Clutter 1.7.x introduced CLUTTER_CAIRO_FORMAT_ARGB32: which can be used when
sharing textures/data with cairo without having to do check the
byte order and choose the appropriate format by hand.
https://bugzilla.gnome.org/show_bug.cgi?id=654577
Update the sample to be more up to date with respect to Shell practices,
and make it look a bit prettier. Additionally, change the file extract
code so that it's easier to update and add new files later.
https://bugzilla.gnome.org/show_bug.cgi?id=653206
Ideally, this would be an entirely-JS implementation, but we have a
couple of issues with gjs and gobject-introspection to work around, so
we need a ShellMountOperation class for the time being.
This first commit implements the show-processes dialog, with a system
modal style very similar to the EndSession dialog.
Implementations of ask-question and ask-password will follow shortly.
https://bugzilla.gnome.org/show_bug.cgi?id=653520
The sniffer is a simple helper process, activated as a DBus service,
that tries to crawl as many files as possible in the provided target
directory (i.e. the new mount's root), for a maximum amount of time -
which is set here to 1.5 seconds (i.e. it will crawl either all the
files in the directory tree, or as many as it can before the specified
timeout expires).
Crawled files are ordered by their content type, and a generic estimation
of the type of files composing the directory is returned to the caller,
using generic 'x-content/*' mimetypes.
The process will then set an autoquit timeout on itself, which can be
disabled by setting the env variable HOTPLUG_SNIFFER_PERSIST for
debugging purposes. The HOTPLUG_SNIFFER_DEBUG env variable can also be
set to enable debugging output.
https://bugzilla.gnome.org/show_bug.cgi?id=653520
==17386== 1,669 (88 direct, 1,581 indirect) bytes in 1 blocks are definitely lost in loss record 4,090 of 4,151
==17386== at 0x4C24AF4: calloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==17386== by 0x691B099: g_malloc0 (in /usr/lib/libglib-2.0.so.0.2800.8)
==17386== by 0x692006A: g_option_context_new (in /usr/lib/libglib-2.0.so.0.2800.8)
==17386== by 0x5124C57: meta_get_option_context (in /usr/lib/libmutter.so.0.0.0)
==17386== by 0x401D4F: main (in /usr/bin/gnome-shell)
https://bugzilla.gnome.org/show_bug.cgi?id=654269
The gconf keys used to restore GNOME aren't in a proper GNOME3 environment.
To mimic what GNOME3 gnome-session does would be extremely complicated, so
just launch the system gnome-shell.
https://bugzilla.gnome.org/show_bug.cgi?id=654527
The cogl path pads the corners out to the maximum corner radius to make the
math and painting logic easier. Unfortunately, when the radius exceeds the
actor's halfsize, the padding ends up interfering with other corners, creating
a big mess of rendering errors.
It'd be extremely complicated to fix this properly in the Cogl code,
so take the Cairo fallback.
https://bugzilla.gnome.org/show_bug.cgi?id=649513
Currently, any cases of overlapping corners were just ignored and rendered incorrectly.
Implement the corner overlap algorithm as specified by the W3C to fix this.
https://bugzilla.gnome.org/show_bug.cgi?id=649513
Unfortunately, gjs cannot handle binary C strings directly from
gobject-introspection. Add a simple workaround method in C to help
us save random files from the web.
https://bugzilla.gnome.org/show_bug.cgi?id=653989
Remove ShellGlobal's monitor-related methods, and have
Main.layoutManager provide that information instead. Move
Main._relayout() to LayoutManager, and have other objects connect to
the layout manager's 'monitors-changed' signal to know when the screen
geometry has changed.
https://bugzilla.gnome.org/show_bug.cgi?id=636963
shell_util_get_file_display_name_if_mount() uses a nautilus setting
to use a name for $HOME which is consistent with the file manager.
But while nautilus was ported to GSettings a while ago, we are still
trying to access the old GConf setting.
https://bugzilla.gnome.org/show_bug.cgi?id=653511
Only skip the areas of the scrollbars when they are invisible
and add take the horizontal scrollbar into account as well
when calculating the faded area.
https://bugzilla.gnome.org/show_bug.cgi?id=651866
jhbuild will default to srcdir != builddir in the future for various
reasons.
The enum generation rules were just broken and both did
cd $(srcdir) and $(addprefix $(srcdir)).
Also, remove an unnecessary $(addprefix) from the GIR sources; thanks
to Dan Winship for pointing out that make will look in both srcdir and
builddir, so it's not necessary to add a prefix explicitly. Doing so
breaks obviously when adding the sourcedir to a builddir file.
https://bugzilla.gnome.org/show_bug.cgi?id=653199
Using the list of stylesheets loaded with st_theme_load_stylesheet(),
one can build an StTheme that is completely identical to the previous
one, except for one property (application-stylesheet).
This allows rt and the user-theme extension to work while respecting
the theming of other extensions.
https://bugzilla.gnome.org/show_bug.cgi?id=650971
As the Shell does more than observing channels (users can interact with them),
it has to be an Handler as well. We have to make sure that all the new
incoming text channels are handled by the Shell by default, so we make it an
Approver as well.
