Since all windows are now MetaWindows as well as compositor
windows, there's no reason to keep a compositor-specific hash table
mapping from XID to MutterWindow.
This reduces complexity and removes a call to XQueryTree that could
potentially produce a BadWindow error if not error-trapped.
https://bugzilla.gnome.org/show_bug.cgi?id=613398
meta_compositor_process_event uses meta_error_trap_push/pop for
every event processed by meta_compositor_process_event which isn't needed
and can cause performance problems.
https://bugzilla.gnome.org/show_bug.cgi?id=613398
- Specify a minimum version of Clutter-1.2.0
- Remove conditionalizatin and always use Clutter-1.1 framebuffer
API rather than raw GL fbos
- Replace deprecated cogl_material/texture_unref() with
cogl_handle_unref()
https://bugzilla.gnome.org/show_bug.cgi?id=610862
The change to reduce GConf trips by using
gconf_client_all_entries() broke the fallback to builtin values
because update_binding() was no longer called for bindings not
found in GConf. Fix this by keeping track of the bindings we
find from GConf in a hash table, then looping through and setting
all the bindings at the end.
This also improves efficiency by avoiding a linear scan for each
binding in GConf.
https://bugzilla.gnome.org/show_bug.cgi?id=609710
When we are reading bindings initially, update_list_binding() needs
to be passed the correct "string list type" since we are calling
it with a list of strings instead of a list of GConfValue.
https://bugzilla.gnome.org/show_bug.cgi?id=609101
The older code relied on Clutter providing default texture coords for any
layers for which texture coords were not specified, which does not work as
of Clutter 1.1.6 (due to commit 8b950bdc87).
https://bugzilla.gnome.org/show_bug.cgi?id=609657
Treat the empty string the same as "disabled" for GConf keybinding
keys. gnome-keybinding-properties was changed to write disabled
keys as the empty string a year or so ago.
https://bugzilla.gnome.org/show_bug.cgi?id=559816
Whether Metacity honors a raise request from an application should
not be affected by the raise_on_click setting; remove a check that
seems to have been added in error.
https://bugzilla.gnome.org/show_bug.cgi?id=445447
When we are moving a window with a modal dialog to a different
workspace, meta_workspace_focus_default_window() can be called
with 'not_this_one' being the focused modal dialog.
Since the ancestor of that window is also being moved, we must
not focus it as an alternative to the current window; this will
cause windows to be moved back and Metacity to get into an
inconsistent confused state.
https://bugzilla.redhat.com/show_bug.cgi?id=237158https://bugzilla.gnome.org/show_bug.cgi?id=598995
We currently allow XRaiseWindow when the same application (defined
by the window group) is focused, but the kind of old applications
that XRaiseWindow are frequently not setting the window group.
Expand the check to allow the same X client (defined by the looking
at client ID) to raise windows above the focus window.
https://bugzilla.gnome.org/show_bug.cgi?id=567528
When started without session related command line parameters (e.g. from
gnome-session), metacity picks up client_id from the DESKTOP_AUTOSTART_ID
environment variable. Unfortunately, there is no way to distinguish if this was
passed from a config file, representing old saved session, or generated by
gnome-session, therefore load_state is attempted in each case. If the client_id
is generated, there will be no session file, and metacity will report a
warning.
Just remove the warning so that users won't always find a warning at the
start of their .xsession-errors.
https://bugzilla.gnome.org/show_bug.cgi?id=577576
The atk-bridge GTK+ module opens its own display; if we get an
XIOError on that display, we shouldn't abort with a meta_bug()
but just exit normally. Also fix a segfault if we got an XError
for that display.
https://bugzilla.gnome.org/show_bug.cgi?id=604319
Add a configure switch:
--with-libcanberra=[yes/no/auto]
(defaulting to auto); if libcanberra is not found or explicitly
disabled, then the default system bell will be used for the bell
sound and no switch workspace sound is played.
https://bugzilla.gnome.org/show_bug.cgi?id=609585
These sounds are good candidates for caching in the sound server, to
save a bit of CPU and make reaction faster. Hence, tell libcanberra to
cache them.
https://bugzilla.gnome.org/show_bug.cgi?id=609585
libcanberra generates specific tooltip popup sounds and for that
recognizes the tooltip windows by the GtkWindowTypeHint set for them.
