The cache had the size 9, which was "big enough" in the past, but when
more ways pipelines could be constructed, the size was not enough. The
need to increase the cache size was hard to spot though, since adding
pipeline flag didn't give any hints about the cache being directly tied
to these flag values.
So, when enough flag bits were set when attempting to retrieve and put a
pipeline in the cache, it'd instead overwrite some arbitrary stack
memory, which would sooner or later result in a memory corruption
induced crash. Valgrind could not detect this particular memory
corruption, as it messed up stack memory, not e.g. freed heap memory, so
it instead got confused and thought plain stack values were unreadable.
Fix these two issues by making the cache size the combination of all
pipeline flags + 1, so that we can safely put any flag combination in
the cache.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1747>
If `meta_xwayland_start_xserver()` returned via an error path, some of
the socket FDs were leaked.
Fix that, and add `steal_fd()` calls to make it clearer that FDs passed
to `g_subprocess_launcher_take_fd()` are transferred to it. There were
no bugs with how `GSubprocessLauncher` was being used.
Spotted while working on
https://gitlab.gnome.org/GNOME/glib/-/issues/2332.
Signed-off-by: Philip Withnall <pwithnall@endlessos.org>
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1738>
Each next and current scanout buffer has a reference on them making sure
they stay alive. When dumb buffers were used on the secondary GPU state,
this didn't happen, leading to crashes due to unref:ing one time too
many, with backtraces such as
0) g_type_check_instance_is_fundamentally_a ()
1) g_object_unref ()
2) secondary_gpu_release_dumb ()
3) import_shared_framebuffer ()
4) update_secondary_gpu_state_post_swap_buffers ()
5) meta_onscreen_native_swap_buffers_with_damage ()
6) cogl_onscreen_swap_buffers_with_damage ()
7) swap_framebuffer ()
8) clutter_stage_cogl_redraw_view_primary ()
9) clutter_stage_cogl_redraw_view ()
10) _clutter_stage_window_redraw_view ()
11) handle_frame_clock_frame ()
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1746>
gnome-shell has this hack where it sets the environment variable
"NO_AT_BRIDGE" to "1" before calling meta_init() and then unsets it
after meta_init() returns.
This variable being set to "1" will then cause the ATK bridge in
at-spi2-gtk to fail to load, which GTK then ignores. This is on purpose,
since accessibility is supposed to be done done by GNOME Shell via
Clutter, not via GTK.
The problem is that, now, by default, setting "NO_AT_BRIDGE" to
"1" during meta_init() only has the desired effect on an X11 session,
where we always connect to the X11 server on startup (i.e. during
meta_init()). With Xwayland on-demand, we do not attempt to create the
GDK display during meta_init(), thus this hack falls apart.
Since there are no real altenatives to this hack, just move it to
mutter, which have a better idea when GDK displays are created or not.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1744>
Commit afa43154 tried to make sure the focus was properly changed when
calling focus_default_window() by checking the focused window just after
trying to set the focus.
However, the X11 “Inter-Client Communication Conventions Manual” version
2.0 (ICCCM 2 for short) states that some X11 client may want to use a so
called “globally active input” model in which case the client expects
keyboard input and set input focus even when it's not one of its own
window.
To comply with this, when dealing with such clients, mutter will not
change the focus and send a WM_TAKE_FOCUS message instead.
That mechanism will defeat the logic introduced by commit afa43154
because the focused window is not changed in this case. As a result, the
input focus will fallback to the no-focus window.
To avoid this, only check that the focus change occurred for windows
using a synchronous focus model.
v2: Split specific test for "globally active input" model (Florian).
v3: Remove the check for window->unmanaging which is useless (Jonas).
Fixes: afa43154 - "core: Make sure focus_default_window() worked"
Close: https://gitlab.gnome.org/GNOME/mutter/-/issues/1620
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1716>
X11 clients can use different models of input handling, of which some
may not result focus being set synchronously.
For such clients, meta_focus_window() will not change the focus itself
but rely on the client itself to set the input focus on the desired
window.
Add a new MetaWindow API to check when dealing with such a window.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1716>
Without these devices, things that depend on the existance of input
device classes won't know about the existance of e.g. pointer devices,
if the only pointer device is from a virtual one.
This requires handling situations where e.g. a device doesn't have a
device node thus can't be matched against a udev device.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1688>
Libinput will queue a few initial events when a seat is assigned to the
udev backend; a result of it probing udev adding detected devices. For
us to see these events, we need to dispatch libinput before going idle,
as nothing will show up on the libinput file descriptor until something
else (e.g. keyboard event or mouse movement) wakes us up.
Do this by adding a prepare() function to the libinput GSource, that
checks whether there are any events in the queue already, and return
TRUE if so is the case, causing us to dispatch before going fully idle.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1688>
The API for NotifyKeyboardKeycode() does not mention what type of
keycode is expected to be submitted.
