Whether relative pointers are supported doesn't depend on the input
backend. This change moves the instantiation of the extension to the
WaylandServer class to make code more consistent.
With the LEDs enum being defined in kwinglobals.h, wayland_server.h
won't need to include keyboard_input.h, which is good for compilation
times and wayland_server.h will drag less stuff, e.g. QtWidgets (input.h
includes QAction)
Currently, cpp source files are included partially in effect sub-folders
and src/effects/CMakeLists.txt, which is really confusing and hard to
follow.
With this change, effect targets will be defined in their own subfolders.
This makes build files more straightforward.
Effect sub-targets are eventually merged into a monolithic
kwin4_effect_builtins target.
Another reason for modularizing builtin effects is that no two effects
can generate moc files with the same name atm even though they are in
different sub-folders. This can be potentially useful in the future, e.g.
making effects static plugins in order to decouple builtin effects from
libkwin (every effect subfolder would need a main.cpp or plugin.cpp file,
which will include the associated moc file).
The GLTexture class is forward declared and the ShowFpsEffect class has
a QScopedPointer<GLTexture> field. We either need to include the
kwinglutils.h header file or add a destructor that does nothing and
define it in the cpp file, where kwinglutils.h is included.
At the moment we'll be setting the YInverted setting, but in practice
that won't have any effect as it only changes the render matrix and
we'll end up streaming inverted textures.
This change addresses it by rendering it into another texture first to
resolve this situation and then download that new texture instead.
The English word "pack" is not really the correct word for these
actions, and does not succeed in communicating what they will do. Since
the actions simply move the active window as far as it will go in the
specified direction, the actions can be renamed to say that instead.
Also rename the action names in the code to match their new UI text for
clarity.
With lv3:ralt_alt ("Right Alt never chooses 3rd level") option set, we
get more layouts from libxkbcommon than it was configured, see:
https://github.com/xkbcommon/libxkbcommon/issues/262
It might be correct lib's behavior, still.
The extra layouts are redundant, so we strip them out from usual usage.
BUG: 440027
This change merges the two OpenGL backends into one making the current
default of GLCore the overall default. It becomes the first context to
try to create. If it fails, it will automatically fall back to the
(previous) OpenGL 2 backend.
Reasoning: the differentiation of OpenGL 2 and 3 is a very technical one
and hard to understand for users. It is not obvious which one is better
or should be used. This results in many user discussions like "Which
backend to use?"
Back when the OpenGL 3 backend was introduced the dedicated feature made
sense. It was a new code base using new driver features. Nowadays the
code base in KWin is robust and mature and so are the drivers. A driver
advertising support for OpenGL 3 will support OpenGL 3. We don't have to
plan for driver breakage in this area any more.
Also our code evolved through the context attribute builder which gives
us the possibility to more easily fall back in case we cannot create the
context. Thus the need to select the backend is not so important as it
used to be when the feature got introduced.
If a user still wants to force OpenGL2, it is still possible by setting
the appropriate environment variables like MESA_GL_VERSION_OVERRIDE.
This change brings the improvement that the backend selection is now
completely removed from the compositing KCM.
Currently, the pointer is moved to the center of the workspace. However,
on a multi-monitor setup, the workspace center may not be the same as
the center of the output it's on. That's the case with my setup -
1920x1080, 1920x1080, 1080x1920 (the monitors are laid out in a row from
left to right).
This change improves the default placement of the pointer by actually
moving it to the center of the screen that contains the workspace center
so the pointer doesn't end up at some "random" position from the user's
point of view.
Edge geometry has width and height, so we need the correction for
approach geometry of bottom and right corners.
The other edges are fixed as well.
BUG: 442973
On X11, there are four input models. With some input models, it's okay
if the window manager calls XSetInputFocus(), with others, the wm has to
ask the client to make a XSetInputFocus() request.
If kwin wants a client to take input focus, kwin will add the client
to the should_get_focus list, which contains all the windows that
are about to get input focus. Clients are popped from the list upon
receiving FOCUS_IN events.
A client will be added to the should_get_focus list even if kwin thinks
that the client already has input focus because communication between
the wm and xorg is async, anything can happen with input focus in meanwhile.
On the other hand, the wm may sometimes focus the null window if no
window should contain the input focus. The issue is that the
should_get_focus list is not cleaned up in that case, which can lead to
Workspace::mostRecentlyActivatedClient() returning wrong client and
possibly other async related issues.
