Summary:
With nouveau driver it can happen that KWin gets frozen when first trying
to render with OpenGL. This results in a freeze of the complete desktop
as the compositor is non functional.
Our OpenGL breakage detection is only able to detect crashes, but not
freezes. This change improves it by also added a freeze protection.
In the PreInit stage a thread is started with a QTimer of 15 sec. If the
timer fires, qFatal is triggered to terminate KWin. This can only happen
if the creation of the OpenGL compositor takes longer than said 15 sec.
In the PostInit stage the timer gets deleted and the thread stopeed
again.
Thus if a freeze is detected the OpenGL unsafe protection is written into
the config. KWin aborts and gets restarted by DrKonqui. The new KWin
instance will no longer try to activate the freezing OpenGL as the
protection is set.
If KWin doesn't freeze the protection is removed from the config as
we are used to.
Check for freezes for the first n frames, not just the first
This patch changes the freeze detection code to detect freezes in the
first 30 frames (by default, users can change that with the
KWIN_MAX_FRAMES_TESTED environment variable). This detects
successfully the freezes associated to nouveau drivers
in https://bugzilla.suse.com/show_bug.cgi?id=1005323
Reviewers: davidedmundson, #plasma, #kwin, graesslin
Reviewed By: #plasma, #kwin, graesslin
Subscribers: luebking, graesslin, kwin, plasma-devel, davidedmundson
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D3132
Summary:
Rational: unredirect fullscreen windows is a weird beast. It's intended
to make fullscreen windows "faster" by not compositing that screen. But
that doesn't really work as KWin jumps out of that condition pretty
quickly. E.g. whenever a tooltip window is shown. KWin itself has a
better functionality by supporting to block compositing completely.
The complete code was full of hacks around it to try to ensure that
things don't break.
Overall unredirect fullscreen has always been the odd one. We had it
because a compositor needs to have it, but it never got truly integrated.
E.g. effects don't interact with it properly so that some things randomly
work, others don't. Will it trigger the screenedge, probably yes, but
will it show the highlight: properly no.
By removing the functionality we finally acknowledge that this mode is
not maintained and has not been maintained for years and that we do not
intend to support it better in future. Over the years we tried to make
it more and more hidden: it's disabled for Intel GPUs, because it used
to crash KWin. It's marked as an "expert" option, etc.
It's clearly something we tried to hide from the user that it exists.
For Wayland the whole unredirect infrastructure doesn't make sense
either. There is no such thing as "unredirecting". We might make use
of passing buffers directly to the underlying stack, but that will be
done automatically when we know it can be done, not by some magic is
this a window of specific size.
Test Plan:
Compiles, cannot really test as I am an Intel user who never
had that working.
Reviewers: #kwin, #plasma, #vdg
Subscribers: kwin
Tags: #kwin
Differential Revision: https://phabricator.kde.org/D2180
So far if the Scene creation failed kwin_wayland went into a shutdown,
but didn't succeed because the thread to start Xwayland was already
running: it froze.
This change introduces a new signal in Compositor: sceneCreated. The
startup of Xwayland is bound to this signal. If it gets fired KWin can
startup Xwayland. If it does not get fired, KWin terminates correctly.
Each of the backends becomes a plugin. This allows kwin_wayland to load
the requested plugin and kwin itself doesn't need to link all the
libraries needed. E.g. libdrm is no longer linked if running kwin_x11.
Also this allows to create backends for the non-standard EGL platforms
(examples could be raspberrypi or Android devices).
For Xwayland we need to have the Scene (and EglDisplay) created prior
to starting Xwayland and having X11. This requires creating the
Compositor before creating Workspace and starting Xwayland.
To support this the startup of Compositor is split into two parts:
prior and after Workspace creation.
The change might also be interesting for the kwin_x11 case as it could
result in the compositor being up in a quicker way.
The Compositor is destroyed before the Client and Decorations are
destroyed on shutdown. This meant the Decorations reacted needlessly
on the alpha channel supported. E.g. Aurorae recreated the Decoration
and most likely crashed in Qt.
