Canceling the animation in the animationEnded handler triggers a crash.
This is due to multiple lists being iterated and manipulated at the same
time.
This adds a test case which simulates the crashy situation.
REVIEW: 126975
So far the key handler in the InternalWindowEventFilter used the
PointerInputRedirection's internal window. This had the result that
key events were only delivered to an internal window if the window
was under the cursor.
This change tries sending the event to the latest created and visible
window. Thus e.g. with nested context menus it goes to the current
sub menu as expected. The return value of sendEvent is used to filter
out the event.
This makes QCursor::pos and QCursor::setPos function correctly. KWin
actually wouldn't need it as KWin has the KWin::Cursor replacement, but
it allows Qt internal API to have it function correctly and also the
zoom effect does use QCursor::setPos.
The mapping is slightly inspired by the mapping in QtWayland.
But the mapping in QtWayland seems wrong. E.g. there is a linux kernel
button called BTN_BACK which is not mapped to Qt::BackButton.
Anyway we are not really interested in the mapping being 100 % correct
for the case in KWin. KWin internally uses only very few mouse buttons
and all others are only relevant to figure out whether buttons are
pressed. The button code itself is passed to the seat with the native
code.
The MoveResizeWindowTest is extended by a test case to verify that
the move only ends once all mouse buttons are released. So far this
is not yet the case as KWin has an incorrect mapping of buttons to
Qt::MouseButtons.
So far KSldApp was always either in state AcquiringLock or Unlocked
during the tests. Due to a fix in WaylandServer it now can also enter
the Locked state. But this is timing related and also depends on whether
the greeter works at all. E.g. on build.kde.org the greeter fails to
start, so it never enters the Locked state.
The adjusted test now considers that the state might have changed to
Locked and expects one additional signal to be emitted.
CI system is adjusted, so that OpenGL should work. At least latest
run didn't show the EGL warnings in the lock screen test. So let's
try to enable again. It's possible that this fails horribly. If that's
the case I'll revert again.
The dontCrashGlxgearsTest is kept on QPainter as locally it crashes
on teardown (needs fixing).
The logic should not be tied to whether libinput is used. It's relevant
for all Wayland backends whether they use libinput or not.
In addition this should generate a pointer motion event, so that proper
processing can take place and we get proper pointer enter events.
The test removes the second screen while the cursor is on it. This
should warp the pointer to the center of first screen and trigger
a focus enter event.
As can be seen by the expect failures currently it's bound to libinput
and also doesn't process the event as if it were a pointer event.
If the pointer is warped the position change should be treated like
a change coming from the input device. Our normal processing should
take place.
A problem in this case is the timestamp to pass to the wayland server.
Normally our timestamps come from the backend/libinput and we don't
know the next one. As an intermediate solution we just use the last
timestamp on the seat. In future a solution could be to not use the
backend's timestamp at all, but have our own timestamp handling.
When warping a pointer through Cursor::setPos it should be processed
just like any other pointer event. It should generate enter/leave event,
create motion events, etc. This is currently not the case as the test
shows.
A new test case which ensures that when stacking order changes the
pointer focus gets re-evaluated and updated. I was positively surprised
to notice that this already works.
Noticed two other problems while writing the test case:
* warping pointer does not re-evaluate the pointer pos
* deleting a ShellSurface (client) does not destroy the ShellClient (server)
When the screen gets locked any existing sequence gets cancelled
and the focused touch surface gets reset. While screen is locked
touch events are filtered to only go to lock screen or input methods.
Test case is added for touch event during lock screen.
Reviewed-By: Bhushan Shah
Instead of only making the active client the focused keyboard surface,
the method now also performs the lock screen security restriction.
Also just like udatePointerWindow the method becomes public, so that
it can be used from the LockScreenEventFilter and is connected for
lock state changes. This means as soon as the screen locks the current
focused keyboard surface will get a leave event and get an enter event
once the screen unlocks.
The auto test is adjusted to verify these new conditions.
Reviewed-By: Bhushan Shah
InputRedirection connects to lockStateChanged to udate the current
pointer window. This way we can ensure that the current pointer
surface gets reset as soon as the screen locks (c.f. the expect
fail in the autotest) and also that it restores to the surface under
the mouse once the screen is unlocked.
