
* Set min_render_time_allowed_us = 0 as well, to avoid unthrottled
frame events. (Robert Mader)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3174>
Instead of g_get_monotonic_time. This makes sure last_presentation_time_us
advances by refresh_interval_us.
Doesn't affect test results at this point, but it will with the next
commit.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3174>
When more than one refresh interval has passed since
last_presentation_time_us.
I honestly can't tell if the previous calculation was correct or not,
but I'm confident the new one is, and it's simpler.
v2:
* ASCII art diagram didn't make sense anymore, try to improve
(Ivan Molodetskikh)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3330>
Every `mtk_x11_error_trap_push()` must be paired
with an `mtk_x11_error_trap_pop[_with_return]()` call
otherwise all future errors will be caught and ignored
even if they shouldn't be.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3328>
Certain kernel drivers can take an unreasonably long time to
complete mode setting operations. That excessive CPU time is charged
to the process's rlimits which can lead to the process getting killed
if the thread is a real-time thread.
This commit inhibits real-time scheduling around mode setting
commits, since those commits are the ones currently presenting as
excessively slow.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3037
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3324>
At the moment if a thread is made real-time there's no going back,
it stays real-time for the duration of its life.
That's suboptimal because real-time threads are expected by RTKit to
have an rlimit on their CPU time and certain GPU drivers in the kernel
can exceed that CPU time during certain operations like DPMS off.
This commit adds two new ref counted functions:
meta_thread_{un,}inhibit_realtime_in_impl
that allow turning a thread real-time or normally scheduled. At the same
time, this commit stores the RTKit proxy as private data on the thread
so that it can be reused by the above apis.
A subsequent commit will use the new APIs.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3324>
Most of the code writes "real-time" as "realtime" not "real_time".
The only exception is one function `request_real_time_scheduling`.
This commit changes that function for consistency.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3324>
If we queued a mode set, but didn't end up compositing all frames, we'll
have pending mode sets in a hash table waiting to be applied. If we
before all monitors again try to reconfigure things we should drop the
old pending mode sets and start fresh.
We already do this when we're doing so when generating views, but when
just unsetting modes, we didn't, so fix that.
Related: https://bugzilla.redhat.com/show_bug.cgi?id=2242612
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3318>
We had a function called "reset_modes()" on MetaRendererNative, but what
it expected to do was to unset all modes on all CRTCs. Despite this, it
had code to unset modes on unconfigured CRTCs, probably because it was
used for multiple things in the past.
Make this a bit easier to follow by renaming the function
"unset_modes()" and fold the function doing the unsetting into the
function itself.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3318>
Intel has started to advertise big gamma LUT sizes on some hardware
because the hardware supports segmented LUTs. This means they have a lot
more precision at certain segments then others. The uAPI can't expose
this functionality meaningfully so they chose to expose a huge number of
TAPs to sample from to their segmented LUT.
This increase in uAPI LUT size resulted in stack overflows because we
allocated the LUT on the stack. This commit moves it to the heap
instead.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3064
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3322>
1. Centralize stride calculation in one function.
2. For dmabufs query the stride instead of assuming a certain value.
3. For system memory buffers use the pixel format to calculate the
stride.
4. Stop negotiating `SPA_PARAM_BUFFERS_size` and
`SPA_PARAM_BUFFERS_stride`.
2. fixes an actual bug where we reported wrong max buffer sizes,
resulting in crashes in Gstreamer when doing area screencasts on AMD
GPUs.
The reasoning for 4. is that the values were possibly wrong for
dmabufs as the negotiation happens before we create any buffers.
Further more neither Mutter nor the common consumers required it.
The later either ignore the values (OBS), always accept (gstpipewiresrc)
them or calculate the exact same possibly wrong values (libwebrtc).
Closes: https://gitlab.gnome.org/GNOME/gnome-shell/-/issues/6747
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3311>
With EI support wired to XTEST, and oeffis being enabled in Xwayland
means that XTEST will always go through the XDG portal.
While this the intended behavior for the general use case of Xwayland
running rootless on a desktop compositor, that breaks when Xwayland is
running on a nested compositor, because the portal is for the entire
session and not limited to the nested Wayland compositor.
Enable XDG portal support in Xwayland only when we managed to connect
to the GNOME session manager, which means we are running in a full
desktop session, and not in any form of nested mode.