With this change the signal gets disconnected and the Decorations
just don't do anything.
Move the buffer-swap-pending state from the compositing backends to
the Compositor class. The Compositor is the only class that needs to
access the state, and this way it to do it without calling through
a chain of virtual functions. This commit adds two new functions to
Compositor; aboutToSwapBuffers() and bufferSwapComplete(). The
backends call these functions to set and reset the buffer-swap-pending
state.
This commit also renames a number of functions and variables to make
their meaning clear.
The act of promoting the contents of the back buffer to become the
contents of the front buffer is referred to as posting the buffer,
presenting the buffer, or swapping the buffers; rendering the buffer
is what paintScreen() does.
Similar to the already existing DBusInterface wrapper for the
org.kde.KWin interface a new CompositorDBusInterface is introduced for
org.kde.kwin.Compositing.
That way the DBus interface is split from the implementation and DBus
specific methods are no longer required in the Compositor class.
The deprecated DBus methods
* toggleCompositing(bool)
* setCompositing(bool)
are removed.
REVIEW: 118463
The egl wayland backend registers for the callback for a rendered frame.
This allows to throttle KWin's compositor so that we don't render frames
which wouldn't end up on the screen.
For this the Scene provides a method to query whether the last frame got
rendered. By default this returns true in all backends. The Egl Wayland
backend returns true or false depending on whether the callback for the
last frame was recieved.
In case the last frame has not been renderd when performCompositing is
tried to be called, the method returns just like in the case when the
overlay window is not visible. Once the frame callback has been recieved
performCompositing is invoked again.
The CM selecton withdraw delay was introduced to mostly unburden
plasma from recreating the theme on temporary changes, but since
plasma also watches some properties hinting supported features
and acts in consequence, this isn't sufficient and actually causes
two theme changes instead of one in the case of a regular suspend
REVIEW: 110232
CCBUG: 179042
stealing it from ourself is reported to be able
to confuse KWindowSystem about the state.
Also it causes minor but unnecessary overhead.
REVIEW: 110231
CCBUG: 179042
The define KWIN_SINGLETON adds to a class definition:
public:
static Foo *create(QObject *parent = 0);
static Foo *self() { return s_self; }
protected:
explicit Foo(QObject *parent = 0);
private:
static Foo *s_self;
There is an additional define KWIN_SINGLETON_VARIABLE to set a different
name than s_self.
The define KWIN_SINGLETON_FACTORY can be used to generate the create
method. It expands to:
Foo *Foo::s_self = 0;
Foo *Foo::create(QObject *parent)
{
Q_ASSERT(!s_self);
s_self = new Foo(parent);
return s_self;
}
In addition there are defines to again set a different variable name and
to create an object of another inheriting class.
All the classes currently using this pattern are adjusted to use these
new defines. In a few places the name was adjusted. E.g. in Compositor
the factory method was called createCompositor instead of create.
REVIEW: 109865
With Qt5 QCursor does no longer provide ::handle() which was used to
set a cursor on a native XWindow for which we do not have a QWidget.
Also KWin has had for quite some time an optimized version to get the
cursor position without doing XQueryPointer each time ::pos() is called.
These two features are merged into a new class Cursor providing more or
less the same API as QCursor.
In addition the new class provides a facility to perform mouse polling
replacing the implementations in Compositor and ScreenEdges.
For more information about the new class see the documentation for the
new class in cursor.h.
When turning off the compositor do not release the compositor selection
directly but delay it through a timer. The idea is that the internal
change when e.g. restarting the compositor or switching from XRender to
OpenGL should not be visible to the outside world.
This hopefully makes restarting the compositor more robust in Plasma due
to the SelectionWatcher sometimes reporting incorrect results.
When restarting KWin the change does not matter as the selection gets
force claimed by the new instance anyway.
