NOTE: this is not working completely yet, lots of code is still ifdefed
other parts are still broken.
The main difference for the new decoration API is that it is neither
QWidget nor QWindow based. It's just a QObject which processes input
events and has a paint method to render the decoration. This means all
the workarounds for the QWidget interception are removed. Also the paint
redirector is removed. Instead each compositor has now its own renderer
which can be optimized for the specific case. E.g. the OpenGL compositor
renders to a scratch image which gets copied into the combined texture,
the XRender compositor copies into the XPixmaps.
Input events are also changed. The events are composed into QMouseEvents
and passed through the decoration, which might accept them. If they are
not accpted we assume that it's a press on the decoration area allowing
us to resize/move the window. Input events are not completely working
yet, e.g. wheel events are not yet processed and double click on deco
is not yet working.
Overall KDecoration2 is way more stateful and KWin core needs more
adjustments for it. E.g. borders are allowed to be disabled at any time.
This is going to be a controversal change. It enforces KWin decorations
on all client side decorated windows from GTK+. Unfortunately we are
caught between a rock and a hard place. Keeping the status quo means
having broken windows and a more or less broken window manager due to
GTK+ including the shadow in the windows. This is no solution.
Enforcing server side decorations visually breaks the windows. This is
also no solution. So why do it?
It's our task to provide the best possible user experience and KWin is
a window manager which has always done great efforts to fix misbehaving
windows. One can think of the focus stealing prevention, the window rules
and lately the scripts. The best possible window management experience is
our aim. This means we cannot leave the users with the broken windows
from GTK.
The issues we noticed were reported to GTK+ about 2 months ago and we are
working on improving the situation. Unfortunately several issues are not
yet addressed and others will only be addressed in the next GTK+ release.
We are working on improving the NETWM spec (see [1]) to ensure that the
client side decorated windows are not in a broken state. This means the
enforcment is a temporary solution and will be re-evaluated with the next
GTK release. I would prefer to not have to do such a change, if some of
the bugs were fixed or GTK+ would not use client-side-decos on wms not
yet supporting those all of this would be a no issue.
For a complete list of the problems caused by GTK's decos see bug [2] and
the linked bug reports from there.
The change is done in a least inversive way in KWin. We just check for
the property _GTK_FRAME_EXTENTS and create a Q_PROPERTY in Client for it.
If we add support for the frame extents in future we would also need
this. So it's not a change just for enforcing the decoration.
The actual enforcing is done through a KWin script so users can still
disable it.
REVIEW: 119062
[1] https://mail.gnome.org/archives/wm-spec-list/2014-June/msg00002.html
[2] https://bugzilla.gnome.org/show_bug.cgi?id=729721
Use the timestamp from the xcb event which triggers the update whenever
possible. If we don't have access to the latest event, let's at least
update our own xTime prior to using it.
Slightly unrelated change included: Group switches the userTime from
XLib datatype to xcb datatype.
BUG: 335637
REVIEW: 118456
Qt doesn't like that we reparent the decoration using low level xcb
calls. So let's use a QWindow wrapper for the frame and let Qt do
the reparenting itself.
BUG: 334768
REVIEW: 118159
Only used for the delayedMoveResizeTimer as timeout slot. Code is small
so a lambda makes more sense. At the same time the code is slightly
improved to ensure that startDelayedMoveResize is never called while
the timer is already active.
This means that mousePressEvents are now required to come from the
decoration.
REVIEW: 117843
The sync protocol with e.g. Qt 4 windows is broken if our app time is
older than the one of the last sync alarm event. Thus we keep a timestamp
in the syncRequest struct of the last sent sync request. If the timestamp
is newer than our xTime when sending the next request, we update the
xTime to ensure that we have a new timestamp again.
BUG: 333512
REVIEW: 117734
So far the Unmanaged got released after an XCB_UNMAP_NOTIFY. This event
gets created after xcb_unmap_window or after xcb_destroy_window. In the
latter case the window is already distroyed and any of KWin's cleanup
calls will cause a BadWindow (or similar) error.
The idea to circumvent these errors is to try to wait for the
DESTROY_NOTIFY event. To do so the processing of the release is slightly
delayed. If KWin gets the destroy notify before the delay times out the
Unamanged gets released immediately but with a Destroy flag. For this a
new enum ReleaseToplevel is introduced and Unmanage::release takes this
as an argument instead of the bool which indicated OnShutdown. Also this
enum is added to Toplevel::finishCompositing so that it can ignore the
destroyed case and not generate an error.