This trivial patch simply sets the hint for the self-drawn tooltips
metacity uses.
https://bugzilla.gnome.org/show_bug.cgi?id=609585
Right now metacity issues only 1 bell event per second. This is
feels buggy when triggering multiple alarm sounds in a terminal.
This patch simple increases the limit to 1/100ms. 100ms is probably a
good choice since the HIG recommends that all user reaction should
happen within 100ms. With this applied pressing 'Left' in gnome-terminal
feels much more responsive.
https://bugzilla.redhat.com/show_bug.cgi?id=498608https://bugzilla.gnome.org/show_bug.cgi?id=609585
* src/core/bell.c: Don't force CA_PROP_CANBERRA_ENABLE to 1.
That was a misunderstanding on my part, and makes it impossible
to get rid of the bell.
svn path=/trunk/; revision=4165
https://bugzilla.gnome.org/show_bug.cgi?id=609585
* src/core/bell.c (meta_bell_set_audible): Now that we are
using libcanberra, don't tell the X server to play the system
bell internally.
svn path=/trunk/; revision=4141
https://bugzilla.gnome.org/show_bug.cgi?id=609585
metacity tries to do the right thing, by preloading all the relevant
directories before getting the keys one-by-one, but GConfClient isn't actually
smart enough to avoid server roundtrips in this case. That should certainly be
fixed in GConf.
In the meantime, here is a patch that reworks the metacity prefs initialization
to avoid roundtrips for individual keys anyway, by using
gconf_client_all_keys().
https://bugzilla.gnome.org/show_bug.cgi?id=574121https://bugzilla.gnome.org/show_bug.cgi?id=607746
The ICE connection is opened by libSM; we can't just close it when
we get an IOError on the ICE connection; instead call SmcCloseConnection()
and mark the connection as closed. This will prevent a segfault if we
exit out of the metacity main loop and get to meta_finalize().
https://bugzilla.gnome.org/show_bug.cgi?id=604867
It makes more sense because the grave key is close
to the tab and escape keys which the other cycle
keybindings use.
This always works better for gnome-shell, which
switchings between applications by default with alt-tab.
The user can now alt-tab to the application they want,
and then move their finger to the grave key to select
the window they want.
Windows demanding attention should never appear in the alt-tab list
unless they're of a type which might have appeared there anyway. This
solves a problem under AWN where docks which were marked as demanding
attention appeared in all alt-tab lists; they were irrelevant and it
was impossible to remove them from the lists.
svn path=/trunk/; revision=4123
Lines where x1==x2 or y1==y2 may have the second element null. Lines
where both are null, and the width is zero, are points. This speeds
things up surprisingly much.
svn path=/trunk/; revision=4119
(This is inspired by Metacity commit 45cbaa2 by Thomas Thurman, but
much simpler - the use of g_date_strftime("%Y") ended up being just
%d for all 90+ current translations)
meta_workspace_set_builtin_struts() is slightly expensive; it involves
discarding all our cached computed information about the layout of the
workspace. So catch calls to set_builtin_struts() that don't change
anything.
https://bugzilla.gnome.org/show_bug.cgi?id=609546
Since meta_workspace_invalidate_work_area() frees the edges
workspace->screen_edges and workspace->monitor_edges, we must clean up
our cached edge resistance data when the invalidate_work_area() is
called on the active workspace, or when the workspace changes.
Make the computation of the edge resistance data lazy so that it
will be recomputed the next time we try to access it.
meta_display_compute_resistance_and_snapping_edges() is made
private to edge-resistance.c
Invaliding the data when active workspace changes also will improve
correctness for edge resistance when the current workspace changes
during a grab operation. (Even with this fix we still don't try to
handle window positions changing during a grab operation; that can't
cause a crash since, unlike screen and monitor edges, the window edges
are freshly allocated, it will just cause slight oddness in that
corner case.)
Root cause tracked down due to much effort by Jon Nettleton.
https://bugzilla.gnome.org/show_bug.cgi?id=608800
When putting 32-bit properties into longs on 64-bit architectures,
XGetWindowProperty assumes the values are supposed to be signed, and
so it sign-extends values greater than 0x7fffffff. So if they *aren't*
supposed to be signed, we need to chop off the high bits ourselves.
