Running gnome-shell under valgrind, I saw the attached invalid write.
Basically we can destroy a window during event processing, and the old
window_remove_filters simply called g_free() on the filter, ignoring
the refcount. Then later in event processing we call filter->refcount--,
which is writing to free()d memory.
Fix this by centralizing list mutation and refcount handling inside
a new shared _gdk_window_filter_unref() function, and using that
everywhere.
==13876== Invalid write of size 4
==13876== at 0x446B181: gdk_event_apply_filters (gdkeventsource.c:86)
==13876== by 0x446B411: _gdk_events_queue (gdkeventsource.c:188)
==13876== by 0x44437EF: gdk_display_get_event (gdkdisplay.c:410)
==13876== by 0x446B009: gdk_event_source_dispatch (gdkeventsource.c:317)
==13876== by 0x4AB7159: g_main_context_dispatch (gmain.c:2436)
==13876== by 0x4AB7957: g_main_context_iterate.clone.5 (gmain.c:3087)
==13876== by 0x4AB806A: g_main_loop_run (gmain.c:3295)
==13876== by 0x8084D6B: main (main.c:722)
==13876== Address 0x1658bcac is 12 bytes inside a block of size 16 free'd
==13876== at 0x4005EAD: free (vg_replace_malloc.c:366)
==13876== by 0x4ABE515: g_free (gmem.c:263)
==13876== by 0x444BCC9: window_remove_filters (gdkwindow.c:1873)
==13876== by 0x4454BA3: _gdk_window_destroy_hierarchy (gdkwindow.c:2043)
==13876== by 0x447BF6E: gdk_window_destroy_notify (gdkwindow-x11.c:1115)
==13876== by 0x43588E2: _gtk_socket_windowing_filter_func (gtksocket-x11.c:518)
==13876== by 0x446B170: gdk_event_apply_filters (gdkeventsource.c:79)
==13876== by 0x446B411: _gdk_events_queue (gdkeventsource.c:188)
==13876== by 0x44437EF: gdk_display_get_event (gdkdisplay.c:410)
==13876== by 0x446B009: gdk_event_source_dispatch (gdkeventsource.c:317)
==13876== by 0x4AB7159: g_main_context_dispatch (gmain.c:2436)
==13876== by 0x4AB7957: g_main_context_iterate.clone.5 (gmain.c:3087)
https://bugzilla.gnome.org/show_bug.cgi?id=637464
The old functions to get core pointer and devices list are gone as
well. This slice is entirely replaced internally by multidevice
handling and may just go.
This function will enable events for all devices of a given
GdkInputSource, either these available at the time of the call,
or these that are connected in the future.
This function may be used to know the hardware device that triggered
an event, it could resort to the master device in the few cases there's
not a direct hardware device to relate to the event (i.e.: crossing events
due to grabs)
gdk_enable_multidevice() has been replaced with gdk_disable_multidevice(),
so applications may call that function if they want to go back at the
previous behavior.
There would be usually little reasons to call that function, unless the
application is doing X calls itself that count on old fashioned core
devices.
One less magic function. Also refactored it to make it easier to
implement. It now returns TRUE if it beeped and FALSE if it failed to do
so. A default implementation exists that just returns FALSE for all the
backends that can't beep windows (read: everything but X11 with XKB -
and why on earth do keyboard libs implement beeping?)
Trying to get rid of all the _gdk_windowing_something() functions that
we expect backends to magically know about and instead put them in a
proper interface (mostly GdkWindowImplClass).
... instead of _gdk_drawable_ref_cairo_surface() where appropriate.
Also, don't implement the drawable->create_cairo_surface vfunc anymore.
This is in preparation for the split of GdkWindow from GdkDrawable.
You are not allowed to track surfaces from GDK or draw outside of expose
events. So we can remove ugly hacks needed previously. See
https://bugzilla.gnome.org/show_bug.cgi?id=606009 for the introduction
of this workaround.
An event filter may add or remove filters itself. This patch does
two things to address this case. The first is to take a temporary
reference to the filter while it is being used. The second is
to wait until after the filter function is run before determining
the next node in the list to process. This guards against
changes to the next node. It also does not run functions
that have been marked as removed. Though I'm not sure if this
case can arise.
https://bugzilla.gnome.org/show_bug.cgi?id=635380
Add signal GdkWindow::create-surface which allows to use any
surface type as storage for offscreen windows.
Test the new signal in tests/gdkoffscreenbox.c
Now that we don't create pixmaps anymore, this function is not needed
anymore. The indirection it did previously is now basically moved to
gdk_window_create_similar_surface()
Now the window background is a cairo_pattern_t. The backends will try to
set this as good as they can on the windowing system, but no guarantees
are made on wether the windowing system supports the pattern.
Also gets rid of GDK_NO_BG as undefined behavior is not a good idea to
support, and GDK_NO_BG effectively made the window's contents undefined.
It wasn't effectively used in GTK anyway.
The function converts the given surface into an alpha bitmap mask. This
is mostly useful for setting shape regions.
Also adds a new internal function _gdk_cairo_surface_extents() that
computes a surface's extents.