So, clarify in the API that the keycode to submit is expected to be an
evdev keycode.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1732>
Currently, when a remote desktop user submits a keycode, it will be
interpreted differently, when using the x11 session, instead of a
wayland session.
In a wayland session, submitting a keycode will have the expected
result (as if the key was pressed locally).
In a x11 session, this is not the case. Instead of getting the expected
key, some other key will be pressed (or sometimes even none).
The reason for this is that the native backend interprets the keycode
as evdev keycode and the x11 backend interprets the keycode as xkb
keycode.
To ensure that both backends produce the same behaviour when submitting
a keycode, fix the x11 backend to always interpret the keycode as evdev
keycode, instead of a xkb keycode.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1732>
GObject signals pass the emitting GObject as the first argument to
signal handler callbacks. When refactoring the grab-op-begin/end signals
to remove MetaScreen with commit 1d5e37050d,
the "screen" argument was replaced with a "display" argument instead of
being removed completely. This made us call the signal handlers with two
identical MetaDisplay arguments, which is very confusing and actually
wasn't handled in a grab-op-begin handler in gnome-shell.
So fix this by not adding the MetaDisplay as an argument to those
signals, GObject will take care of that for us.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1734>
We're going to round the workspace backgrounds in the new overview for
gnome-shell 40.
So far corner-rounding was only possible for StWidgets because the
rounded clipping was done using cairo drawing. We now need rounded
clipping for ClutterActors too because backgrounds are drawn using
ClutterActors (or more specifically a ClutterContent). To implement
that, first a ClutterOffscreenEffect subclass together with a fragment
shader from GSK (see gskSetOutputColor() [1] in the GSK GL renderer
code) was investigated, and while that was generic and worked quite
well, it was extremely slow for the case of drawing wallpapers because
of all the FBOs that had to be allocated.
This is the new, more performant approach: Use the same fragment shader,
but perform the rounded clipping right in MetaBackgroundContent while
we're painting the wallpaper. This has almost no performance impact,
with the downside of not being a generic solution.
To allow for rounded clipping not only at the edges of the wallpaper,
but using any given bounding rectangle, the API exposes not only the
radius, but also a bounding rect.
[1] https://gitlab.gnome.org/GNOME/gtk/-/blob/master/gsk/resources/glsl/preamble.fs.glsl
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1717>
d42f1873fc introduced a semi circular
reference between the CoglFramebuffer, and CoglJournal, where
CoglJournal would keep a reference on the CoglFramebuffer when there
were any entries in the journal log.
To avoid risking leaking these objects indefinitely, when freeing
objects without doing anything that triggered a flush, CoglFramebuffer
had a "filter" on cogl_object_unref() calls, which knew
about under what conditions CoglJournal had a reference to it. When it
could detect that there were only the journal itself holding such a
reference, it'd flush the journal, effectively releasing the reference
the journal held, thus freeing itself, as well as the journal.
When CoglFramebuffer was ported to be implemented using GObject instead
of CoglObject, this "filter" was missed, causing not only awkward but
infrequent leaks, but also situations where we'd flush journals when
only the journal itself held the last reference to the framebuffer,
meaning the journal would free the framebuffer, thus itself, in the
middle of flushing, causing memory corruption and crashes.
A way to detect this, by asserting on CoglObject reference count during
flush, is by adding the `g_assert()` as described below, which will
assert instead cause memory corruption.
void
_cogl_journal_flush (CoglJournal *journal
{
...
_cogl_journal_discard (journal);
+ g_assert (journal->_parent.ref_count > 0);
...
}
Fix this by making CoglFramebuffer the owner of the journal, which it
already was, and remove any circle referencing that was there before, as
it is not needed given that the CoglFramebuffer pointer is guaranteed to
be valid for the lifetime of CoglJournal as the framebuffer is the owner
of the journal.
However, to not miss flushing before tearing down, which is important as
this flushes painting calls to the driver that is important for e.g.
using the result of those journal entries, flush the journal the first
time cogl_framebuffer_dispose() is called, before doing anything else.
This also adds a test case. Without having broken the circular
reference, the test would fail on g_assert_null (offscreen), as it would
have been "leaked" at this point, but the actual memory corruption would
be a result of the `cogl_texture_get_data()` call, which flushes the
framebuffer, and causes the 'mid-flush' destruction of the journal
described above. Note that the texture keeps track of dependent
framebuffers, but it does not hold any references to them.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1474
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1735>
Currently, the documentation for NotifyPointerAxis() just mentions that
a smooth scroll event is emitted.
However, this is not entirely correct. For each NotifyPointerAxis(),
mutter emits an emulated discrete scrolling event based on the
submitted accumulated smooth scrolling deltas.
Additionally, it doesn't mention how the motion deltas need to be
interpreted.