We don't have such madness on Wayland as the compositor is in charge of
handling input focus.
This change makes Workspace::focusToNull() clear the should_get_focus,
which is reasonable. We need to deactivate "in-flight" focus requests.
This also fixes the bug where fullscreen Wayland windows don't go above
docks and panels due to Workspace::mostRecentlyActivatedClient() returning
bad client.
BUG: 439405
BUG: 395919
Currently, kwin crashes at shutdown because the idle poller plugin is unloaded when exit handlers are run, after the waylandServer() is destroyed. This results in null dereferencing.
BUG: 443268
Xwayland will re-create the wl_surface object if the X11 window is
unmapped and mapped. That, and the fact that the order in which the
WL_SURFACE_ID client message event is received and the wl_surface object
is created is undefined can cause the following bug:
* WL_SURFACE_ID is received
* the old wl_surface object is destroyed, m_surfaceId is reset to 0
* new wl_surface is created but because m_surfaceId is 0, it won't be
associated with the x11 window
This change ensures that kwin will associate the wl_surface with x11
window by making it not reset cached surface id when the old wl_surface
is destroyed.
However, we cannot leave m_surfaceId as is because wayland aggressively
re-uses object ids so kwin can associate wrong surface with x11 window.
To prevent that, this change also makes Toplevel::setSurface() reset
cached surface id.
CCBUG: 442936
CCBUG: 426069
It's not necessary, the placeholder output already gets created in DrmBackend::removeOutput
if necessary. More improtantly it's missing the check for shutdown, which
may cause issues if the computer is turned off while no outputs are connected.
clientArea() was changed to forbid "-1" screen ids, but it seems like
the thumbnail aside effect can pass it to refer to the active screen.
This change makes the thumbnail aside effect handle "-1" screen id case
explicitly. It will be also useful for EffectScreen transition.
BUG: 443166
If the virtual desktops rule is created, only the last virtual desktop
the window is on is going to be checked.
With this, all virtual desktops that the window is currently on will be
checked.
Technically, it's an api breaking change, but the window rules kcm is
the primary user of queryWindowInfo() so I don't think it's a big deal.
KPluginMetaData::findPlugins only returns valid json metdata, the previous
check was accidentally kept for the condition that the plugin provides a theme engine.
BUG: 442978
FIXED-IN: master
Blobs are not reference counted if used by other drm master, if kwin
re-uses a deleted blob in an atomic commit, it will fail. For example,
on my computer, this happens when kwin starts after xorg.
Besides that, kwin may try to destroy blobs that it doesn't own, which
is not fatal but it's strange to do so.
CCBUG: 442603
CCBUG: 439873
While findWorkingCombination should never fail, in the case it does
KWin should not crash. To achieve that simply restore the old config
in case of failure.
CCBUG: 439873
When libinput tears down, it may access the Session object. This change
re-jitters the shut down logic so the Session object is guaranteed to be
valid when libinput stuff gets destroyed.
BUG: 442104
ksmserver provides the backing window (the hidden black one) for the
lockscreen. We don't want to animate this fading out.
It no longer provides the logout prompt since several plasma releases.
DrmGpu::updateOutputs is now changing the state of some drm properties
without changing the fitting output state (namely RenderLoop inhibition)
which can lead to rendering being inhibited indefinitely in some cases
Software-based clipping is naive. It maps the clip region from the
global screen coordinates and intersects it with window quads. If the
window is transformed, the quads won't be clipped as expected.
Unfortunately, the OpenGL scene enables software-based geometry clipping
if the screen transformed. It's not clear why it does so, perhaps as a
performance optimization? Either way, this produces incorrect results
when the screen is scaled.
BUG: 440940
This makes ExpoLayout easier to maintain. As is, the main issue with it
is the number of layout algorithms.
With this, the ExpoLayout is going to have two layout algorithms - one is
very simple and the other not so.
Startup code in plasmashell was changed so xsetroot is not called
anymore, which is sort of fine.
Unfortunately (or not?), it exposed a bug in kwin. Cursor::x11Cursor()
only works in the standalone X11 session.
On Wayland, Cursor::x11Cursor() will return XCB_NONE which results in
seeing cross cursor when there should be arrow cursor.
