Summary:
This change standardizes the behavior regarding the destructor request.
The destructor should destroy the resource and nothing else. The
Wayland library invokes the static unbind method once the resource is
destroyed. The implementation provided by Resource::Private::unbind
triggers a delete later on the Resource. So there is no need to trigger
a deleteLater from the destructor request callback.
This change adds a generic implementation to Resource::Private which is
now used by all inheriting classes replacing the custom implementations.
Test Plan:
For a few Resources the test is extended to ensure that the Resource
gets deleted on server side.
Reviewers: #plasma
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1679
The text input change creates an additional serial, so the check
for last generated serial on the Display fails. The test is now
adjusted to the new semantics. A more reliable way would be to
verify the serial on the SeatInterface, though.
Summary:
This change introduces support for text input. Text input allows to
compose text on the server (e.g. through a virtual keyboard) and sent
the composed text to the client.
There are multiple interfaces for text input. QtWayland 5.6 uses
wl_text_input, QtWayland 5.7 uses zwp_text_input_v2.
wl_text_input is from pre Wayland-Protocols times and considered as
UnstableV0 in this implementation. The other interface is UnstableV2.
Unfortunately the V2 variant is not yet part of Wayland-Protocols, but
used in Qt.
The implementation hides the different interfaces as good as possible.
The general idea is the same, the differences are rather minor.
This means changes to how interfaces are wrapped normally. On client
side in the Registry a manager is factored which represent either of
the two interfaces. Similar on the server side Display's factory method
takes an argument to decide which interface should be factored. This
way a user of the library can expose both interfaces and thus be
compatible with Qt 5.6 and Qt 5.7 onwards.
Reviewers: #plasma
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1631
Summary:
When destroying a SurfaceInterface all callbacks are getting destroyed.
This used to iterate over the callbacks and performing
wl_resource_destroy on them. This triggered the destroy handler which
removes the resource from the callback list. Which means removing from
the list we are iterating on. This could result in a double delete or
accessing invalid memory.
This change copies all callbacks to a temporary list and clears the
normal lists. So the destroy handler does no longer modify the lists
currently being iterated on.
Test Plan: Added a test case which crashed with previous code
Reviewers: #plasma
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1677
The ShellSurface might have the resource destroyed, before being deleted,
so there is a short time frame where resource might be null.
Crash was caught by KWin.
Reviewed-By: sebas and notmart
Summary:
So far the server component performed manual cleanup in some cases
when a client disconnects. But this is not needed: the Wayland library
calls the static unbind methods which do cleanup. If we cleanup ourselves
this can result in double deletes in the worst case, so let's only use
the Wayland functionality.
Adjusted:
* RegionInterface
* SurfaceInterface
* ShellSurfaceInterface (doesn't take a parent anymore)
* DpmsInterface
* QtSurfaceExtensionInterface
* KeyboardInterface
* PointerInterface
* TouchInterface
* DataOfferInterface
* PlasmaShellSurfaceInterface
For each adjusted case a test case is added to verify that the cleanup
works. Exceptions are DpmsInterface as the actual Resource is not exposed
at all in the Server component and DataOfferInterface as that is server
side created.
Reviewers: #plasma
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1640
From time to time the test is failing on build.kde.org with an ASAN
heap-use-after-free error. From my investigation this seems to be caused
by the OutputDevice being constructed on the stack and being destroyed
while handling Wayland events, but before all are handled.
The test mostly operates on the changed signal. There is also a done
signal emitted later on. Wayland sends the done after a set of changes
is transmitted. Thus the test is adjusted to wait for done instead of
changed. So we can ensure that all events are handled before the object
gets destroyed.
I have never been able to reproduce the problem locally, so I cannot
guarantee that the issue is solved for good. If it still happens more
investigation will be needed.
Summary:
When the model gets destroyed the lambda connections were still invoked
and could cause crashers.
Test Plan: Test case added and each one verified that it crashed
Reviewers: #plasma, hein
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1624
Summary:
There is a possibility that a PlasmaWindow is unmapped when the
PlasmaWindowModel gets created. In this situation the unmapped
PlasmaWindow will be deleted in the next event cycle. So far
PlasmaWindowModel didn't handle this situation and the model might
hold deleted objects due to this.
This change addresses this potential problem and ensures the model
gets updated when a PlasmaWindow is deleted.
Test Plan: Test case which exposes the problem is added
Reviewers: #plasma, hein
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1622
Summary:
So far when the active PlasmaWindow got unmapped or destroyed, the
PlasmaWindowManagement didn't update the activeWindow. This means it
could expose a deleted object through it's API which could result in
a crash.
