It turns out we really have to ignore grab/ungrab events or we'll
report double crossing events when we grab or ungrab.
However, we also can't ignore crossing events from grabs from other clients
as that leads to missed enter/leave events on e.g. alt-tab in metacity.
Fortunately we now track grabs very precisely, so we know with certainty
whether we have a grab at the time (serial) of the native crossing events,
and only if we do we ignore them.
If we get crossing events with subwindow unexpectedly being NULL
that means there is a native subwindow that gdk doesn't know about.
We track these and forward them, with the correct virtual window
events inbetween.
This is important to get right, as metacity uses gdk for the frame
windows, but gdk doesn't know about the client windows reparented
into the frame.
For instance if we grab the pointer and then check if its grabbed
so that we know to ungrab we don't care that the grab is not
yet active, so report the steady state (i.e. the last grab)
Having GDK_WINDOW_CHILD windows with root as the parent apparently works,
and metacity uses it. The current gdk_window_get_toplevel() returns the
root window for that, which is wrong, so we check that explicitly.
This causes all sorts of weirdness with pointer_over_window
being the rootwindow and then crashing gdk_window_get_toplevel() later.
With this metacity stops crashing madly.
To avoid drawing the window background of other windows in the area
where the window was we set the bg to none recursively. However, this
is quite costly it the moved window has many siblings. Furthermore, it
is uncommon that siblings overlap, so this cost has little gain.
So, we only set bg None on the parent, which means that there will
be some more flicker in the uncommon case of overlapping siblings.
These are sent when someone else grabs the pointer, and we don't
want to miss these expose events. For instance, we missed enter
and leave events on alt-tab.
There were some issues with these wrt out-of-sync grab information
in the client, but that should now be handled. So, it should work
or at least be fixable if we find some bug.
After a successful grab/ungrab we wait for an xserver
roundtrip until we change the tracked grab in GdkDisplay.
This way that data is always up-to-date wrt events comming in.
Get rid of invalidate_maybe_recurse and process_updates. Implement
_gdk_windowing_{before,after}_process_all_updates(), and keep track of
when we're inside process_all_updates in the common code so we know
when to flush windows. Implement
_gdk_windowing_window_process_updates_recurse by means of
setNeedsDisplayInRect: displayIfNeeded. Use the added window argument
in begin_paint_region to get the right window (the paintable is always
the impl window now).
Replace them with two new functions
_gdk_windowing_{before,after}_process_all_updates() that are called
around the guts of gdk_window_process_all_updates(). Add empty ones
for X11 (nothing more needed), quartz ones will be implemented next.
We use this in the added windowing function
_gdk_windowing_window_process_updates_recurse. The X11 implementation
just calls _gdk_window_process_updates_recurse directly, but at least
quartz will need to do some more work.
This fixes a problem where we used to set them on a native window, but we
never unset it becase when the pointer moved to another native window
we just set the cursor on that window. Always setting on the toplevel
fixes this.
This reverts commit 7cc15ec6ea1504133dfe6febbdb12615550bb966.
Its risky to convert all the events in a go like this, as it
increases the out-of-order issues. It also isn't a full solution
to the motion hint issue as it will only work for the events
we happen to convert. It would be better to use serials to
handle motion hinting.
This is the first stage in tracking keyboard grabs in the common code.
This lets us handle destroying or unmapping virtual window with a
keyboard grab.
If we only convert the first then motion hint emulation won't
work since we don't see the next motion even until we've
fully handled this one.
However, this changes a behaviour that has been like this since
the mists of time. I don't know if it could cause other issues.
I haven't seen any yet though.
There was a performance problem with the old flicker fixing
approach. For moved windows we copied the window data to the double
buffer pixmap and then back to the window with the rest of the
expose data. In some cases the copy from window data to pixmap was
very slow because the pixmap was allocated in system memory and
the window in video memory.
The new approach is to delay all window moves and then replay them
after the expose has drawn to the double buffer pixmap but before
drawing it to the window. Furthermore, we remove all exposed areas
from the destination of the delayed moves so we won't copy something
just to then immediately draw over it.
This makes scrolling in firefox fast, and it makes tests/flicker not
show any (detectable) flicker.
We return the raw window drawable, so its likely the app will do some
weird stuff to it, like draw using non-gdk operations. We don't want
the app to see any half-drawn state, so flush everything.
This fixes a scroll issue in firefox at least.
This fixes a mismatch in save/release gstate for the CGContext. If
this function is ever used for pixmaps, we need to add support for
that, probably will have to do so soon.
The one that puts windows outside the screen and moves them in when
showing. It might not be needed anymore and the workaround doesn't
work with the client-side window branch anyway because the window is
already mapped when we get showed for some reason.