(Most CARDINAL-valued properties only end up using small values
anyway, so it doesn't matter, but _NET_WM_WINDOW_OPACITY uses the full
range, and so was previously failing on 64-bit machines.)
https://bugzilla.gnome.org/show_bug.cgi?id=605678
The CGL_* defines in COGL were always meant to be private and should
have never been exposed in the first place. The API in COGL has been
updated to never require them starting from 1.1, but using the original
GL symbols has always been the intent of the API.
This commit removes the CGL_TEXTURE_RECTANGLE_ARB usage in favour of the
ARB-sanctioned GL_TEXTURE_RECTANGLE_ARB enumeration value.
Signed-off-by: Emmanuele Bassi <ebassi@linux.intel.com>
https://bugzilla.gnome.org/show_bug.cgi?id=607398
PFNGLACTIVETEXTUREPROC (a GL-1.2 addition) was inadvertently missing
from some recent versions of Mesa (like that in Fedora 11.) Use
the identical PFNGLACTIVETEXTUREARBPROC instead.
the mutlitexture and texture_rectangle extensions have recently
been incorporated into the GL core; fixes needed to work with
libGL that proceeds that:
GL_TEXTURE_RECTANGLE_ARB - use _ARB name
glActiveTextureARB() - use get_proc_address
https://bugzilla.gnome.org/show_bug.cgi?id=602870
Add MutterTextureTower, an abstraction for getting a image with
the right level of detail for rendering at a particular scale,
by manually scaling down by powers of two.
This results in much better looking scaled window images when
mipmaps can't be used with texture_from_pixmap (which is the
typical case for current GL drivers.)
When framebuffer objects are available, they are used to do
the scaledown using the GPU without having to pull the data
back from video memory. A software codepath is also available
for the case when FBO's are not present, though performance
will suffer
https://bugzilla.gnome.org/show_bug.cgi?id=601032
The commit that removed metacity-dialog added a global SIGCHLD handler
that caused problems by (a) calling waitpid(-1) and thus breaking
g_child_watch for everyone else, and (b) doing too much from a signal
handler and sometimes causing deadlocks (bug 596200).
This removes the global handler and has each zenity user create its
own child watch to watch for exit. (It also fixes the window class of
the zenity dialogs, so that meta_window_present_delete_dialog() will
work again.)
Having a MetaDisplay::window-demands-attention signal allows to deal with
windows demanding attention in a cetralized fashion.
The signal is emitted when a window is created with initial demands-attention
state and/or when the state changes later on.
Based on original patch by Jon Nettleton.
https://bugzilla.gnome.org/show_bug.cgi?id=597052
With the change from bug 582639, we no longer call the reload
functions for properties that are not initially set, so the
initialization of fields in window.c has to match what
window-props.c would set for a missing property.
There was only one discrepancy, window->input, which needs
to be set to TRUE by default (or a window missing a WM_HINTS
property won't get focus); we also add explicit initializers
for a couple of fields that were getting 0-initialized
to the correct default value of FALSE for consistency with
the explicit intialization of the rest of the fields.
Bug reported by Dominique Leuenberger
https://bugzilla.gnome.org/show_bug.cgi?id=601228
For some consumers it's significantly more convenient to be able
to directly connect to a signal on the Window to know when
Mutter is done with it, rather than having to connect to each
Workspace object (and handle workspace additions, etc.).
Similarly, add window-created which acts globally.
https://bugzilla.gnome.org/show_bug.cgi?id=598289
When we focus a window on a different desktop, and the calc_showing
idle that hides/shows the windows gets run before we get focus events
back from X, we think that we are hiding the window with the focus
so we focus a "random" window to avoid leaving the user with no focus.
Work around this temporarily by checking display->expected_focus_window;
this isn't a perfect fix because there are cases where
display->expected_focus_window corresponds to a window we tried to
focus in the past but failed, but it makes things work fairly well.
https://bugzilla.gnome.org/show_bug.cgi?id=597352
When we create the timeline dummy timeline to ensure that our later
functions that should be run during repaint get called called, pass in
G_MAXUINT to make the duration very long, not 0. (It will get reset
whenever there is no repaint later to run, so the fact that G_MAXUINT
is only ~40 days isn't a problem.)
This fixes a warning from Clutter, but also a real problem.
There was a problem where if, for example, a restack was triggered
out of a clutter event handler, then after Clutter processed the
events, it would proceed immmediately on to repaint the stage without
ever returning control to the GLib main loop. So even though we
had an idle handler installed with a higher priority than the
Clutter stage repainting the clutter stage repainting would happen
first and we'd get a wrong frame.