This change moves xcb_cursor_t look up code from X11Cursor to the base
Cursor class. In hindsight, I would like to introduce a window manager
class where the xcb cursor and other x11 specific code can be moved in
the future for better encapsulation of platform-specific code.
CCBUG: 442539
For some reason, layers are change after applying Show Desktop effect, so
we will have AboveLayer for our popups and Desktop, and NotificationLayer
for Dock which hides us.
Also, OSD and critical notifications were always hidden by our popups.
This patch solves both the issues.
Usual notifications will still be hidden, but hopefully it's not a big
issue.
BUG: 442605
reinterpret_cast<>() will fail if the types we cast from and to have
mismatching sizes.
Unfortunately, there are platforms that have Window and
EGLNativeWindowType of different size. This results in compilation
errors.
In order to work around those problematic platforms, this change
replaces reinterpret_cast cast with a c style cast.
EffectQuickView is used by Aurorae. As long as Aurorae decorations are
rendered by kwin, it's not a problem. However, kde-gtk-config kded
module can also render them. This creates a problem. If effects object
is not created, accessing any of its getters or setters will result in a
segmentation fault.
This change rewrites the share context check so the effects object is
not accessed anymore.
One could argue that kde-gtk-config has to provide a dummy EffectsHandler
but it doesn't use effects and libkwineffects is not a dependency of
KDecoration2. So, providing a dummy EffectsHandler makes no sense.
In hindsight, we need to revisit the usage of EffectQuickView in Aurorae
as accessing the EffectsHandler in EffectQuickView is a totally valid
usecase and integration of QOpenGLContext.
BUG: 441585
When running in plasma, the animation speed slider is in General
Behavior KCM.
The Defaults button should not reset the animation speed slider because
it's hidden.
BUG: 442600
With per-screen rendering, the projection matrix is no longer created
with the workspace geometry, i.e. all screen geometries united, so the
center of the workspace geometry may not map to (0, 0) in the clip
coordinates.
GLRenderTarget::virtualScreenGeometry() can be used to query the rect
that was used to create the projection matrix.
BUG: 442770
KMS can only ever have one frame pending at the moment. If we
update the outputs while a pageflip is still pending on any output
the atomic commit will fail with EBUSY, which both invalidates
the testing for output configuration and makes applying the wanted
configuration fail.
This can be removed again once KMS gains the ability to do mailbox
presentation; that will likely still take a while though.
BUG: 442677
Currently, data from client buffer EGLStream is copied to a surface
texture. An fbo is used for that purpose. The main issue with it is that
it doesn't restore the old framebuffer binding.
The surface texture can be updated in the middle of a compositing cycle.
If the framebuffer binding is not restored, any window that must be
rendered in an offscreen texture won't be rendered to the offscreen
texture.
This change makes EglStreamSurfaceTextureWayland restore the fbo binding
so the DeformEffect and software screen rotation work as expected with
the proprietary NVIDIA driver.
BUG: 442697
This makes the compositingPossible() output more elaborative,
and adjusts their priorities to qCWarning.
I thought about dropping the "Compositing is not possible"
(qCCritical) message. It's not critical for X11 and Wayland
depends on OpenGL. OTOH it's from a different component. So in
the end I settled for qCWarning again; there is no real way
to know, if the platform will consider this a critical problem
(and eventually abort).
Then I saw X11StandalonePlatform::compositingNotPossibleReason,
which has more user friendly, QT Richtext / HTML encoded output,
but this seems overkill for the terminal; now I would like to
know, where this is actually used...
While at it report "manually" suspended compositing via qCInfo
instead of qCDebug.
After updating my Debian buster to bullseye, I noticed missing
shadows / composite for KDE, which made overlapping windows hard
to recognize.
.xsession-errors just had "kwin_core: Compositing is not possible".
Using the suggested QT_LOGGING_RULES="kwin_*.debug=true" didn't
provide any more clues; nothing I tried did.
Eventually I ended reading the kwin source and found, that
disabled composite via kwinrc :: [Compositing] :: OpenGLIsUnsafe
was not reported.
IMHO all these disabling cases should use qCCritical(KWIN_CORE),
like the "Compositing is not possible" message, but for the time
being, this simply logs the OpenGLIsUnsafe reason, just like all
the others.
With the recent AbstractOutput changes, Workspace::clientArea() overload
that takes only the window and no additional output was changed to
return clientArea(opt, window, window->output());
prior to that, it had been looking up the output at the center of the
frame geometry.