CCBUG: 179042
REVIEW: 106844
The new methods suspend and resume are meant to provide a better way to
influence the current compositing state than toggleCompositing. In
addition an overload setCompositing(bool) is added. The resume method is
implemented in a way that it can be used to try to start the compositor
again in case it failed.
Internally the method suspendResume is dropped as it does the same as
setCompositing just with inverted binary logic and worse name. The
compositingToggled signal is now emitted from within setup and finish to
ensure that especially the compositingToggled(true) signal is only
emitted if the Compositor could start.
Also the updateCompositingBlocking is adjusted to use the new dedicated
suspend and resume methods instead of the toggle method.
REVIEW: 106273
Two new interfaces are introduced:
* org.kde.kwin.Compositing
* org.kde.kwin.Effects
The Compositing interface is generated from scriptable elements on the
KWin::Compositor class and the Compositor is exported as /Compositor.
It provides the general Compositing related D-Bus methods like whether
the compositor is active and toggling and so on.
The Effects interface is generated from scriptable elements on the
KWin::EffectsHandlerImpl class and the instance is exported as /Effects.
It provides all the effects related D-Bus methods like loading an effect
or the list of all effects.
This removes the need to have all these methods provided on the global
org.kde.KWin interface. For backwards compatibility they are kept, but
no longer provided by the Workspace class. Instead a new DBusInterface
is generated which wrapps the calls and delegates it to one of our three
related Singleton objects:
* Workspace
* Compositor
* EffectsHandlerImpl
The Compositor class actually behaves like a Singleton so it should be
one. Therefore four static methods are added:
* self() to access the Singleton
* createCompositor() to be used by Workspace to create the instance
* isCreated() to have a simple check whether the Singleton is already
created
* compositing() as a shortcut to test whether the compositor has been
created and is active
The isCreated() check is actually required as especially Clients might
be created and trying to access the Compositor before it is setup.
It seems not all Qt installs will automatically #include QElapsedTimer
from QtCore/QTimer, this caused a build failure on my system and on a
RHEL 6.2 VM I've been testing on.
Checking the Qt docs, QBasicTimer also has a separate include so
although this didn't cause a build failure, I've thrown in its
separate #include as well. I have not checked for other #include errors,
and a very quick search on b.k.o for bugs mentioning "build" did not
return any bugs to close.
I'm pretty sure this build failure applies only to master but I haven't
checked thoroughly.
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.
The DBus signal which causes KWin to reinitialize the Compositor
is moved into the Compositor as everything can be handled from
there as well. This comes together with moving the restartKWin
functionality into the Compositor as it is only relevant there.
Restart will only happen if the wrong Qt graphicssystem is used
for the chosen compositing backend.
All the custom slot did was printing a debug statement and
calling finish. We do not need this debug statement. The times
of Compositors not part of Window Managers are over, so it is
extremely unlikely that we lose the ownership without KWin
going down anyway.
The Scene has always been created and destroyed inside what is
now the split out compositor. Which means it is actually owned
by the Compositor. The static pointer has never been needed
inside KWin core. Access to the Scene is not required for the
Window Manager. The only real usage is in the EffectsHandlerImpl
and in utils.h to provide a convenient way to figure out whether
compositing is currently active (scene != NULL).
The EffectsHandlerImpl gets also created by the Compositor after
the Scene is created and gets deleted just before the Scene gets
deleted. This allows to inject the Scene into the EffectsHandlerImpl
to resolve the static access in this class.
The convenient way to access the compositing() in utils.h had
to go. To provide the same feature the Compositor provides a
hasScene() access which has the same behavior as the old method.
In order to keep the code changes small in Workspace and Toplevel
a new method compositing() is defined which properly resolves
the state. A disadvantage is that this can no longer be inlined
and consists of several method calls and pointer checks.
Replaces the member variable which is actually not needed as a
pointer to the Workspace can always be retrieved through the
singleton accessor of Workspace.
All Workspace functions which were implemented in the file composite.cpp
were moved to an own class Compositor. The header entries were moved as well.
All functions calls are updated.