This change addresses the problem by updating the active window when
a window gets unmapped or destroyed.
Test Plan: Tests added which exposed the problem
Reviewers: #plasma, hein
Subscribers: plasma-devel
Tags: #plasma
Differential Revision: https://phabricator.kde.org/D1621
Summary:
The protocol is extended by a dedicated destructor request. When a
PlasmaWindow is umapped we no longer destroy the resource directly,
but only send the unmap. The client is then supposed to clean up
(which it already did in that case) and will invoke the destructor.
The PlasmaWindowInterface object will be automatically deleted after
the unmap once all resources bound for it are destroyed.
The tests are extended by two new test cases which triggered protocol
errors on the client side prior to this change.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1594
Summary:
Exposes closable state in the window model and adds tests. This was
included in the protocol and interface, but missing from the model.
Test Plan: Autotest extended.
Reviewers: graesslin
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1435
Summary:
Analogous to NET::ActionChangeDesktop.
KWindowInfo::actionSupported(NET::ActionChangeDesktop) is hardcoded
to return true in kwin, but that's not how it should be; as this will
be fixed later the Wayland protocol needs this state bit as well for
parity.
Test Plan: PlasmaWindowModel test is extended.
Reviewers: graesslin
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1448
Summary:
Adds states and request methods to determine whether a window can be put into
move or resize modes, and request move and resize modes, respectively.
request* naming pre-discussed with Martin. I chose to add *Mode to be more
explicit as well as avoid namespace conflicts with possible other Move/Resize
methods in the future.
Since these are not toggleable states, there is no requestToggle* methods.
Protocol version is not bumped (also pre-discussed with Martin) since we have
pending changes already bumping to 3.
Depends on D1417.
Test Plan: PlasmaWindowModel test extended.
Reviewers: graesslin
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1432
Summary:
This adds support for 'shadable' and 'shaded' states to the protocol and
to the client and server classes, as well as the window model.
Test Plan: The PlasmaWindowModel test has been extended to test the new states.
Reviewers: graesslin
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1417
Summary:
Surface/SubSurface and Shadow handle the case that an attached buffer
gets destroyed by the client. So far we didn't have this code covered,
but it's rather important as incorrect reference counting can hit
asserts.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1403
Summary:
Basic functionality is covered. Changing of shadow elements not covered,
there seems to be lacking server API for that - no change signal.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1398
Summary:
Now also tests:
* requesting mode from client side
* update supported change on server side
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1357
If the focused pointer and keyboard surface is the same we use pointer
clicks as a hint to which child surface should have keyboard focus.
Keyboard focus handling for sub surfaces is rather limited overall.
We just don't have a good model on how to determine which child surface
should get the keyboard focus. When passing focus to a surface there
is no way to know which of the sub-surfaces should get the focus.
Ideally the client should handle this, but that's just not the case.
The best we have is a reference through the pointer. But that's of
course also limited. Keyboard focus passed to the surface for another
reason (Alt+Tab) cannot select the proper sub-surface without interaction
from another input device.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1330
Summary:
The idea behind this change is to make the existance of sub-surfaces
an implementation detail for pointer events. The user of the library
does not need to care about on which sub-surface the pointer is on.
It only needs to care about the main surface and passes the focus to
the main surface.
Internally the PointerInterface takes care of sending the enter to
the sub-surface at the current pointer position. Also whenever the
pointer position changes, the PointerInterface evaluates whether it
triggered a change for the focused sub-surface and sends enter/leave
events accordingly. If the focused sub-surface does not change, it
sends motion events as normally, but of course under consideration
of the sub-surface position.
Overall this means that from pointer usage perspective a user of the
library doesn't need to care about the fact that there are sub-surfaces
at all. The library does the correct thing for it.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1329
Summary:
The new method returns the (child) surface at a given surface position
taking care of stacking order, whether surfaces are mapped, etc.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1319
Summary:
This change introduces a damage tracking feature in SurfaceInterface.
So far the SurfaceInterface only exposes the damage compared to the
last attached buffer. But this is not always usefull for the user of
the library. E.g. if:
* server renders
* client damages buffer and commits
* client damages buffer and commits
* server wants render
In this situation the last damage information is not helpful to the
user of the library. It would need the combined damage information over
all attached buffers.
The new API combines the damage of the two commits in the example above.
The user of the library can then call resetTrackedDamage once it
processed the current damage (e.g. by updating the OpenGL texture).