As it turns out, AbstractOutput::sendToOutput() blocks geometry updates,
which in its turn means that Toplevel::output() will be updated only
after geometry updates are unblocked. For the most part, it's not a big
deal until you need to use Workspace::clientArea(opt, const Toplevel*)
Since the output won't be updated until geometry updates are unblocked,
clientArea(opt, const Toplevel *) may return outdated info.
One could argue that we just simply need to update m_output similar to
m_frameGeometry when geometry updates are blocked, but... it's not going
to work on wayland! On Wayland, GeometryUpdatesBlocker is totally
unnecessary. Even if a window is resized, m_frameGeometry will be left
unchanged until the client repaints the window.
Instead, one need to pass the AbstractOutput to clientArea() if it's
important.
This change makes X11Client::changeMaximize() query the MaximizeArea of
the output containing the center of the move resize geometry, similar to
what the XdgToplevelClient does.
BUG: 442534
There are some unresolved issues where driving outputs can fail because
of bandwidth constraints. These don't appear to happen with implicit
modifiers, or at least they don't happen as often.
libinput will send release and proximity out events after the device is
disconnected.
This just takes into account that tablet might be nullptr.
BUG: 442573
The stored geometry is irrelevant in the window is maximized.
This is especially relevant if we:
* open a virtual keyboard
* rotate the display (then it looks ok as it was already addressed)
* then remove the keyboard.
At that point we want to reset to the available area rather than the
previous.
By supporting all the required attributes in the DecorationThemeMetaData class
the custom struct can be dropped.
Task: https://phabricator.kde.org/T14744
We can just put the parameter in the variant list. Because the method is virtual, the keyword
parameter must exist in the method definition.
Task: https://phabricator.kde.org/T14744
b38bb416 introduced flicker when restarting compositing because the
DrmGpu::findWorkingCombination method usually creates new buffers for
the commit, without rendering into them. Instead of that, re-use
existing buffers where possible
The configs returned by eglChooseConfig are sorted from highest to lowest buffer sizes.
Filter the configs to find a suitable candidate that matches the requested attributes of the surface.
This fixes the corrupted Aurorae window decorations on Wayland with the proprietary Nvidia driver.
Conceptually, a configure event inherits its parent's state and adds
some of its own. This allows the compositor to skip intermediate
configure events in ack_configure handler and jump to the last one.
Currently, the only field that XdgSurfaceConfigure objects need to
inherit is flags. The geometry info and window states are filled in by
role commit implementations to their latest values.
XdgSurfaceConfigure::flags indicates if the configure event moves the
window. This flag is important to resolve conflicts between geometry
updates initiated by the user and the ones that are made in response to
acknowledged configure events, e.g. after maximizing the window, etc.
(effectively, if the user moves a window, kwin will cancel scheduled
moves in configure events)
If configure flags are not inherited, we can end up with the following
case:
* configure event A (flags: {position})
* configure event B (flags: {})
If the client acknowledges configure event B, kwin will skip configure
event A, and thus it won't move the window to the right place. This is
the root cause of fullscreen mode misbehaving with apps such as google
chrome.
discardPixmap() increases the reference counter, but there's no matching
place where the ref count is decreased, which results in the previous
pixmap not being released even if it's not needed anymore.
Instead of setting pipelines one by one, use DrmGpu::updateOutputs to
set all the outputs on a GPU with a single atomic commit. This makes
the switch both faster, more reliable and in case the other drm master
changes the output configuration, prevents blanking.
eglCreateWindowSurface() wants a Window (unsigned long), but with
EGL_NO_PLATFORM_SPECIFIC_TYPES, EGLNativeWindowType is defined as an
opaque pointer, i.e. void*.
BUG: 440372
If a property blob got created by another drm master we can't rely on
re-using it, so create a new blob of our own with the data copied over
when the kernel destroys it.
Also free property blob objects when we don't need them anymore.
BUG: 440294
At the moment we were checking for size difference between the buffers
which works for rotation but not for flipped or up-side down screens.
This changes how we check it by comparing the effective transform vs the
supported one in drm to see if they differ.
Fixes https://invent.kde.org/plasma/plasma-phone-components/-/issues/112
Do not discard a shortcut immediately when it is already registered
in KGlobalAccel, because there could be possible dangling shortcuts
if they weren't properly cleaned-up (ex. after a crash).