Fix this by introducing the idea of "later functions", which abstract
the idea of "doing something later" away from g_idle_add() and use
a combination of GLib idle functions and Clutter "repaint functions"
to get our callbacks triggered at the right time, even when they
are installed from a clutter event handler.
https://bugzilla.gnome.org/show_bug.cgi?id=596334
This also resolve a FIXME where MUTTER_PRIORITY_BEFORE_REDRAW
could starve stage repainting.
The return value of XGrabKeyboard() wasn't actually being assigned
to the 'result' variable so we didn't notice when grabbing the
keyboard failed.
https://bugzilla.gnome.org/show_bug.cgi?id=596343
We need a way to indicate to gnome-control-center that we want the
keybindings capplet to show the Window Manager keybindings for Metacity;
do this through a _GNOME_WM_KEYBINDING property we put on the
_NET_SUPPORTING_WM_CHECK window and set to Mutter,Metacity.
See http://bugzilla.gnome.org/show_bug.cgi?id=594066 for the
gnome-control-center part of this.
http://bugzilla.gnome.org/show_bug.cgi?id=594067
It seems a bit cleaner to make the MUTTER_DEBUG_XINERAMA variable
that sets up fake Xinerama take effect even if Xinerama is active;
this means we don't count on Xinerama (or Xrandr if we switch tot
that) special casing the case of one monitor.
http://bugzilla.gnome.org/show_bug.cgi?id=593404
Unminimize minimized windows passed to meta_workspace_activate_with_focus()
by calling meta_window_activate() on them instead of meta_window_focus()
and meta_window_raise(). This fix makes sense because for the existing
usage inside Mutter meta_workspace_activate_with_focus() is never called
on a minimized window and for calls from outside Mutter there is no
point in focusing a minimized window without unminimizing it first.
Add a doc comment to meta_workspace_activate_with_focus().
http://bugzilla.gnome.org/show_bug.cgi?id=592393
The changes to enforce single handling of all key events were breaking
custom-alt-tab keypress handlers, since that code was assuming that
key event would get to process_tab_grab(), and then maybe to
process_event() and then to the plugin's xevent_filter to detect a
key release.
We centeralize all of this handling into process_tab_grab() and either
- Invoke a custom handler for the key press
- Select the current window on modifier release by calling a new
pseudo-binding "tab_popup_select"
- Cancel the grab on an unbound key by calling a new pseudo-binding
"tab_popup_cancel"
http://bugzilla.gnome.org/show_bug.cgi?id=590754
mutter_plugin_begin_modal() and mutter_plugin_begin_modal() allow putting
a plugin into a "modal" state. This means:
- The plugin has the keyboard and mouse grabbed
- All keyboard and mouse events go exclusively to the plugin
mutter-plugin.[ch]: Add public API
compositor.c compositor-private.h: Implement the API
mutter-plugin-manager.c: When reloading plugins, make sure none of them
are modal at that moment, and if so force-unmodal them.
common.h: Add META_GRAB_OP_COMPOSITOR
display: When display->grab_op is META_GRAB_OP_COMPOSITOR forward relevant
events exclusively to the compositor.
http://bugzilla.gnome.org/show_bug.cgi?id=590754
Only process each key event once. If all keys are grabbed, then
don't also look for handlers for a key shortcut after processing
the grab op. If all keys are grabbed or we find a key shortcut,
don't pass the event on to the compositing mananger.
http://bugzilla.gnome.org/show_bug.cgi?id=590754
The previous notification code was attempting to use the "modified"
boolean returned from set_title_text, but "that boolean doesn't mean
what you think it means". It actually means "I truncated the title".
Just always notify, it's far simpler than trying to compute
when we don't need to, and callers can compress if they really need
to.
mutter-window.c originally grew an #include "window-private.h" for
window->override_redirect, but that was just fixed. However since
then it also ended up relying on a few other minor private bits.
To fix that, add meta_window_is_mapped, promote meta_window_toplevel_is_mapped
to public, and use the public MetaDisplay accessor.
The functionality to propagate errors for other displays to other
a "foreign error handler" was Soeren's compositor and is no longer
being used. Remove it.
(Now that error.h is being installed and scanned, we need to either
do this or add XErrorEvent to xlib-2.0.gir and rename ErrorHandler
to MetaErrorHandler. This way is a bit simpler.)