* add per-display gdk_x11_display_error_trap_push()
(X11-specific because gdk_error_trap_push() probably
should have been)
* make gdk_error_trap_push() handle only GDK displays
not displays opened without a GDK wrapper
* make gdk_error_trap_pop() and gdk_x11_display_error_trap_pop()
automatically sync only if needed, so manual gdk_flush() is not
required
* add gdk_error_trap_pop_ignored() which just asynchronously
ignores errors, so never needs to sync
* add G_GNUC_WARN_UNUSED_RESULT to plain pop(), because
if you use plain pop() and don't need the return value,
the async gdk_error_trap_pop_ignored() should be used
instead. This results in lots of warnings to clean
up in a later patch.
The main objective here was to avoid the need to sync just
to ignore an error. Now, syncing is automatic, and only
happens when we need to know the error code.
https://bugzilla.gnome.org/show_bug.cgi?id=629608
While X11 surfaces can be resized, this is not the case for Quartz
surfaces. Instead of resizing we will invalidate the surface instead.
By giving _gdk_windowing_set_cairo_surface_size() a boolean return
value, we can signal back whether or not resizing was possible. If not
possible, we invalidate the surface.
When a cairo surface is requested for direct window access (i.e. not
when double-buffering) we can't really track when the actual drawing happens
as cairo drawing is not virtualized. This means we can't properly flush
any outstanding window moves or implicit paints.
This actually causes problems with e.g. abiword (bug #606009) where they
draw without double-buffering. If you press down it scrolls the window
and then draws the caret, but the caret drawing does not flush the
outstanding move from the scroll, so the caret gets drawn on the wrong
screen.
We fix this by never allowing either implicit paints or outstanding window
moves on impl-windows where any windows related to it has an outstanding
direct cairo surface. Luckily this is not very common so in practice this
doesn't matter much.
This has two advantages:
1) In many backends, this is faster as we can terminate the window
hierarchy traversal earlier
2) When used in gdkdisplay.c::get_current_toplevel() to get the
current toplevel that has the pointer we now correctly return
a toplevel with the pointer in it where the pointer is inside
some foreign subwindow of a toplevel window.
The second advantage fixes some bugs in client side event generation
when the pointer is inside such a foreign child window.
For toplevels, never apply clip as shape, instead apply shape.
This way we don't have to re-set it all the time as the window size
changes. Furthermore, this change fixes unsetting a shape on a
toplevel window which didn't actually unset the shape before.
Additionally we never apply clips as shape if the shape would just
be the same as the regular window size. This means we won't unnecessarily
add a useless shape to most native child windows (and additionally this
helps apps that do weird X stuff that don't expect these shaped windows).
Some applications make weird assumtions on Gtk+ that do not work anymore
with the new client-side windows support. For instance SWT/Eclipse reorders
the stacking order of the X windows directly without telling gdk this,
which breaks gdk drawing as gdk now relies on knowing the stacking order
for window clipping.
This introduces a GDK_NATIVE_WINDOWS environment variable, which if set
causes Gtk+ to always use native windows. Its more compatible with
pre-csw Gtk+ behaviour if you do weird X-specific hacks, although it does
limit the size of GdkWindows to 65535x65535.
If there are outstanding moves in an area that intersects
the source of an outstanding move we need to move the invalid
area correspondingly, otherwise we will expose the wrong area
as the outstanding move copy will happen before we expose
the invalid area.
This moves the native show/hide calls to the generic code
for calculating viewable rather than in its own separate code
called from gdk_window_show/hide. This simplifies the code a bit,
but most significantly it means things are correctly shown when
they become viewable for other reasons than a show/hide call.
For instance, this fixes bug 590442 (gvim embedding) where the
toplevel GtkPlug is mapped by the embedder and we didn't previously
pick up that the native children became viewable and should be shown.
If we get a nonlinear enter/leave notify on the toplevel we need
to set nonlinear in all the events we send, even if the in-toplevel
tree is linear.
This fixes combobox menus popping down immediately when you click
(not hold). (bug #587559)
we now use gdk_offscreen_window_set_embedder() instead of a signal
to get the parent. This also replaces set_has_offscreen_changes.
Rename "parent" in all embedding related names to "embedder" to make it
more obviously different than the normal parent.
Rename gdk_window_get_offscreen_pixmap to gdk_offscreen_window_get_pixmap
to match the other offscreen calls.
Rename gdk_window_offscreen_children_changed to gdk_window_geometry_changed
as this is more descriptive.
Doing this directly had some issues with picking going recursive in
clutter-gtk. Furthermore, doing it in an idle means we can coalesce
multiple calls (which is common due to widget changes) in the same
toplevel to just one call.
3 signals are for offscreen windows
get-offscreen-parent: Get the parent window an offscreen is embedded in
to-parent: Convert coordinates from offscreen to parent
from-parent: Convert coordinates from parent to offscreen
1 signal is for the window embedding offscreens:
pick-offscreen-child: This picks what (if any) offscreen is at a specific position
The last signal is only used if you call gdk_window_set_has_offscreen_children
to tell gdk that the window has embedded offscreen children.
Add get-pointer signal for offscreen window pointer getting
Apps using offscreen windows can connect to get-pointer on offscreen
windows in order to make gdk_window_get_pointer() return correct
values.
Add get-offscreen-parent signal
Add signals for from-parent and to-parent coordinate mapping
Add pick-offscreen-child signal
We returned the innermost child that has the pointer, which is not right.
Only the direct child that has the pointer in it should be reported (if any).
Unfortunately the old GdkWindowObject is public and accessed
from macros, etc. So, we publish a limited copy of GdkWindowObject
and use the full one internally when building gdk.
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.