The relevant code was not yet lock screen aware and performed an
early exit. Part of the code was fine, e.g. findToplevel is lock
screen aware. So this change adjusts the methods for updating the
internal window and decoration to be lock screen aware, that is they
get reset. With that updatePointerWindow is also lock screen aware.
Thus the LockScreenFilter can also use updatePointerWindow just like
the normal handling and does not need to reimplement parts of it. As
it now relies on other code being correct it has an additional check
to verify that the current pointer surface is a surface which is allowed
to get events. If it isn't the events are not forwarded.
Reviewed-By: Bhushan Shah
The test creates a QRasterWindow which through KWin's internal QPA
is considered an internal window. In the test methods we simulate
various pointer events (enter/leave, press/release, wheel).
First test case is whether wheel events are forwarded correctly.
From the code it seems to me like up/down is inversed. Needs manual
testing.
As this requires working decorations it's possible that this test
will fail on build.kde.org.
Interesting approach I would not have expected to work. A dummy Effect
class is created in the test and an instance is passed to
EffectsHandler's startMouseInterception. It doesn't verify whether it's
an Effect it knows or has created, so it's totally happy with the dummy.
It shows that motion and press/release are passed to the Effect when
screen is not locked and doesn't while the screen is locked.
On the CI system our tests failed due to the greeter failing to start
due to OpenGL problems. This was because we waited for a window to show
which never happened. Thus the test failed.
This change makes use of the new lockStateChanged signal to determine
when the screen gets locked/unlocked.
It's still possible that the test fails as I'm not able to reproduce
the failure condition on the CI system.
First test case is to ensure that pointer motion events trigger a
leave event on the surface the pointer is on.
The test case shows errors in the input handling.
More tests need to be added.
Trigger quick tiling by moving the window. For moving the window only
keyboard keys are used.
The test experienced some problems with the Outline triggering crashes.
To work around them the test disables the Outline by specifying an
invalid configuration.
Ensures that all Wayland objects are destroyed and the cleanup handling
is performed before tearing down the Compositor. This fixes for example
a crash if a Surface with a Shadow is still around at tear down.
WaitForFinished blocks our main thread, but Xwayland wants to talk
to Wayland and blocks as well. So let's ensure events are processed
while terminating Xwayland.
Some effects do X11 calls in their cleanup code through external
libraries (e.g. KSelectionOwner). As we cannot control that we need
to ensure the effects are unloaded prior to destroying the Xwayland
connection.
We need to destroy the ClientConnections we create. Also we need
to disconnect our Xwayland error handling before destroying it, otherwise
it would trigger the abort for crashing Xwayland.
During Compositor tear down Xwayland is already destroyed. Thus it
doesn't make sense to try to delete the support properties: either
it freezes in xcb or it crashes because the connection is null.
At the same time we also ensure that the connection internally is
reset to null. Note: the one in kwinglobals.h caches and could cause
use-after-free errors. Any tear-down code must be migrated to
kwinApp()->x11Connection().
We need to destroy the compositor after Xwayland terminated and after
the internal Wayland connection is destroyed. This means when destroying
the Workspace we may no longer destroy the Compositor at the same time.
Also we need to ensure that other tear down functionality doesn't call
into the no longer existing internal client connection.
With this change kwin doesn't crash when exiting with Wayland and/or
X11 windows still open.
-use qstringliteral only when necessary (i.e. not in concat or comparison)
-use qbytearray instead of qstring when dealing with latin1 input and output (glplatform)
-use qstringref to extract numbers from strings (glplatform)
-define qt_use_qstringbuilder to optimize all string concatenations
-anidata: use ctor init lists, add windowType member initialization
REVIEW: 125933
Methods are no longer virtual. The only x11 specific usage in these
methods (resizeInc) is replaced by a virtual method. Default resize
increments is QSize(1,1) for AbstractClient.
Don't emit both geometryShapeChanged and geometryChanged: the one
is set up to call the other.
Also adjust tests because maximize changes triggers too many geometry
changed signals.
Method no longer virtual and only implemented in AbstractClient.
The implementaton works in a generic way nowadyas.
Added an autotest for the basic packTo behavior for packing against
a screen border. Packing towards other clients still needs adjustments
in the Placement code.
The signals operate on AbstractClient nowadays, so we can have one
implementation for both Client and ShellClient.
Only X specific connections are only done for Client.
So far only moving through useractions menu is possible and only through
cursor control (mouse events are lost).