REVIEW: 117422
Adds NET::WM2BlockCompositing to the Client's properties which allows to
read the state from the NETWinInfo object and get updates without having
to resolve the atom ourselve.
REVIEW: 117561
Excluded are the signals to Appmenu as that's currently excluded from
build.
Private slots with only one connection are turned into lambdas.
REVIEW: 117355
Instead of passing the macro based Predicate to findClient it now
expects a function which can be passed to std::find_if.
Existing code like:
xcb_window_t window; // our test window
Client *c = findClient(WindowMatchPredicated(window));
becomes:
Client *c = findClient([window](const Client *c) {
return c->window() == window;
});
The advantage is that it is way more flexible and has the logic what
to check for directly with the code and not hidden in the macro
definition.
In addition there is a simplified overload for the very common case of
matching a window id against one of Client's windows. This overloaded
method takes a Predicate and the window id.
Above example becomes:
Client *c = findClient(Predicate::WindowMatch, w);
Existing code is migrated to use the simplified method taking
MatchPredicate and window id. The very few cases where a more complex
condition is tested the lambda function is used. As these are very
local tests only used in one function it's not worthwhile to add further
overloads to the findClient method in Workspace.
With this change all the Predicate macro definitions are removed from
utils.h as they are now completely unused.
REVIEW: 116916
As can be seen in [1] the patches to KWin were in CVS HEAD before the
protocol got standardized and it never got any adoption. It's neither in
the NETWM spec, nor implemented in Qt4 nor in Qt5. KWin did not even add
the protocol to the NET::Supported property.
Thus it doesn't make much sense to keep a protocol which nobody speaks.
Still the code around the protocol is kept and also the names are kept.
Only difference is that Client::takeActivity got removed and the code
moved to the only calling place in Workspace. Motivated by that change
the enum defined in utils.h is moved into Workspace, it's turned into
a proper QFlags class and used as a type in the method argument instead
of a generic long.
[1] https://mail.gnome.org/archives/wm-spec-list/2004-April/msg00013.html
REVIEW: 116922
Major new functionality is xkbcommon support. InputRedirection holds an
instance to a small wrapper class which has the xkb context, keymap and
state. The keymap is initialied from the file descriptor we get from the
Wayland backend.
InputRedirection uses this to translate the keycodes into keysymbols and
to QString and to track the modifiers as provided by the
Qt::KeybordModifiers flags.
This provides us enough information for internal usage (e.g. pass through
effects if they have "grabbed" the keyboard).
If KWin doesn't filter out the key events, it passes them on to the
currently active Client respectively an unmanaged on top of the stack.
This needs still some improvement (not each unmanaged should get the
event). The Client/Unmnaged still uses xtest extension to send the key
events to the window. So keylogging is still possible.
InputRedirection keeps track of the Toplevel which is currently the one
which should get pointer events. This is determined by checking whether
there is an Unmanaged or a Client at the pointer position. At the moment
this is still slightly incorrect, e.g. pointer grabs are ignored,
unmanaged are not checked whether they are output only and input shapes
are not yet tracked.
The pointer events are delivered to the Toplevel as:
* enter
* leave
* move
* button press
* axis event
Nevertheless move events are still generated in InputRedirection through
xcb test for simplicity. They are still send to the root window, so all
windows get mouse move.
Button press and axis are generated only in the implementations of the
event handlers and delivered directly to the window, so other windows
won't see it.
In KCommonDecoration the OnAllDesktops button gets hidden or shown
depending on the number of desktops. For that KDecoration is extended
by a new property which delegates to the bridge to return whether
onAllDesktops is available. In KWin Core this is implemented using
the number of desktops.
FEATURE: 321611
FIXED-IN: 5.0.0
REVIEW: 116076
This provides a new protocol intended to be used by auto-hiding panels
to make use of the centralized screen edges. To use it a Client can
set an X11 property of type _KDE_NET_WM_SCREEN_EDGE_SHOW to KWin.
As value it takes:
* 0: top edge
* 1: right edge
* 2: bottom edge
* 3: left edge
KWin will hide the Client (hide because unmap or minimize would break
it) and create an Edge. If that Edge gets triggered the Client is shown
again and the property gets deleted. If the Client doesn't border the
specified screen edge the Client gets shown immediately so that we
never end in a situation that we cannot unhide the auto-hidden panel
again. The exact process is described in the documentation of
ScreenEdges. The Client can request to be shown again by deleting the
property.
