Creative people managed to create an X11 display and a Wayland display
at once, thereby getting EGL and GLX involved in a fight to the death
over the ownership of the glFoo() symbolspace.
A way to force such a fight with available tools here is (on Wayland)
running something like:
GTK_INSPECTOR_DISPLAY=:1 GTK_DEBUG=interactive gtk4-demo
Related: xdg-desktop-portal-gnome#5
Goals:
1. Provide as much information as possible in the error message, so
users can try to fix their system themselves.
2. Try to formulate the error message in a way that explains that this
is not something GTK can fix, but a lower layer problem.
Related: #4193
Usually the "dnd-finished" signal will be used to unref the GdkDrag. In
those cases, we would lose the object, so that when we do the final
drag_drop_done() afterwards, we wouldn't have a remaining reference.
With the reference guard, this now works.
nvidia sets the default draw buffer to GL_NONE if EGL contexts are
initially bound to EGL_NO_SURFACE which is exactly what we are doing. So
bind them to GL_BACK when drawing, as they should be.
See https://phabricator.services.mozilla.com/D118743 for a discussion
about EGL_NO_CONTEXT and draw buffers.
Now that we have the display's context to hook into, we can use it to
construct other GL contexts and don't need a GdkSurface vfunc anymore.
This has the added benefit that backends can have different GdkGLContext
classes on the display and get new GLContexts generated from them, so
we get multiple GL backend support per GDK backend for free.
I originally wanted to make this a vfunc on GdkGLContextClass, but
it turns out all the abckends would just call g_object_new() anyway.
Instead of
Display::make_gl_context_current()
we now have
GLContext::clear_current()
GLContext::make_current()
This fits better with the backends (we can actually implement
clearCurrent on macOS now) and makes it easier to implement different GL
backends for backends (like EGL/GLX on X11).
We also pass a surfaceless boolean to make_current() so the calling code
can decide if a surface needs to be bound or not, because the backends
were all doing whatever, which was very counterproductive.
The code to create and manage a fake egl surface to bind to is
complex and completely untested because everyone seems to support this
extension.
nvidia and Mesa do support it and according to Mesa devs, adding support
in a new driver is rather simple and Mesa drivers gain that feature
automatically, so all future drivers shoould have it.
... or more exactly: Only use paint contexts with
gdk_cairo_draw_from_gl().
Instead of paint contexts being the only contexts who call swapBuffer(),
any context can be used for this, when it's used with
begin_frame()/end_frame().
This removes 2 features:
1. We no longer need a big sharing hierarchy. All contexts are now
shared with gdk_display_get_gl_context().
2. There is no longer a difference between attached and non-attached
contexts. All contexts work the same way.
The vfunc is called to initialize GL and it returns a "base" context
that GDK then uses as the context all others are shared with. So the GL
context share tree now looks like:
+ context from init_gl
- context1
- context2
...
So this is a flat tree now, the complexity is gone.
The only caveat is that backends now need to create a GL context when
initializing GL so some refactoring was needed.
Two new functions have been added:
* gdk_display_prepare_gl()
This is public API and can be used to ensure that GL has been
initialized or if not, retrieve an error to display (or debug-print).
* gdk_display_get_gl_context()
This is a private function to retrieve the base context from
init_gl(). It replaces gdk_surface_get_shared_data_context().
Create it during init and then reuse it for all contexts.
While doing that, also improve error reporting - that's not used yet but
will in later commits.
If the pointer capability is added, pointer swipe and pinch gestures
will be created. However, if the pointer capability is removed, the
gesture objects won't be destroyed.
If the pointer capability is removed and added several times in a row,
for example due to plugging and unplugging physical mouse, this can lead
to leaking the old gesture objects.
In order to prevent that, this change makes the seat destroy swipe and
pinch gestures when the pointer capability is withdrawn.
We must call gdk_drag_drop_done() when the drag ends,
successfully or not. Without this, we get an unwarranted
emission of ::cancel after a successful drop.
Since only the first call to gdk_drag_drop_done() is taking
effect, it is safe to call as a fallback, after emitting
::dnd-finished. If the application connects to that signal
and calls gdk_drag_drop_done() itself, its call will take
precedence.
This matches what the X11 implementation does.
The releasing of grabs while a button is pressed (e.g. after starting dnd, or
dragging the window, or going to overview with a pressed button, etc...) was
generalized here in https://gitlab.gnome.org/GNOME/gtk/-/merge_requests/1879.
However we shouldn't break all grabs here. In the case of grabbing popups,
compositors will still emit crossing events between client surfaces (e.g.
popping up and selecting a menu item via press-drag-release), breaking all
grabs here means inconsistent client state, that was
https://gitlab.gnome.org/GNOME/gtk/-/issues/2746.
That was fixed in mutter, by essentially making implicit grabs
owner_events=FALSE, however that breaks the mentioned use pattern entirely.
Mutter is changing this behavior back, so GTK should handle these crossing
events.
The grab that we are interested in breaking here is the implicit pointer
one. Popups will be dismissed via other means if the compositor says their
active grab needs breaking. This still leaves dnd/move/resize drags in
one place, while not allowing #2746 to happen with popups.
Remove a boatload of "or %NULL" from nullable parameters
and return values. gi-docgen generates suitable text from
the annotation that we don't need to duplicate.
This adds a few missing nullable annotations too.
Width and height of a GdkMonitor are derived via wl_output which
talks about physical dimensions of a device and compositors usually
implement this as the untransformed values (e.g. weston, wlroots).
Since the GTK client has no way to figure out if a monitor was rotated,
transform the physical dimensions according to the applied wayland
transform to have the physical dimensions match the logical ones.
Mutter flips the physical dimensions itself but doesn't announce the
transform so this shouldn't break anything there.
Rewrite this in a way that doesn't depend on kernel
header defines at the time the wayland scanner was run.
This was causing the build to break on Centos 8, where
a bunch of fourcc formats are missing.
When building for homebrew/linuxbrew on Ubuntu 16.04, memfd_create() is
not available and causes the build to fail.
This adds a proper check for the function.
When we don't get stettings from the portal, the current
fallback is 'awful fonts'. There is no need for that. Instead,
set the fallback values to grayscale antialiasing with slight
hinting.
... until all globals have been received.
The dependency tracking introduced in 4e9be39518 only allows to
specify required globals and processes the closures as soon as
the requirements have been met. There are, however, also optional
dependencies - most notably the primary_selection protocol.
Currently we rely on the fact that compositors like Mutter announce
it before `wl_seat`, even though the order is not specified in
the spec.
Process globals closures only after all globals have been announced,
so optional dependencies can be accommodated.
Closes https://gitlab.gnome.org/GNOME/gtk/-/issues/3791
This lets the NGL backend be selected instead of the Cairo backend on
devices which expose both GL and GLES, but have better support of GLES.
Tested on a PinePhone.