A basic first autotest is added to validate the moving of Windows.
The problem we had was closing a glxgears through an Aurorae theme
crashed KWin inside QtQuick.
This test case simulates the sequence:
1. starts glxgears
2. wait till we have a Client for it
3. send mouse move to guessed close button position
4. send mouse press/release at that position
5. verify the window is closed
6. verify glxgears exits
With the given commit reverted this crashes, with it in place it passes.
Please note: on CI it might fail as glxgears is not yet installed. [1]
Also we cannot enforce using Aurorae from the test yet, though on
the CI system it should get picked automatically as no other deco
plugin should be installed.
[1] Sysadmin ticket already created
If the size is the same it's basically just a window movement. That's
nothing we need to roundtrip to the client, but can adjust the geometry
change directly.
The quick tiling test is adjusted to test this together with
sendToScreen. Each window is also sent to the next screen to verify the
state doesn't change and geometry is updated.
Note: the flag for quick maximization seems to get lost in this setup.
Very basic: all screens have same size and are ordered from left to
right. It's mostly meant to allow easy test cases with multi-screen.
The quick tiling test demonstrates how it's used.
Last commit removed a setGeometry call that I thought was only used to clear an edge (which we don't want) but was in fact crucial to the next test.
REVIEW: 125635
This is needed to properly restore to a valid geometry after quick
maximizing. Note: this is not yet perfect, actually it should be done
after initial placing, which means we do need a manage method like
Client.
The autotest for quick tiling is extended to cover maximize changes
and cover this case.
Base test verifies the quick tile positions. No maximization or direct
state changes tested yet.
Required to have Toplevel and AbstractClient exported. Otherwise we
cannot use the SignalSpy.
Add action type to screen edge show to allow raise/lower as well as
autohide
Add an action type to screen edge show to allow raise/lower as well as
autohide. This uses the same atom, using a mask to separate type and
location.
The logic for handling geometry changes is moved from the screenedge to
the client so that we can handle both types without screenedge needing
to know what the raise is for.
REVIEW: 124272
As expected in a025791d7b the tests
fail on the CI system due to llvmpipe. Let's try whether enforcing
O2 makes them run, if not update to switch to Q is following.
If the surface indicates that it doesn't want keyboard focus we shouldn't
give it keyboard focus.
This was a problem with Kate's autocompletion tool tip windows.
REVIEW: 125553
This is the beginning of a new testing era for KWin: finally we are
able to test against a running KWin. This works by making use of the
new virtual framebuffer backend for Wayland. It starts a specific
Application subclass which is mostly a fork of ApplicationWayland.
The individual tests are able to influence the socket name and the
size of the virtual screen. This is supposed to be done in
initTestCase. To know when KWin is fully started one can use the
workspaceCreated signal of KWin::Application. KWin is not started in
another process, but the kwin library is used, so the test has pretty
much full introspection to everything going on inside KWin. It can
access the Workspace, WaylandServer, fake input events through
InputRedirection and so on.
Once the test KWin is running it's possible to connect to it using
KWayland::Client library. This allows to introspect the Workspace
to see whether all worked as expected (e.g. correct stacking order,
active window and so on).
This first autotest is mostly meant to illustrate how to setup a
test and how one can use KWayland::Client to interact with the mock
KWin. For more tests it is suggested to move the connections to the
Wayland server in the init() and cleanup() methods.
The change also affects the qpa plugin: the specific check to only
run in binaries called kwin_wayland doesn't hold any more. This can
now be overwritten by an env variable.
Please note that this first test will probably fail in the CI system
as it might not have XWayland which is needed by KWin.
With Qt 5.5 the physicalSize of a screen is broken if the X system does
not provide the XRandR extension. This caused our screen edge test to
fail on the CI system (Xvfb) due to approach window being based on the
dpi.
The problem itself is addressed for Qt in:
https://codereview.qt-project.org/126808
This change just worksaround till the change has made it into our CI
system (at which point the expected fails will break).
Reviewed-By: David Edmundson
The test put a reference to the XCB::Window as data
to the _net_wm_transient property, not the window ID
This "works" (you get various garbage transients for
the leader in addition) as long as the window id itself
is the first item in the structure (because that is what
the test resolves to verify what it has done) bug fails
when the XCB::Window structure changes and anything
is the first item (or, btw., when the compiler feels
to re-align the structure and adds some padding...)