Instead we delay the check for conflicts with other client shortcuts
to the second loop, which we know is up to date.
BUG: 442215
FIXED-IN: 5.22.80
QSGImageNode ignores QSGTexture's filter, one has to specify it using
QSGImageNode::setFiltering().
In hindsight, we need to figure out how to integrate the lanczos filter
with window thumbnails and use it instead.
BUG: 441709
If the buffer size changes after creating or destroying the decoration,
the surface item will discard the pixmap.
If the buffer size remains the same, we can continue using old pixmap.
The current implementation sometimes will leave the screens unusable,
therefore KWIN_DRM_SW_ROTATIONS_ONLY=1 is default.
Use KWIN_DRM_SW_ROTATIONS_ONLY=0 to test hardware DRM operations.
xcb_send_event always copies 32 bytes, so we have to pad all xcb_*_event_t
to 32 bytes to avoid leaking uninitialized stack memory. I found this
problem while running kwin_x11 on a CHERI-RISC-V system (which has bounded
pointers). The xcb_send_event() implementation has a memcpy() that was
copying 32 bytes but the event passed was a bounded to 28 bytes, so this
resulted in a run-time exception in X11Client::sendClientMessage().
The same problem exists in Selection::sendSelectionNotify(), but this time
we could end up copying up to 8 bytes since xcb_selection_notify_event_t
is only 24 bytes.
This disclosure of uninitialized data could in theory have a security
impact if it leaks a pointer value (e.g. a return address) as part of an
exploit chain that needs to bypass ASLR. However, the selection notify
events go directly to the XServer and you most likely already have a
serious problem if an attacker has full control over the XServer. It is
possible that the configure notify events go directly to an untrusted
client, but even if they do this leak is not directly exploitable.
See also https://gitlab.freedesktop.org/xorg/lib/libxcb/-/issues/18
An AbstractDragTarget is introduced. This contains either the DataDevice we are
dragging to or an Xwl bridge.
We set this on Seat along with the active surface.
In future this also allows getting rid of the move filter.
If a virtual desktop has empty name, kwin is going to end up in an
infinite cycle where VirtualDesktopManager::save() updates the
_NET_DESKTOP_NAMES property and then in the next event loop cycle,
RootInfoFilter::filter() is going to call VirtualDesktopManager::save()
because the _NET_DESKTOP_NAMES property has been changed.
This regression has showed up just now because some parts of kwin used
to increase virtual desktop count to create new desktops. Recently, that
code was ported to VirtualDesktopManager::createVirtualDesktop().
The main difference between createVirtualDesktop() and setCount() is
that the latter will set the desktop name to "Desktop N", the former
will use user-specified name.
If the user-specified name is empty, the createVirtualDesktop() method
should assign "Desktop N" name to the virtual desktop according to the
docs. However, it's not the case and this change addresses that.
Not all combinations of connectors, crtcs and planes will work
on all hardware, so we need to test the pipelines before using
them.
BUG: 433107
CCBUG: 435265
In order to accomodate crtc or primary plane changes on outputs we
need to dynamically recreate rendering resources with EglStreams as
the stream is tied to the crtc or plane currently in use.
Instead of using setTransformInternal directly, have DrmOutput
set it in updateTransform, like with the other setters. Also use
updateTransform instead of setTransformInternal in DrmBackend, so
that hardware rotation can be used where possible.
This decouples the management of Shadow from the scene window and allows
multiple items share the same Shadow.
Currently, kwin has a single scene graph, but it makes sense to create a
scene graph per output as they could have different layers, etc. This
would also allow QtQuick share more textures with kwin, which is worth
doing for optimization purposes in the future.
This makes shadow texture coordinates consistent with surface and
decoration texture coords and allows reusing window quads between
kwin items and qtquick items more easily.
This allows the decoration buttons to be updated accordingly
when the status changes.
We also need to explicitely update allowed actions when rules
are applied, as `isCloseable` is not a settable property, but
an allowed action
BUG: 424370
FIXED-IN: 5.22.90
The check for this rule was missing on window initialization,
so on Wayland it wasn't being applied until something would
trigger the check again.