When we first start up, we do not want to run effects on any pre-exising
windows (this is either the case we are starting up and there are no windows,
or we are replacing an exisint window manager, or worse, we crashed, and we
just want to get to the desired desktop as quick as possible).
Dithered about where to place the check; putting into the plugin manager
reduces the number of places (and files) in which it needs to be done.
It's useful for plugins to be able to get access to the
startup-notification data that Mutter already has. Add
an accessor and change signal when recieve an event.
When window initially maps, use the more recent of NET_WM_USER_TIME and
startup notification timestamps to compare against last known user action to
decide whether to focus the window or not. Once we show the window, clear
the initial_timestamp_set flag, so the startup notification timestamp is not
taken into account again.
Based on patch for metacity by Alexander Larsson
http://bugzilla.gnome.org/show_bug.cgi?id=573922
We store a pointer to the texture independently of the ClutterContainer
internals, and rely on the pointer remaining valid until we run dispose.
Since we also provide public API to access this pointer, we should not
rely on the reference ClutterContainer holds to ensure that texture will
not be destroyed (e.g., some nasty developer could reparent the texture).
We were freeing the description string in dispose and not setting it to NULL,
thus leaving around a dangling pointer for the duration of the disposal.
This commit moves the free into the finalize vfuction, where it belongs.
Many override-redirect windows (including the Metacity UI windows!)
will have NET_WM_WINDOW_TYPE_NORMAL set on them because of shared
code paths with normal windows in toolkits.
Some current Compositor plugins (default plugin and gnome-shell)
check type == NORMAL to determine if to run effects. While fixing
such plugins to also check if the window is override-redirect is
posisble, it seems cleanest to simply not allow any of the decorated
window types to be set on an override-redirect window and to force
these types to META_WINDOW_OVERRIDE_OTHER. This will prevent other
similar problems from showing up in the future.
http://bugzilla.gnome.org/show_bug.cgi?id=590971
When calculating maximum permissible size of our frame window, we need to
avoid an overflow if the application set its max size hint to INT_MAX.
http://bugzilla.gnome.org/show_bug.cgi?id=590627
As a sideffect of commit a576f7a1ea, override
redirect windows of type _NET_WM_WINDOW_TYPE_NORMAL do not have their
features recalculated during MetaWindow construction (same as regular
windows of type _NET_WM_WINDOW_TYPE_NORMAL), so we need to set the initial
values accordingly.
Although the spec designates some window types as typically used for
override redirect windows, it does not prohibit the use of these with
managed windows, so we should not abort if we encounter one of these.
http://bugzilla.gnome.org/show_bug.cgi?id=583870
As with other events, we want to pass through scroll events (button 4/5 presses)
to the compositor, whether or not they are associated with a particular MetaWindow;
do this by simply falling through to the normal code path instead of
separately delivering the events to the compositor.
http://bugzilla.gnome.org/show_bug.cgi?id=588232
Setting the size of the texture causes the minimum and preferred width and
height values to be fixed at the set value. The normal requisition functions
of ClutterTexture will already report the size of the texture pixmap as the
natural size, but also allow scaling down as needed if less space is
available. We don't need that here, but we want to allow someone to make
a ClutterClone of the texture actor.
With recent changes, Clutter no longer sets up the viewport correctly,
unless it receives ConfigureNotify events. If there is a plugin with
an xevent_filter function, then it's that plugins responsibility to pass
the event to Clutter if it doesn't want it. If there is no plugin,
or the plugin doesn't have an xevent_filter function, then we should
call clutter_x11_handle_event() ourselves.
http://bugzilla.gnome.org/show_bug.cgi?id=589419
When a windows contents or shape changes, we schedule a redraw
with clutter_actor_queue_redraw(); we need to queue the redraw
on the shaped texture rather than on the window actor to support
cloning of just the shaped texture without the shadow: that
is, the shaped is what is really changing and it may be
visible via a clone even if the MutterWindow itself is not
visible.
http://bugzilla.gnome.org/show_bug.cgi?id=589429
MetaPreview is only built into libmutter-private, and not included in
the mutter executable. Linking mutter against libmutter-private was
inadvertently added when the introspection build process was set up,
but isn't actually needed, and if -Wl,-as-needed is added during the
build process, then the libmutter-private dependency will be skipped.