If KWin gets restarted the state is read from the property and it is
tried to create the edge as described.
As this is a KWin specific extension we need to discuss what it means
for Clients using this feature with other WMs: it does nothing. As
the Client gets hidden by KWin and not by the Client, it just doesn't
get hidden if the WM doesn't provide the feature. In case of an
auto-hiding panel this seems like a good solution given that we don't
want to hide it if we cannot unhide it. Of course there's the option
for the Client to provide that feature itself and if that's wanted we
would need to announce the feature in the _NET_SUPPORTED atom. At the
moment that doesn't sound like being needed as Plasma doesn't want to
provide an own implementation.
The implementation comes with a small test application showing how
the feature is intended to be used.
REVIEW: 115910
Client used to have dedicated methods for different icon sizes instead
of combining all pixmaps into one QIcon. This resulted in various parts
of KWin having different access to the icons:
* effects only got one pixmap of size 32x32
* decorations only got the 16x16 and 32x32 pixmaps combined into a QIcon
* tabbox could request all icon sizes, but only as pixmap
Now all sizes are available in one QIcon allowing to easily access the
best fitting icon in a given UI.
The X property _KDE_NET_WM_COLOR_SCHEME can be set on a window and
specifies the absolute path to a .color file describing the color
scheme of the managed client.
The Client reads this property and creates a QPalette from it. If
the property is not set or the value is incorrect, the Client uses
KWin's default palette.
The idea behind this property is to allow an application with a
custom color scheme to tell KWin which color scheme the window
decoration should use. So that the window looks as a solid pattern
again.
KWin core can access the QWindow of the decoration instead of the
QWidget. This is a preparation step to allow QWidget based window
decorations without any QWidgets at all.
KWin core makes already use of this new accessor to get the window Id
which is also on QWidgets provided through the QWindow.
The frameId only makes sense for a Client, in case of Unmanaged the
same window id is used as for the window() handle. Client creates the
frame and destroys it.
Given that it makes sense to let Client manage the frame properly.
The ::frameId() is therefore virtual and as base implementation it
returns the client id. Client reimplements it and returns the proper
frame id.
Method is also implemented in Deleted as it used to be passed to
deleted.
Instead of inspecting the XEvent queue we create a Timer with a
singleshot of 0 msec to move the setActive(false) call to the end of the
event handling. In case there is a matching FocusIn event this will be
handled before the timer fired and can cancel the timer.
Focus out handling used to check the event queue for a matching focus in
event to prevent short flickers when no window is active. This is not
possible with XCB and needs a replacement. Maybe a short timer event.
Interestingly the attribute send_event from XUnmapEvent does not exist in
xcb_unmap_notify_event_t and also the X protocol doesn't know anything
about send event.
At same time also renaming variable to follow naming scheme and have a
sensible name. Also moved default value initialization into initializer
list.
REVIEW: 110283
The type of the transient_for related variables are changed to
xcb_window_t. They cannot be Xcb::Window as we don't take ownership over
the transient for window.
Variables are renamed to m_camelCase to follow naming scheme.
A wrapper for retrieving the TransientFor hint is added to the Xcb
Wrappers.
Using Xcb::Window to wrap this helper window and port all the used XLib
calls to XCB.
Also renaming the variable to m_ and camel case to follow general naming
scheme.
Unfortunately the Xcb::Window wrapper cannot be used for the client
window as the client should not be destroyed by KWin.
All the API calls except XSelectInput are changed to xcb and the name is
adjusted to m_client to follow the naming scheme.
The Xcb::Window nicely encapsulates the created wrapper window. As
almost all code is adjusted, the variable is also renamed to
m_wrapper to follow the normal naming scheme.
For all the decoration updates called from Client into the decoration we
also have a signal being emitted. So turning the pure virtual public
functions into slots means we can just connect our existing signals and
get rid off the deep function calls.
The keepAbove/Below signals are changed to take a boolean argument as
needed by KDecoration and a few emitted signals are moved to a better
fitting location.
REVIEW: 110335
Main motivation for this change is that it's unhandy to have the class
definition in workspace.h and client.h while the implementation is in
events.cpp although nothing in events.cpp uses it directly.