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1281
Summary:
QtWayland doesn't map the parent sub-surfaces in a sub-surface tree.
According to the spec this would mean also the child sub-surface is not
mapped. But being strict according to the spec will make applications
like SystemSettings fail badly. Embedded child windows will not be
rendered and QtWayland is going to hard freeze. This is not acceptable,
thus we need to workaround this QtWayland bug till it's fixed.
It's worth mentioning that Weston as the reference compositor also
doesn't handle this situation according to spec and renders the
sub-surface. See https://bugs.freedesktop.org/show_bug.cgi?id=94735
The difficult part for the workaround is to determine whether a surface
should be considered unmapped. E.g. when the parent gets unmapped we need
to really unmap it. But what's the difference between an unmapped parent
surface which should be considered mapped and an unmapped parent surface
which should be considered unmapped?
The implementation goes with considering a new sub-surface always as
mapped - independently of whether it ever got a buffer attached. As soon
as it had a buffer attached and it gets unmapped again, it will go back
to a standard conform way.
The behavior now is not standard conform, thus the autotest is adjusted
to have QEXPECT_FAIL for the now no longer standard conform areas.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1250
Summary:
If a Surface doesn't have a buffer attached and a null buffer gets
attached the buffer state doesn't really change. Thus neither the
unmapped signal nor the damaged signal should not be emitted.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1261
Summary:
New test case which verifies the behavior when a surface is considered
mapped or unmapped in a sub-surface tree.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1248
Summary:
In a SubSurface tree a Surface is only considered mapped if the Surface
has a buffer applied and the parent Surface is mapped.
The added method implements this check. It's useful for the compositor
to easily check this condition as it allows to easily figure out whether
a SubSurface needs to be rendered and it's also useful for implementing
the input handling as a not mapped sub-surface should not get any input
events.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1247
Summary:
From spec:
If the wl_surface associated with the wl_subsurface is destroyed,
the wl_subsurface object becomes inert. Note, that destroying either
object takes effect immediately.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1211
Summary:
Recursively go up to the main surface, that is the top level surface
which doesn't have a parent.
This is useful to know to which surface tree a sub-surface belongs to.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1207
QtWayland doesn't commit the parent surface when creating a sub-surface.
This results in a QtWayland application to freeze as it renders to the
surface and waits for the frame rendered, which it will never get as the
Compositor waits for the commit on the parent prior to mapping the
sub-surface.
To work around this behavior, we apply the adding/removing directly.
The behavior around this is actually not fully documented, so QtWayland
is not wrong per se. See:
https://lists.freedesktop.org/archives/wayland-devel/2016-March/027540.html
Once this is properly clarified and implemented in the Client, we should
revert this change.
Differential Revision: https://phabricator.kde.org/D1191
The idea behind this signal is to notify whenever the tree of sub
surfaces changes in a way that a repaint of the Surface is required.
Possible situations are:
* surface damaged
* surface unmapped
* subsurface added/removed
* subsurface moved (position changed)
* subsurface stacking changed
Ideally it would be possible to provide the actual area which needs
repainting, but due to the possible complexity of the tree, synced
and desynced changes this doesn't look worth the effort. A user of
the signal might trigger too many repaints with it, but if it really
wants to be only notified about the actual changes, it can just track
the individual sub-surfaces.
Summary:
The wrapper for wl_surface::set_buffer_scale was still missing.
Main reason for implementation is the need for the added auto-test.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1188
DataDevice exposes new methods to get the DataOffer provided through
drag'n'drop and also signals to indicate the various changes on the
DataDevice during drag'n'drop.
This allows to add a first basic auto test for the drag'n'drop
functionality covering both server and client side.
Summary:
The Cursor wasn't properly initialized. E.g. the damage signal didn't
get connected resulting in the server not noticing when the cursor
changes. The damage only got connected if a new cursor got instelled by
the client on the same pointer.
This change ensures that the Cursor is properly initialized by calling
into the same method as when the cursor changed.
The tests are extended by a new test case for damaging the surface.
Reviewers: #plasma
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1022
Summary:
The signal gets emitted whenever the focused PointerInterfaces gets
newly set or reset to nullptr. This is needed to better track the
current cursor image in the compositor.
Reviewers: #plasma, sebas
Subscribers: plasma-devel
Projects: #plasma
Differential Revision: https://phabricator.kde.org/D1007
This patch transports the EDID data base64-encoded over the wire.
Apparently, we can't just send random QByteArrays as "strings" over, it
has to be encoded and decoded. So...