* Don't link mutter (or the test programs) against libmutter-private
* Exclude meta-preview.h from the set of headers we feed into the
introspection build process
Reported by Patryk Zawadzki
http://bugzilla.gnome.org/show_bug.cgi?id=587975
* Make generation of the .gir file work for srcdir != builddir
* Add files to CLEANFILES as needed
* Don't distribute the generated file mutter.schemas
When we are painting a stack of 5-10 maximized windows, the
standard bottom-to-top method of drawing every actor results
in a tremendous amount of overdraw and can easily max out
the available memory bandwidth on a low-end* graphics chipset.
It's even worse if window textures are being accessed over
the AGP bus.
When we have opaque windows, we can go ahead and compute visibility
ourselves (in classic X-server fashion) and use that information to
restrict drawing obscured actors.
* Add MutterWindowGroup - a ClutterGroup subclass with logic
for figuring out obscured regions.
* Add mutter_window_get_obscured_region() to get the region
obscured by that window.
* Add mutter_shaped_texture_set_clip_region() to hint
a clip region to the painting code; this is set based on
the computed visible region of MutterWindowGroup.
* Add tidy_texture_frame_set_needs_paint() to hint that the
paint can be skipped entirely; this is used when we detect
that the window shadow is entirely obscured.
http://bugzilla.gnome.org/show_bug.cgi?id=587344
If we are starting on something other than the first workspace,
meta_compositor_switch_workspace() will be called before
meta_compositor_manage_screen(); guard against that.
Previously, changes to the visibility of a window could be indicated
by meta_compositor_map_window(), meta_compositor_unminimize_window(),
meta_compositor_set_window_hidden(), etc, with the exact behavior
depending on the 'live_hidden_windows' preference.
Simplify this so that visibility is controlled by:
meta_compositor_show_window()
meta_compositor_hide_window()
With an 'effect' parameter provided to indicate the appropriate
effect (CREATE/UNMINIMIZE/MINIMIZE/DESTROY/NONE.)
The map state of the window is signalled separately by:
meta_compositor_map_window()
meta_compositor_unmap_window()
And is used only to control resource handling.
Other changes:
* The desired effect on show/hide is explicitly stored in
MetaWindow, avoiding the need for the was_minimized flag.
At idle, once we calculate the window state, we pass the
effect to the compositor if it matches the new window
state, and then clear the effect to start over for future
map state changes.
* meta_compositor_switch_workspace() is called before any windows
are hidden or shown, allowing the compositor to avoid hiding
or showing an effect for windows involved in the switch.
http://bugzilla.gnome.org/show_bug.cgi?id=582341
* Handling of post-effect cleanups for MutterWindow are
simplified - instead of trying to do different things based
on the individual needs of different effects, we just wait until
all effects complete and sync the window state to what it
should be.
* On unmap, once we destroy the pixmap, we tell ClutterX11Pixmap
that we've done so, so it can clean up and unbind. (The
unbinding doesn't seem to be working properly because of
ClutterGLXPixmap or video driver issues.)
http://bugzilla.gnome.org/show_bug.cgi?id=587251
Clean up mutter_window_effect_in_progress() by removing the
include_destroy parameter which was used only in one place that
could be easily done otherwise. (There was another use in
mutter_window_sync_actor_position() that had no point and looked
unintended.)
http://bugzilla.gnome.org/show_bug.cgi?id=587251
Add a paint function that checks all windows for repair and
shape updates; this:
- simplifies the logic for when a window needs to be repaired
- avoids duplicate work when we get multiple damage effects
- avoids the need to look ahead in the event queue
Instead of relying on repair to implicitly resize the
MutterWindow actor, set the size explicitly when the core
code updates the geometry. (This is needed because we haven't
repaired yet when we start an animation, and the animation
may depend on the size to, e.g., rescale from the center.)
Because the core geometry update happens before we start
maximize/unmaximize effects we need to work around this by
passing both the old and new geometry to the compositor.
http://bugzilla.gnome.org/show_bug.cgi?id=587251
Putting hidden windows in the desktop layer is pointless - in
the desktop layer isn't necessary below all visible windows,
and we are hiding the windows by other means. And the movement
isn't reliable because nothing sets stack->needs_relayer, so
windows can get stuck in the desktop layer after being
rehidden.
http://bugzilla.gnome.org/show_bug.cgi?id=587251