BUG: 429171
FIXED-IN: 5.22.90
Do not access fields in DataSourceModel::m_data that are not available
0 0x00007f3445a4bd22 in raise () at /usr/lib/libc.so.6
1 0x00007f3445a3590e in abort () at /usr/lib/libc.so.6
2 0x00007f3446362bb1 in qt_message_fatal (message=<synthetic pointer>..., context=...) at /home/apol/devel/frameworks/qt5/qtbase/src/corelib/global/qlogging.cpp:1914
3 QMessageLogger::fatal(char const*, ...) const (this=this@entry=0x7fff288a4ff8, msg=msg@entry=0x7f3446663028 "ASSERT failure in %s: \"%s\", file %s, line %d") at /home/apol/devel/frameworks/qt5/qtbase/src/corelib/global/qlogging.cpp:893
4 0x00007f344636203e in qt_assert_x(char const*, char const*, char const*, int) (where=<optimized out>, what=<optimized out>, file=<optimized out>, line=<optimized out>) at /home/apol/devel/frameworks/qt5/qtbase/src/corelib/global/qglobal.cpp:3366
5 0x00007f344973e118 in QVector<QByteArray>::at(int) const (this=0x55f42c20d338, i=7) at /home/apol/devel/kde5/include/QtCore/qvector.h:449
6 0x00007f34497305df in KWin::DataSourceModel::data(QModelIndex const&, int) const (this=0x55f42c20d320, index=..., role=6) at /home/apol/devel/frameworks/kwin/src/debug_console.cpp:1657
7 0x00007f3447674ed1 in QModelIndex::data(int) const (arole=6, this=0x7fff288a55c0) at ../../include/QtCore/../../../../../devel/frameworks/qt5/qtbase/src/corelib/itemmodels/qabstractitemmodel.h:460
no_use_linear: dropping the gbm GBM_BO_USE_LINEAR flag on allocation
explicit_modifiers: moving from the modifier-less api to the modifier
aware one with explicit modifiers
Removing (WIP) will result in more work for translators. There are other
ways to indicate the effect is not ready for its prime time yet, e.g.
being disabled by default.
Can be useful when investigating clipboard related problems, by checking
which source KWin thinks is the current clipboard. Displays clipboard and
primary selection.
Port the RuleSettings and KCM to store and use a list of virtual desktop
UUIDs, instead of the previous x11 positional id, continuing the work on
This allows to set a rule with several desktops on Wayland.
On X11 it has no visible change for the user, but internally it uses the
more modern concept, helping to simplify the related code.
The relevant key on kwinrulesrc changes from `desktop` to `desktops`.
A kconf_update script handles the migration.
The present windows allows the user to search windows by their caption
or window role. This change brings that functionality to this effect.
BUG: 441302
According to the spec, enter has to be send before position.
Sending position first seems to confuse clients, before this
change X windows only seemed to get data when the pointer left and
entered them for a second time. Now it works straight away.
BUG:437406
This lays down some groundwork for realtime gestures in Wayland,
so that gestures that are 1:1 with user motion on a touchpad are
now possible to implement.
Due to earlier commits, this is mostly just glue code to make a
convenient API.
Gestures implemented with this API are four-finger gestures, to
avoid conflicting with apps that may use two or three-finger
gestures.
The previous implementation added noise in linear space, which resulted in
the effect becoming more pronounced on black backgrounds. This patch
changes the process to be applied in perceptual space, by making the noise
addition pass a separate draw call and disabling GL_FRAMEBUFFER_SRGB during
that.
After this change, noise will look much more suppressed and almost never
grainy. This change also changes the range of the noise from
[-strength..strength) to [0..strength), as blending can only be either
additive or subtractive. As a result, users might need to ramp up their
noise parameter after this change.
v2: Add more explanation around the draw call.
v3: Fix noise not fading out with the fade out effect.
v4: Restore an accidentally removed comment.
v5: Add CCBUG.
v6: Rebase.
v7: Fix a formatting issue.
CCBUG: 409620
Originally, there was ShellClient class that was the predecessor to
WaylandClient subclasses. If the main surface is unmapped, the
corresponding instance of ShellClient would be kept around. That had
been changed with the introduction of XdgToplevelClient and
XdgPopupClient.
Since ShellClient had been kept alive even if the corresponding surface
were unmapped, it did make sense to setup window management integration
when the surface is mapped.
However, this also means that plasma will not know about initially
minimized windows.
Since unmapped windows are handled differently now, we can setup window
management after the client performs the initial commit. If the main
surface is unmapped, the XdgToplevelClient object will be destroyed and
with it, the window management integration.