* base64-encode the data before sending to the client
* base64-decode it on the client side
* document the above, fix documentation woes in the xml definition
* change test accordingly
The test data used was actually invalid, it's a base64 string of the
actual data, so fix the tests (which actually breaks it), and encode on
the server-side and decode on the client side.
REVIEW:126380
So far we only supported mapping global to surface-local coordinates
using a 2D-offset. With this change it's possible to register a
QMatrix4x4 to describe the transformation for going from global to
surface-local coordinates in a full 3D space.
The existing 2D-offset is transformed to use the new matrix based
variant describing a translation.
REVIEW: 126271
this exposes the geometry of taskbar entries in
plasma-windowmanagement, in order to make the
minimize effects possible.
unlike on X11, it takes relative positions and
it has one geometry per panel, making possible
to have multiple taskbars working.
REVIEW:125871
A transient surface can indicate through the flags that it does not
want to accept keyboard focus. This is now exposed through a dedicated
method.
REVIEW: 125552
So far transient was a mutual exclusive mode causing a transient window
to not be able to be fullscreen. This seems wrong. Let's have transient
still as a dedicated mode allowing the window to be maximized and/or
fullscreen. Only popup stays a dedicated mode.
REVIEW: 125468
On client side a setTransient method is added which wraps the semantic
of wl_shell_surface_set_transient.
On server side both set_transient and set_popup are implemented, though
for popup only the transient part is implemented. In particular the grab
is not yet handled and also no popup done is provided.
For the transient on server side the flags are ignored. Main reason is
that Qt does not use the flag, so testing whether it works is tricky
(needs a test application).
REVIEW: 125223
* Raises wl_seat supported version to 4 in both server and client
* Raises wl_keyboard supported version to 4 in wl_keyboard
* wl_pointer and wl_touch are still on version 3
* Raises minimum Wayland version to 1.6
a protocol to activate the blur behind windows and to
optionally set a sub region of the window where to apply
the blur to, in case the window is shaped
REVIEW:125015
This change addresses the sometimes cumbersome to use API of Registry.
So far to bind an interface one needed to connect to the signals when
setting up the registry. There was no other way to get the name and
version at a later point. This made it difficult to use one Registry
in different parts of an Wayland application.
This change adds a convenient:
struct AnnouncedInterface { quint32 name; quint32 version;};
AnnouncedInterface interface(Interface) const
QVector<AnnouncedInterface> interfaces(Interface) const
The first one is meant for interfaces like wl_shm which is normally only
announced once. The latter one is meant for interfaces like wl_output
for which normally multiple interfaces can get announced.
Those methods can be used without having to connect to the announced
spy.
REVIEW: 124734
Apparently we need to add a small delay after starting weston otherwise
it fails to connect on the CI system.
Additionally we switch to using headless-backend.so which should also
make the setup more robust.
Running real world applications (Xwayland) showed that it doesn't like
at all that we do a dispatch when we are going to flush. This caused
in a very reliable manner a "Connection closed" error in XWayland, taking
down the client and in return the (xwayland-enabled) server.
Anologous to the implementation in Pointer:
* signal entered and left
* Surface *enteredSurface()
The passed in keys are not yet passed along and
currently ignored.
Now the code handles correctly the attaching of a null buffer
(emits a signal unmapped) and if a commit doesn't have a pending
buffer it won't be reset. Damage requests are ignored if no buffer
has been attached yet.
If one passes StartMode::ConnectClientsOnly the socket in
XDG_RUNTIME_DIR is not created and instead one can only connect to
this server using the Display::createClient method. A very nice
solution to just have a Server and one dedicated Client talk to each
other.
The selection is supposed to be sent to the DataDeviceInterface just
before getting keyboard focus. In order to do that the SeatInterface
keeps track of the DataDeviceInterface which is the current selection
and the DataDeviceInterface of the focused keyboard client.
SeatInterface friends DataDeviceManagerInterface so that the latter
can register each created DataDevice for the SeatInterface.
The button state is a seat-global state and not a per pointer state.
All pressed/released and axis events are moved to the SeatInterface
and just invoke the related method on the focused surface pointer.
Makes PointerInterface more like other Interface classes wrapping
wl_resource. The most important change is the handling of the
focused surface. This is now kept in the SeatInterface and can also
be set if there is no PointerInterface for the client yet.
The unit tests had to be adjusted and some are also disabled as the
button events are not yet moved into SeatInterface.
This method is supposed to return the PointerInterface for the current
focused surface. At the moment it just creates the one global
PointerInterface. The existing SeatInterface::pointer method got
removed as that is actually wrong usage.