By getting it out of workspace.h we get the header a little bit smaller
which should improve compile time given that it's included almost
everywhere.
In events.cpp the enum usage is changed to NETWinInfo as that's the class
where they are defined.
RootInfo does no longer hold a workspace pointer. Where it's needed it
uses the singleton accessor of Workspace.
REVIEW: 110199
REVIEW: 103948
BUG: 91703
BUG: 299245
FIXED-IN: 4.11
- The setting is ignored, the decoration always gets a "true" for it
- moving a maximized window requires breaking a "strong" snap (1/16 of screen height - unless you use quick maximization)
- all snapping is done towards the client, not the frame
- QuickTileMode is exported to the decoration (just as the maximizeMode) so that it can fix the bordersize alongside that.
not that i really like using QWidgetAction, but it'll
prevent the popup from autoclosing.
Introduce activityUpdateBlocking to prevent users from
removing the popup under their fingertips
BUG: 283309
FIXED-IN: 4.10.2
REVIEW: 107762
The Client::cursor property is changed from QCursor to Qt::CursorShape
and renamed to m_cursor (as all usages are adjusted).
This property is mostly used to define the cursor on e.g. the extended
deco border window. To make this easier a XDefineCursor replacement is
added to xcbutils.h both as a static method and as a member function to
Xcb::Window.
Using Xcb::Window to manage the xcb_window_t and simplify the code - no
longer need to check whether the input_window is valid before calling e.g.
map, as that's handled in Xcb::Window.
One XLib usage for setting cursor is still present. This will be ported
once all the QCursor::handle() get removed.
REVIEW: 108771
Most windows use the hostname in WM_CLIENT_MACHINE, but there are windows
using the FQDN (for example libreoffice). So instead of "foo" it is
"foo.local.net" or similar. The logic so far has been unable to properly
determine whether windows with FQDN are on the local system.
In order to solve this problem the handling is split out into an own
class which stores the information of hostname and whether it is a local
machine. This is to not query multiple times. To determine whether the
Client is on the local system getaddrinfo is used for the own hostname
and the FQDN provided in WM_CLIENT_MACHINE. If one of the queried
names matches, we know that it is on the local machine. The old logic to
compare the hostname is still used and getaddrinfo is only a fallback in
case hostname does not match.
The problem with getaddrinfo is, that it accesses the network and by that
could block. To circumvent this problem the calls are moved into threads
by using QtConcurrent::run.
Obviously this brings disadvantages. When trying to resolve whether a
Client is on the local machine and a FQDN is used, the information is
initially wrong. The new ClientMachine class emits a signal when the
information that the system is local becomes available, but for some
things this is just too late:
* window rules are already gathered
* Session Management has already taken place
In both cases this is an acceptable loss. For window rules it just needs
a proper matching of the machine in case of localhost (remote hosts are
not affected). And the case of session management is very academic as it
is unlikely that a restoring session contains remote windows.
BUG: 308391
FIXED-IN: 4.11
REVIEW: 108235
This implements an optimization similar to one in compiz which defers updating
the backing X window during a window move until the move is terminated. This
helps alleviate some choppiness when using composite + vsync.
REVIEW: 107256
A decoration can provide the AbilityAnnounceAlphaChannel in addition to
AbilityUsesAlphaChannel. If this ability is provided the decoration can
enable/disable the use of the alpha channel through setAlphaEnabled().
The base idea behind this mechanism is to be able to tell the compositor
that currently alpha is not needed. An example is the maximized state in
which the decoration is fully opaque so that there is no need to use the
translucency code path which would render all windows behind the deco.
In addition also the blur effect honors this setting so that behind a
known opaque decoration no blurring is performed.
Oxygen is adjusted to disable translucency in maximized state and Aurorae
is adjusted to allow themes to enable/disable translucency. For Plastik
translucency and with that also blurring is disabled.
REVIEW: 106810
The only task of the PaintRedirector is to redirect the painting of the
window decorations into Pixmaps. So it should actually do this by also
handling the four pixmaps for the decoration. This simplifies the code
as all the logic concerning redirecting the painting is now grouped
together.
Furthermore the PaintRedirector is now a child of the decoration widget,
which means it gets automatically destroyed whenever the decoration is
destroyed - the Client does not have to care about it.
Also the PaintRedirector gets only created if the Compositor is active as
it is not needed in the non-compositing case.