So, document the NotifyPointerAxis() notification better by mentioning
the emulation of discrete scroll events, how these discrete scroll
events are calculated and how the motion deltas need to be interpreted.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1727>
When a remote desktop user emits a virtual smooth scrolling event, a
smooth scroll event, that is not emulated, is emitted and on occasion
a discrete scroll event, that is emulated, is emitted.
As base for the discrete scrolling event, the smooth scrolling steps
are accumulated.
When the accumulated smooth scrolling steps surpass the
DISCRETE_SCROLL_STEP, the discrete scrolling event is emitted.
Currently, mutter uses for DISCRETE_SCROLL_STEP the value 10, which is
a terrible value to work with, especially for high resolution mouse
wheels.
When a triple resolution mouse wheel is used, each scrolling step will
have the value 3 1/3.
Three of such events won't however surpass the DISCRETE_SCROLL_STEP.
To fix this situation, add DBL_EPSILON to the calculation step, when
checking for the discrete scroll event to ensure that 3 smooth scroll
events, with each having the value 3 1/3, emit a discrete scrolling
event.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1727>
MetaVirtualInputDeviceX11 currently doesn't handle smooth scroll events
at all.
So, if a user of the remote desktop API uses smooth scroll events, then
only the wayland backend handles these events.
The user of the remote desktop API however, might not know which
backend is being used and actually the user should not even have to
care about it.
Actual smooth events cannot be emulated in the X11 events.
What can be done however is accumulating smooth events and then when
the accumulated steps surpass the DISCRETE_SCROLL_STEP value, emit a
discrete scroll event.
So, do exactly that, to make smooth scroll events work when the remote
desktop API is used with the x11 backend.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1727>
With 734a185915 an optimization was
introduced to only pick on events which can actually cause the pointer
to move. In case of touch events, the first event (TOUCH_BEGIN) will
already move the touchpoint though, and we'll send our crossing
CLUTTER_ENTER event to the actor this TOUCH_BEGIN happened on.
So fix this embarrassing bug that caused touch input to break by also
picking to find an event-actor on TOUCH_BEGIN events.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1733>
Aside from ENTER/LEAVE, there are only two kinds of events that can move
the pointer, motion events and touch update events. Everything else
keeps the pointer at it's current position.
The reason we pick inside _clutter_process_event_details() is that we
want to set the event actor. Now if an event can't move the pointer, it
also can't change the event actor (well, it can subsequently by
triggering changes to the scenegraph, but that's handled elsewhere), so
there's no need to pick a new event actor when we get those events.
Instead, simply reuse the actor that's already associated with the
current input device as the event actor for non MOTION/TOUCH_UPDATE
events.
Events where a device or a touchpoint goes away (like DEVICE_REMOVED or
TOUCH_END/CANCEL) also affect picking, they don't need a repick, but
instead the actor associated with the device/touchpoint needs to be
unassociated. This is ensured by invoking remove_device_for_event() on
those events and will not be affected by this change.
This should improve performance while scrolling quite a bit, since
scroll events come in unthrottled and we now no longer do a repick on
each one of those.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1729>
Given X11 nature, the pointer "leaves" the stage anytime it wanders into
a client window, or any other areas that are not deemed part of the
stage input region.
Yet we want to stay correct in those situations, e.g. have the clutter
side reasonably in sync, picking and highlighting to work properly, etc.
In order to achieve that, emulate motion events on XI_RawMotion. These
are as much throttled as our pointer tracking for a11y, in order to avoid
too many XIQueryPointer sync calls. This emulation only kicks in anytime
that X11 notifies us that we are not "on" the stage.
This replaces some sync_pointer() calls in GNOME Shell code that are
there just to compensate for this trait of X11, e.g. in the message tray
code.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1659>
Ensure we issue a motion event for the current pointer position,
as there might be situations where compositor modals get X grabs
from other clients stacked on top, or missed events in between
otherwise.
Ensure the Clutter state is still up-to-date afterwards here. This
replaces some sync_pointer() calls done in GNOME Shell code, always
done after modality changes.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1659>
Do these Wayland operations (that apply on both native and nested backends)
in the MetaCompositorServer subclass. We want to add more backend specific
behavior here in the future.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1659>
In the case that DnD is performed and succeeds, we want to release
the grab early, and let the transfer IPC happen in the back. For
that to happen without a hitch, drag source and offer must be left
related to each other after undoing the grab, even though the default
ungrabbing code does that automatically (indirectly, by unsetting the
drag focus).
In these cases, we used to manually unset the current source, so
this decoupling was skipped. Notably, one missed case is X11 DnD,
so we might end up with the situation there that DnD did succeed,
transfer is ongoing, but the source and offer are already decoupled,
this confused the machinery and missed the finishing XdndFinished
to be emitted to the X11 drag source.
The prior commits prepared for this source/offer decoupling being
a manual operation, this commit avoids doing this automatically
when ungrabbing.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1720>
This object is just being detached, with no code unref'ing it. Do
this whenever the XDnD selection goes unowned, usually a good
indication that the drag source no longer is one.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1720>