So let's fix the test by at least passing the proper
data reference.
is FIXED-IN: 5.4.1
REVIEW: 124888
also see REVIEW: 124864
# The first commit's message is:
do not switch desktop on resizing windows
REVIEW: 123599
# The 2nd commit message will be skipped:
# fix screenedge flipping merge
KWin::connection() uses the property to resolve the value instead of
using QX11Info. In practice this doesn't change anything at the moment,
but allows kwin_wayland to provide an xcb connection without depending
on QX11Info.
As we cannot make xcb_connection_t* available as a metatype, the
property's type is set to void*.
KWin::rootWindow() uses the property to resolve the value instead of
using QX11Info. In practice this doesn't change anything at the moment,
but allows kwin_wayland to provide a root window without depending on
QX11Info.
KGlobalAccel sets the timestamp as a property and we need to set our
x11Time to it otherwise following keyboard grabs might fail.
Requires 61e2a156678eef033b2629f7c72530dc78d7c3ac in kglobalaccel.
A wrapper class for MotifHints is added to xcbutils. This class manages
the information about the read Motif hints, so that Client doesn't need
to have a copy of the read states.
The class is designed in a way that during Client::manage we get rid of
another roundtrip.
REVIEW: 122378
It's a convenient class to encapsulate the ICCCM WM_SIZE_HINT.
Instead of exposing just the properties it provides accessors for
the interesting parts and applies sanity checks.
The test is flaky during the wait for Xephyr. Giving it more time should
hopefully make it more reliable.
The better way would be to switch to the command line argument
-displayfd pipeFd
unfortunately the Xephyr on the CI system does not yet support this
command line argument.
A new implementation of the Screens interface is added which uses XRandR
directly instead of relying on QDesktopWidget. The implementation is
provided in a new implementation file screens_xrandr.cpp.
XRandRScreens comes with a unit test. Unfortunately it's rather difficult
to provide a proper unit test against XRandR. Xvfb (which is obviously
used on the CI system) doesn't provide the XRandR extension. Also on a
"normal" developer system one would not want to just execute the test as
the results are not predictable (number of available outputs?) and the
test would mess up the setup resulting in nobody wanting to execute the
test.
As a solution to both problems the unit test starts Xephyr as a nested
X server. This allows to have at least some limited tests against XRandR.
Nevertheless there are a few things which I was not able to test:
* multiple outputs
* no output at all
The nested X Server approach makes the interaction rather complex. Qt
opens it's connection against the main X Server thus QX11Info provides
a wrong connection and also KWin::connection() which is heavily used by
xcbutils and thus all the RandR wrappers have the wrong connection. To
circumvent this problem the test is GUILESS. In case it would call into
any code using QX11Info, it would probably either runtime fail or crash.
REVIEW: 117614
The new test does not cover ScreenEdges completely, so far the
following areas are handled:
* creating of the edges
* reserving of edges
* trigger callback
* cursor pushback
* blocking of edges for fullscreen active clients
The default was set that the grid size for one desktop is (1, 2)
which doesn't make any sense at all - it should be (1, 1).
This most likely only affects the unit test as in production the
default layout is taken fron NETRootInfo.
This is a very interesting auto test as Screens uses both Workspace
and Client. Thus it operates in the "impossible to mock" area.
The solution is to provide mock includes in autotests and ensure that
when building the auto-test the mock header includes will be picked
first. There is now a mock class for Workspace and Client providing
just the API pieces used inside Screens.
As Screens is abstract and we cannot properly interact with
QDesktopWidget there is also a MockScreens class inheriting from Screens
and mocking the required functionality (by just operating on a list of
QRects).
The auto-test itself is only performing checks on the abstract class.
The mock class is indirectly tested by Screens calling into the virtual
methods. The test case is not yet complete, but looking quite good
already.
So far the Seat interface is provided together with pointer and
keyboard. As always touch is not yet implemented. The pointer interface
is still lacking the set cursor callback. Keyboard on the other hand is
complete.
Both Keyboard and Pointer have the concept of a focused surface and only
to the bound interface belonging to the same client as the focused
surface events are sent.
The change comes with a set of new auto tests also verifying the client
side which wasn't possible before as we couldn't fake events.
ShellSurfaceInterface is not yet completely implemented. Several parts
are still TODO, e.g. move/resize is missing, setting to maximized is
missing and also flags for fullscreen are missing.
