A motivating example of such problem - Zoom's popups that open on button presses.
Before this fix the popup would flicker and immediately disappear - because the PID is not yet
available for the verification (as the surface has not been associated yet), wlroots would refuse to
focus the popup and instead focus the previous window. This leads QT to interpret this as a sign to
close the popup.
This change moves the PID aqcuisition to an earlier phase - just where the window is created.
Instead, move the wlr_xwayland_surface_set_withdrawn() and
wlr_xwayland_surface_restack() calls to the MapNotify handler with an
override_redirect check, as they are done too early. This mirrors the logic in
the UnmapNotify handler and fixes a bug where wlr_xwayland_surface_restack()
would be called on an o-r window after the following sequence of requests:
- CreateWindow with override_redirect=True
- ChangeWindowAttributes with override_redirect=False
- MapWindow
Fixes: https://gitlab.freedesktop.org/wlroots/wlroots/-/issues/3770
This function allows compositors to set the _NET_WORKAREA property on
the root window. XWayland clients use _NET_WORKAREA to determine how
much of the screen is not covered by panels/docks. The property is used
for example by Qt to determine areas of the screen that popup menus
should not overlap (see QScreen::availableVirtualGeometry).
Translating the right/bottom coordinates from offsets to absolute
coordinates in wlroots (rather than in the compositor) was supposed to
be more reliable, since wlroots had access to the X11 screen size.
It ended up being less reliable, because the screen size values
(xwm->screen->width_in_pixels/height_in_pixels) are not updated when the
output layout changes.
So let's remove the translation from wlroots, and let the compositor
figure it out. From what I can understand of the current XWayland code,
the X11 screen size should generally match the overall wlr_output_layout
bounding box, which the compositor has access to.
Instead of sending one request, waiting for the reply, and
repeating for all properties we're interested in, we can send all
property requests in one go and then wait for the server to reply.
This increases type safety, makes it more obvious that role_data
must represent the role object, and will allow for automatic
cleanup when the resource is destroyed.
This fixes a use-after-free in the Sway patch to filter the Xwayland
shell [1].
The server is destroyed first, then the shell. The Xwayland process
might still be using the shell while running.
When the shell is destroyed, libwayland will invoke the global
filter (to figure out whether to send a wl_registry.global_remove
to clients). Then Sway will compare the client with
wlr_xwayland_server.client. However, at that point, the server is
gone.
Reset the server to NULL so that Sway can check whether the server
is still running.
[1]: https://github.com/swaywm/sway/pull/7647
We'll soon introduce a unified wlr_surface map event. Up until now, compositors
have been using wlr_xwayland_surface's map event to setup various wlr_surface
related listeners (e.g. commit). This will no longer be possible when that
event is moved over to wlr_surface. Introduce new events where the compositor
can add/remove wlr_surface event listeners.
X11 clients expect a ConfigureNotify after a ConfigureRequest. If
the compositor/window manager chooses not to honor the request
(e.g. due to the window being maximized), XWayland will not send a
"real" ConfigureNotify event and the window manager is expected to
send a synthetic event instead. Otherwise, the X11 client is left
waiting and may not repaint its window properly.
For comparison, see Openbox's client_configure() or Weston's
weston_wm_window_send_configure_notify().
v2: Move logic to wlr_xwayland_surface_configure()
This is needed for compositors that want to reserve space for
XWayland panels. Such a feature can be useful in a "transitional"
setup, where only the X11 window manager and compositor is replaced
but other components of an X11 desktop environment are still used.
This change simply reads the X11 property; the compositor is free
to ignore it. Thus, compositors that don't want to support such a
"transitional" feature are not impacted.
v2: Update xwayland_surface_associate()
If a window is unmapped too quickly, we might receive UnmapNotify before
we get the corresponding wl_surface, which will later lead to
associating the same window twice. To fix this, move the NULL surface
check to xwayland_surface_dissociate(), which makes resetting the
unpaired link and the wl_surface object ID unconditional.
Fixes: https://gitlab.freedesktop.org/wlroots/wlroots/-/issues/3552