REVIEW: 106620
The method windowType needs actually two implementations:
* one for Clients
* one for Unmanaged
as for Clients also the window rules are checked and hacks are applied
which is both not needed for Unmanaged windows.
To have the Client specific behavior in windowType the function used to
perform two dynamic_casts which made this method one of the most
expensive during compositing, e.g. for ~1000 frames
* called ~43000 times
* ~85000 dynamic casts
* incl. cost of method: 0.24
* self cost of method: 0.05
* incl. cost of the casts: 0.12
After the change to remove the dynamic casts we have for ~1500 frames
in Client::windowType:
* called ~31000 times
* incl. cost of 0.06
* self cost of 0.02
Calls on Unmanaged and Deleted are so low that we do not need to consider
them.
BUG: 306384
FIXED-IN: 4.10
REVIEW: 106349
For most actions where the compositor needs to perform an action
(e.g. scheduling another repaint) signals were already emitted.
So it's easier to just connect the signals to the Compositor
which in turn makes the code much more readable.
All signals are connected from the Workspace when either the
Compositor gets constructed or a Toplevel gets created.
as the flamewar pointed out, resetShowingDesktop invalidly keeps hidden windows
the reason is that there're several updateVisiblity calls (notably one from the compositor)
which break the showingDesktop state as a side effect (before ::manage() does it's thing)
Since they also invalidate the Withdrawn mapping state, that will fail as isManaged() test
(it's also invalidly used by the compositor to set up the decoration, isManaged() used to be
true because of the updateVisibility() call before) since the result is never Withdrawn
CCBUG: 299655
REVIEW: 105303
Client holds a SharedPointer to the TabBoxClient and only
provides access to a WeakPointer which is passed to TabBox.
ClientModel is adjusted to hold a list of WeakPointers instead
of the direct pointers.
This fixes the following reproducable crash:
1. Configure both primary and secondary TabBox with different
layouts
2. Use primary TabBox
3. Close a window, best the one which used to be active
4. Use secondary TabBox
-> Crash
The reason is that the ClientModel still contains the pointer
to the deleted TabBoxClient in step 3 and while creating the
layout access to the TabBoxClient is needed to get the Client's
icon.
By using the weak pointer it can be ensured that we don't try
to dereference the deleted pointer and prevent the crash.
CCBUG: 290482
CCBUG: 285747
CCBUG: 237345
REVIEW: 105000
Toplevel::setupCompositing returns a boolean value and returns
false in the cases where it has not setup compositing.
This is used by the specialization on Client to not perform the
Client specific setup if Toplevel has not setup.
REVIEW: 104767
TabGroup is not (and should not be) a QObject which
makes it rather useless in scripting. In order to use
window tabs in KWin scripts the methods to interact
with tabs from Client should be used.
REVIEW: 104685
This allows to copy the layer to the deleted window in order to
keep the deleted window in the same layer.
Additionally a new layer is added for unmanaged windows.
Allows to block compositing on Client's from scripts.
Usecase: window rule cannot properly identify buggy
clients and scripting allows to just act on fullscreen
windows.
BUG: 297146
FIXED-IN: 4.9.0
REVIEW: 104448
Client has a property for demands attention and Workspace
is emitting a signal whenever the demands attention state
of any Client changes.
REVIEW: 104204
Property invokes virtual methods returning false by default. Deleted
reimplements the isDeleted and returns true. Client returns true for
isClient. Method is not called isManaged as this is already used
inside Client.
Unlike stated at several places in the code it is not difficult to
setup the connections to all Clients.
It would have been nice if the failed attempts to connect the Clients
would not have made it into the code as emitted signals which are
nowhere used. Not to mention that like in all places the signals to
inform that a state changed were emitted before the state changed was
performed.
For this properties are defined in ClientGroup and several methods
are changed to be slots (to be invokable from scripts). On Client
the clientGroup is exported as a property.
The existing wrapper around ClientGroup is dropped as it is no
longer needed. Interestingly it was wrong anyway as it allowed to
construct a new ClientGroup, which has to be done internally.
At the same time the meta declarations get cleaned up a little bit.
Dropping the wrapper around Client and just exporting the Client's
properties. This breaks all existing scripts as it's now e.g.:
client.caption
instead of
client.caption()
But the first one is the more natural one for JavaScript and also
for everyone writing QML as well.
Setters and signals are mostly still missing in client. Other parts
like ClientGroup must be converted to properties, too.
Martin Gräßlin