The surface test is extended as far as possible.
The Display provides a method to create the shm pool and a
BufferInterface class is added to the server module. It is created
from the SurfaceInterface when a buffer gets attached to the surface.
The BufferInterface can be referenced and once its unreferenced it
sends a buffer release to the client and destroys itself.
For the case that the buffer is a shm buffer the BufferInterface
provides a convenience method to turn it into a QImage.
The auto test for Surface is extended by attaching buffers to the
surface and verifying that the content is correct.
As far as it's currently possible to implement. CompositorInterface is
able to create a surface and emits the created SurfaceInterface. It
does not yet support regions.
The SurfaceInterface is already more complete. It keeps track of the
double buffered states and emits signals when one of the values are
changed after the committing. It supports frame callbacks and has a
hook to mark the frame as rendered.
What's still missing are the regions (as it's not implemented in
CompositorInterface) and attaching the buffer as we do not yet support
creating shm buffers and pools.
The client side test is changed to use our own server and extended to
test damage and frame callback. The test needs to be extended for scale
and transform, but that is still missing in the client side
implementation.
So far this new module contains:
* Display
* OutputInterface
Display manages the server socket and server event loop. In general it's
the entry point to any part of the server.
OutputInterface is the abstraction for the wl_output interface on server
side. An OutputInterface is created through the Display.
The auto tests for ConnectionThread and Output are adjusted to use the
internal server instead of starting Weston. Especially the Output test
could be extended to test much more as we have absolute control over
the server now.
Technically the Surface itself does not have a size, it's the
ShellSurface or the size of the FullScreenShell's Output. But it
simplifies a lot if we keep track of the size in the Surface as that
way we can hide the fact which kind of Shell is used.
The user of the Surface must connect either the FullscreenShell's
Output or the ShellSurface to set the size on the Surface.
The fullscreen shell is only provided with at least Weston 1.5.
It should be checked on runtime, but that's difficult as Weston doesn't
fail if it cannot find the shell.so. So we disable it based on the
assumption that if Wayland library is not 1.5 the Weston is also not 1.5.
A Surface class is split out which holds a wl_surface and supports
attaching a buffer, setting the damage and emitting a signal when the
frame callback got called.
It doesn't come with a unit test yet as it first needs the ShmPool
and Buffer properly split out to easily set it up.
New classes Shell and ShellSurface are created. Both are in shell.[h|cpp]
to indicate their close relationship with the Shell having to create the
ShellSurface.
WaylandBackend is adjusted to hold a Shell* and ShellSurface* instead of
the lower level structs. This also required adjustements to the creation
of the Backend as it now doesn't set a default size any more. Thus the
backendReady signal may not be emitted before the initial configure
event arrived. This also makes it easier to support either the fullscreen
shell or wl_shell at the same time.
Of course a unit test is added for the two new classes. This needs to
be extended once we have more control over the mock Wayland server.
At the same time adding an autotest for the Output, moving the listener
into the Output class and providing enums for Subpixel and Transform.
KWin now requires wl_ouput interface version 2 as that allows us to emit
the changed signal in a better way.
The unit test is not yet capable of testing everything, we need a mock
Wayland server which is more flexible.
The FullscreenShell is a Wayland protocol provided by Weston to have
exactly one surface per output. This is exactly what KWin needs. So
in case the Wayland server we connect to provides the FullscreenShell
we prefer it over the normal Shell and mapping our surface as fullscreen.
The protocol is not yet part of wayland-client library, so the header
and source file needs to be generated. This is done during the build
process using the external tool wayland-scanner. The protocol
description is copied from the Westion 1.5 sources.
REVIEW: 119839
The Wayland::Registry class wraps wl_registry handling. It keeps track
of the interfaces in the registry and emits signals whenever a known
interface gets announced or removed. So far it only tracks the interfaces
which are used and needed by KWin.
The Wayland event queue is moved into a dedicated thread and a
new class is created for just creating the connection and listening
for events. The WaylandBackend creates the thread and uses an event
queue for the main thread.
REVIEW: 119761
This restores the behavior on KWin4: if I set an X property that doesn't have any data on a window, it's still information, so this makes the Xcb wrapper return an empty QByteArray that is not null.
EffectWindow::readProperty() now returns an empty QByteArray constructed the same way as it was in KWin4.
