This code was designed to make sure we didn't commit new buffers whilst
we were waiting for a configure. The way this worked failed in 6 after
kwin does not reply to desired_size changes immediately.
It is uneeded after kwin commit "wayland: Avoid rearranging layer
surfaces when wl_surface size changes" which means if we do submit
frames between size change requests, they'll be ignored. Meaning the
client will eventually get a configure event at the right size.
When a resize is driven client side we wait for the compositor to have a
chance to reconfigure us before submitting the next frame.
Using a blocking round trip caused an issue. Instead block isExposed and
trigger an expose event whilst a sync is in progress.
Qt's resizing is inherently synchronous. When an application calls
QWindow::setGeometry with a new size, we expect it to make that resize.
An expose event of the requested size will be generated.
Wayland is by default async, a client requests a size, and then will be
configured to that size, or potentially another size.
The simplest way to map the two APIs is to roundtrip when the client
wants to resize. This way we can guarantee that a call to
`setGeometry(); update()` will have the server configured size before
the paint.
In practice it's still not perfect due to other issues, but at least
will sort itself out within a frame.
QWaylandLayerShellIntegration has virtual hooks for Xdg Activation.
This is important to hook up in layer shell because activation using
requestActivate in Qt will go through this path. It also means we have
support for us to drop the implementation in KWindowSystem in favour of
calling into Qt.
Without this when krunner expands, the frame after kwin will resize it
back to the last size the layer shell requested to be resized to.
To detect when we should not send an explicit size on window resize
events we check the anchors; if we're constrained on opposing edges we
want the compositor alone to set the size.
QWaylandLayerSurface pulled data from the Window on startup. The Window
pushed data into the QWaylandLayerSurface on changes. Having two
patterns is a sign of something being off.
This moves everything to a single design, pulling from the public
interface. This allows us to drop a code path that meddles with
QWaylandWindow internals.
When a compositor has no outputs listed Qt creates a dummy placeholder
QScreen object that does not represent a wl_output.
This should still be fixed in the clients to not create a view, it's
wasteful and probably will still have plenty of other bugs, hence the
giant warning, but it's still worth guarding.
CCBUG: 439096
In the current implementation we cannot use a LayerShellQt before the
shell surface is created.
At the moment a shell surface is created, the constructor is run and
then QtWayland commits the current state. This means the compositor
configures the window before a client has any chance to set anchors or
margins.
This works whilst we're just being a simple fullscreen window, but won't
scale for plasmashell in the future.
This patch makes LayerShellQt::Window always creatable, and we can set
and cache properties before the platform window is created, just like
one can on QWindow and XDGShell properties.
This also makes it less potentially crashy as ::get always returns a
valid result, and
sets up the public API to be QML-able as an attached property in future.
Co-authored on Aleix's patch for the unit test
In the current implementation we cannot use a LayerShellQt before the
shell surface is created.
At the moment a shell surface is created, the constructor is run and
then QtWayland commits the current state. This means the compositor
configures the window before a client has any chance to set anchors or
margins.
This works whilst we're just being a simple fullscreen window, but won't
scale for plasmashell in the future.
This patch makes LayerShellQt::Window always creatable, and we can set
and cache properties before the platform window is created, just like
one can on QWindow and XDGShell properties.
This also makes it less potentially crashy as ::get always returns a
valid result, and
sets up the public API to be QML-able as an attached property in future.
Co-authored on Aleix's patch for the unit test
David Edmundson