This is determined by simply using the status returned by set_gnome_env()
which will fail if not connected to a GNOME Session manager.
See-also: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1586
See-also: https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/1170
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3047
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3303>
The function set_gnome_env() is used to pass environment variables
though DBus using the "org.gnome.SessionManager".
If that fails, it means we are not running in a full environment, which
might be useful to determine whether Xwayland should enable the portal
support.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3303>
Enabling portal support unconditionally breaks when running nested, so
Xwayland has now a new command line option to explicitly enable support
for XDG portal at runtime.
Check whether the version of Xwayland we are using has such an option.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3303>
This makes sure the new update takes effect over the pending update for
any common properties. It matches the other users of
meta_kms_update_merge_from.
Fixes: 27ed069766 ("kms/impl-device: Add deadline based KMS commit scheduling")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3316>
When a configuration has a fractional scale, but we're using a physical
monitor layout, we can't use the scale, but if we do, we end up with
wierd issues down the line. Just discard the config if we run into this.
Eventually we probably want to store the layout mode in the
configuration so we can handle more seamless switching between physical
and logical layout mode, but first do this.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3057
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3299>
Mutter so far very rarely use the name `self`, so lets describe that in
the conventions document so that we can stay consistent.
Also mention how to deal with abstract/generic object instance pointers
vs sub type ones.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3297>
The surface actors may not have a valid allocation when running the
test. The preferred height, which `clutter_actor_get_size()` returns
in that case, can be wrong in certain cases, making us not add the black
background when it's actually needed.
Query the allocation instead, even at the expense of additional
relayouts.
While on it, sneak it some small cleanups.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3024
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3310>
The absolute modelview contains OpenGL coordinates, which have a higher
chance to not be invertible or, when doing so, introduce rounding
errors. These again often result in relative transforms becoming 3D
instead of 2D, making us miss optimized code paths down the line.
Thus cache stage-relative matrices instead, improving correctness and
possibly performance.
While on it also add some fast paths for cases where we can skip
calculating inverted matrices altogether and change variable names to be
more precise.
Fixes: dfd58ca8f1 ("clutter/actor: Extend caching in apply_relative_transformation_matrix")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3286>
There's two aspects from its documentation
(https://xkbcommon.org/doc/current/group__state.html#gae56031a8c1d48e7802da32f5f39f5738)
affecting us here:
1. "This function is similar to xkb_state_key_get_syms(), but intended for
users which cannot or do not want to handle the case where multiple
keysyms are returned (in which case this function is preferred)."
We are indeed in that field, and have been for a long time.
2. "This function performs Capitalization Keysym Transformations."
This is unlike the xkb_key_get_syms() function that we use, and
convenient here for parity with X11 since it behaves exactly that
way.
Fixes cases where the keysym for some keys is not properly capitalized
when caps lock is toggled, due to the output of capslock+key being
different from shift+key. An example of this is 'é' in french(azerty)
layout (bound to the '2' key). Even though shift+2 outputs '2',
capslock+é should output 'É'.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3058
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3306>
This is similar, but reserved for the crossing events induced by the
input shape changes on our overlay window. The mechanism in the previous
commit does again protect against this, so this mechanism may go away.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3267>
Focus follows mouse is meant to avoid focusing windows that happened
to pop up under the pointer, e.g. due to mapping, workspace changes,
etc... On X11, this has been done since ancient times through a
moderately complex synchronization mechanism, so mutter would know
to ignore crossing events caused on those situations.
This mechanism is much prior to XInput 2 though, where we may know
this in a more straightforward way: If the sourceid of the crossing
event is a logical pointer (i.e. equals deviceid), the crossing event
was triggered logically, and not through user input.
Perform this simpler check, and drop the existing mechanism to
ignore logically induced crossing events.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3267>
The virtual stream source with CURSOR_MODE_EMBEDDED uses
META_STAGE_WATCH_AFTER_PAINT as the callback for recording its frame. In
this stage of the paint though, there is no ClutterPaintContext anymore
(there only is a paint context during the paint, not afterwards).
The callback (actors_painted()) tries to get the redraw clip from the paint
context, and we end up with a NULL pointer crash.
We actually do still have a redraw clip at this point, so because everyone
uses the paint context to get the redraw clip anyway, just pass the redraw
clip to the stage watches directly.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3283>