REVIEW:118645
BUG:335446
This removes all the hacks to add kwin4_effect_ to the name of the Effect
and adjusts the desktop files of the effect configuration's parent
component.
Note: the scripted effects still start with kwin4_effect_ prefix.
REVIEW: 117367
The Xcb::Property can wrap the xcb_get_property call and provides
convenient access methods to read the value of the reply with checks
applied. For this it provides a templated ::value method for reading a
single value or reading an array. There's also a ::toBool and
::toByteArray which performs the conversion directly with default values
for the type and format checks.
Xcb::TransientFor is changed to be derived from Property instead of
Wrapper directly, so that the reading of the property value can be
shared.
Xcb::StringProperty is a convenient wrapper derived from Property to
handle the reading of a string property providing a cast to QByteArray
operator. This replaces the ::getStringProperty from utils. Though the
separator functionality from ::getStringProperty is not provided as that
is only used in one function and handled there.
All the custom usages of xcb_get_property or getStringProperty are
replaced to use this new wrapper. That simplifies the code and ensures
that all properties are read in the same way.
REVIEW: 117574
So far the effects could just use the connection() and rootWindow()
provided by kwinglobals. Thus an internal detail from KWin core is
accessed directly.
To be more consistent with the rest of the API it's wrapped through the
EffectsHandler and with a convenient method in Effect.
The connection() is provided as xcbConnection() to free the very generic
name connection which could create confusion once we provide a wayland
connection to the Effects.
The rootWindow() is provided as x11RootWindow() to indicate that it is
for the X11 world.
REVIEW: 117597
Let's use the available API instead of duplicating code.
Nice side effect: client_machine.cpp doesn't include utils.h any more
which simplifies the unit test.
REVIEW: 117473
This is a specialized subclass of AbstractEffectLoader to load binary
effect plugins. It used the KPluginTrader to find all candidates to load.
The loader is able to detect incorrect ABI versions through the
pluginVersion() and uses the methods exposed by the new
KWin::EffectPluginFactory to check whether the Effect is supported and
should be enabled by default.
The unit test for this loader comes with two plugins: one is able to be
loaded and provides a supported and enabledByDefault method which can be
tweaked during the test to get all the conditions we want to test for.
The second plugin uses an incorrect plugin version and thus cannot get
loaded.
This implementation of the AbstractEffectLoader is able to to load the
scripted KWin Effects. It uses KServiceTypeTrader to find all the
candidates to load.
Effect loading gets split by the kind of effects KWin supports:
* Built-In Effects
* Scripted Effects
* Binary Plugin Effects
For this a new AbstractEffectLoader is added which will have several
sub-classes:
* BuiltInEffectLoader
* ScriptedEffectLoader
* PluginEffectLoader
* EffectLoader
The EffectLoader will be what the EffectsHandlerImpl is using and it just
delegates to the three other types of loaders. Thus the handler doesn't
need to care about the different kinds of effects. The loading is
supposed to be completely async and the EffectLoader emits a signal
whenever an Effect got loaded. The EffectsHandlerImpl is supposed to
connect to this signal and insert it into its own Effect management.
Unloading is not performed by the loader, but by the EffectsHandler.
There is one important change which needs to be implemented: the ordering
cannot be provided by the loader and thus needs to be added to the
Effects directly.
So far only the BuiltInEffectsLoader is implemented. It's not yet
integrated into the EffectsHandlerImpl, but a unit test is added which
tries to perform the various operations provided by the loader and the
BuiltInEffects. The test should cover all cases except the Check Default
functionality which is only used by Blur and Contrast effects. This
cannot be mocked yet as the GLPlatform doesn't allow mocking yet.
Only delegated to Cursor::pos() anyway, so let's just use that directly.
Fixes the annoyances of having to mock it in the unit tests which include
utils.cpp.
REVIEW: 116900
Used by Cursor to properly emit the mouseChanged signal which for
historic reasons includes the keyboard modifiers.
Again some fiddling around with the autotests and kcmrules needed to
make it compile. This needs improvement!
New inheriting class which uses the InputRedirection to track the cursor
position. It doesn't support warping of cursor.
This introduces a slight dependency loop in the startup. Cursor needs to
be created after the WaylandBackend to ensure that the operation mode is
set correctly. But the WaylandBackend itself is accessing Cursor. It
should be safe as inside the WaylandBackend it's only accessed after
callbacks.