There can only be one focused surface per Seat, thus the information
should be hold in the seat.
This only adjusts the API, the actual data is still hold in the
PointerInterface. This still needs adjustment.
The Resource base class is supposed to be used by all interface
classes which get created for a wl_resource.
Most interface classes are adjusted, but there are some exceptions:
* BufferInterface: is different as the wl_resource is already created
* PointerInterface and KeyboardInterface: those two need changes, the
implementation differs from all other interface implementations.
The BufferInterface used nested calls of wl_shm_buffer_begin_access
and allowed multiple different BufferInterfaces to be in a block
between wl_shm_buffer_begin_access and wl_shm_buffer_end_access.
But this is not allowed: nesting is only possible if it accesses
the same buffer!
This resulted in an abort when we accessed two BufferInterfaces
at the same time. The added test case aborts in the previous
version.
The fix now changes the semantic of the method. The BufferInterface
doesn't hold the QImage any more, but creates a new one every time
::data is invoked. The QImage is created with a custom cleanup
function which calls to wl_shm_buffer_end_access. It ensures that
we don't call into wl_shm_buffer_begin_access as long as there is
a valid QImage for any other BufferInterface still around and
instead returns a null QImage.
Thus a user of a BufferInterface::data should destroy the returned
QImage as soon as it's no longer needed or create a deep copy if
it needs to be kept around.
Qt::MouseButton is mapped to the linux buttons. The mapping does
not match the mapping used in QtWayland module [1] as that seems
to be incorrect to me. E.g. Qt::BackButton is not mapped to BTN_BACK.
[1] qtwayland/src/client/qwaylandinputdevice.cpp
Only selection part is implemented, drag'n'drop still needs to be
implemented.
Unit test is not properly testing whether the data can be transferred.
This needs some better architecture to have multiple processes which
perform the source and target part.
This implements the subcompositor and subsurface protocol on both
Client and Server side.
Client:
New classes SubCompositor and SubSurface. The SubCompositor can be
created through the Registry and creates the SubSurface which is
bound to a Surface and has a parent Surface. The SubSurface class
provides convenient wrappers for all calls exposed in the
wl_subsurface interface.
Server:
New classes SubCompositorInterface and SubSurfaceInterface. Support
for all commands is added, though the API probably still can need
some fine tuning. The synchronized vs. desynchronized behavior is
not yet exposed in the API. This could also be delegated towards
the user of the library.
Emitted when the Wayland display is done flushing the initial interface
callbacks, announcing wl_display properties. This can be used to compress
events. Note that this signal is emitted only after announcing interfaces,
such as outputs, but not after receiving callbacks of interface properties,
such as the output's geometry, modes, etc..
This signal is emitted from the wl_display_sync callback.
REVIEW:120471
The test started to fail on my system without changing anything. I
assume my system got an update to latest wayland version and it
looks like the registry doesn't announce anything if nothing is
registered. So now our display at least creates the Shm interface.
In addition the test case is slightly improved with a few smaller
changes added during the debug session.
The EventQueue is a wrapper for wl_event_queue. The usage is for the
case that the connection is in a different thread.
The EventQueue needs to be added to the Registry before Registry::create
is invoked. This ensures that the Registry gets added to the EventQueue
and all objects created by the Registry (and further down the tree)
will be added to the same EventQueue.
The auto-tests which already used a wl_event_queue are transited to the
EventQueue class. That's also the reason why there is no dedicated
unit test: the new code is sufficiently tested by the existing tests.
This makes the API more consistent and allows to properly cleanup
Keyboard and Pointer if the connection dies. Like with Shell and
ShellSurface signals are emitted from Seat when the interface is
going to be released or destroyed. These are connected to the
methods of the created Pointer and Keyboard.
The Buffers are exclusively hold by ShmPool. The user of a Buffer is not
supposed to delete it as a no longer used Buffer should be reused by the
ShmPool.
To make it obvious that the ownership of the pointer is not passed to the
user the return type is changed to QWeakPointer. This also allows the
ShmPool to destroy Buffers as needed.
The unit test found a few problems which are now addressed
* getBuffer did not check the format when reusing a buffer
* creatBuffer used the wrong method on QSize to check whether it is empty
* destroy didn't call destroy on the Buffer. This is now added by moving
the Buffer::Private in a dedicated header which can also be included
from the ShmPool
Ensure that we do not connect to a version which our client library
does not support. If we would allow higher versions we would run the
risk that the wayland library calls into not existing callbacks.