From an user point of view, the only difference is that Empathy's tray icon
will stop blicking when receiving new channels.
We rely on ChannelDispatcher.DelegateChannels() and PresentChannel() to
interact with Empathy. Those methods have been implemented in
telepathy-mission-control 5.9.0 and telepathy-glib 0.15.0.
shell_global_get_memory_info() is a new function which extracts a few
global counters we have already, namely glibc's mallinfo, spidermonkey's
JSGC_BYTES, and gjs' counters for boxed/object/etc wrappers.
There is some slight overlap with perf; ultimately though I'd
like this function to do some more extensive analysis, so it wouldn't
be quite the same.
perf is going to be mainly concerned with how big the whole process
over time is; memory_info is for debugging memory leaks.
https://bugzilla.gnome.org/show_bug.cgi?id=650692
Theme authors now have the power (and responsibility) of creating fade
effects with the new CSS length property '-st-fade-offset'. A value of
0 disables the effect.
This new CSS approach replaces the current programmatic toggle of
the 'vfade' property. A new CSS style class name 'vfade' is used as
a replacement for the old property.
https://bugzilla.gnome.org/show_bug.cgi?id=651813
cogl_read_pixels() used to only support a useless pixel format, but it
will do our preferred format now, so use it rather than doing GL stuff
by hand.
https://bugzilla.gnome.org/show_bug.cgi?id=648758
Remove a workaround for clutter_actor_get_transformed_position() not
working inside paint(), and remove a comment about
ClutterText::position not being properly notified, since it is now.
(However, it doesn't seem worth it to rewrite the code to use
notification, since that would actually end up being more complicated
than the current solution.)
https://bugzilla.gnome.org/show_bug.cgi?id=648758
StScrollBar was tracking whether or not it currently had a valid
allocation, but since Clutter 1.4 there is a method it can call to get
that information instead.
https://bugzilla.gnome.org/show_bug.cgi?id=648758
A new envrionment variable GNOME_SHELL_ENABLE_CLEANUP is added which
causes us to attempt freeing global data. The reason this isn't
enabled by default is that it's a waste of time at best, and at
worst in corner cases could cause crashes which would fill up
crash databases. Better to leave it as a developer-only tool.
Start stubbing out some cleanup in ShellGlobal.
https://bugzilla.gnome.org/show_bug.cgi?id=649517
Move some more environment-initializationy stuff from main.js to
environment.js, and be more careful about not importing shell JS
modules until after the environment has been fully patched.
Change gnome-shell-plugin to call Environment.init() before
Main.start(); this means that Environment.init() now runs before any
shell JS modules (besides environment itself) have been imported.
Make run-js-test create a ShellGlobal and use its js_context, so that
the shell_global_set_property_mutable() stuff in Environment.init()
will work correctly in tests as well.
https://bugzilla.gnome.org/show_bug.cgi?id=649203
The tests were broken again, because since Shell-0.1.gir now has
'shared-library="libgnome-shell.so"', the references to Shell.PerfLog
ended up pulling in libgnome-shell in addition to the copy of
shell-perf-log.c that libjs-test was built with.
Fix all this hopefully forever by just making run-js-test link to
libgnome-shell.
https://bugzilla.gnome.org/show_bug.cgi?id=649203
shell-global had become a dumping ground for functions that didn't
have anywhere else to be. Make shell-util the dumping ground instead,
and have shell-global only have methods that involve the ShellGlobal
object.
https://bugzilla.gnome.org/show_bug.cgi?id=648755
==13810== 11,360 bytes in 1 blocks are definitely lost in loss record 18,574 of 18,765
==13810== at 0x4005447: calloc (vg_replace_malloc.c:467)
==13810== by 0x5191882: standard_calloc (gmem.c:107)
==13810== by 0x51920A7: g_malloc0 (gmem.c:196)
==13810== by 0x4056201: blur_pixels (st-private.c:466)
==13810== by 0x40573B4: _st_create_shadow_cairo_pattern (st-private.c:710)
==13810== by 0x4070746: st_theme_node_paint (st-theme-node-drawing.c:856)
==13810== by 0x3FEFFFFF: ???
https://bugzilla.gnome.org/show_bug.cgi?id=649497
When activating an uninteresting window, the last_user_time isn't updated,
because we aren't tracking the window that the user_time gets updated on.
Hack around this by setting the last_user_time in shell_app_activate when
activating an uninteresting window.
https://bugzilla.gnome.org/show_bug.cgi?id=643302
When activating an uninteresting window, the last_user_time isn't updated,
because we aren't tracking the window that the user_time gets updated on.
Hack around this by setting the last_user_time in shell_app_activate when
activating an uninteresting window.
https://bugzilla.gnome.org/show_bug.cgi?id=643302
This property represents that the widget is being labelled by an
actor. The name is label-actor to avoid problems with the current
StButton:label and StTooltip:label
Before this change, we displayed dialogs on the monitor containing the
focused window if there was any, otherwise on monitor 0. We now use
the primary monitor rather than monitor 0 when no window has focus.
https://bugzilla.gnome.org/show_bug.cgi?id=648305
I unintentionally made .desktop->pid association "win" over
WM_CLASS. Fixing this makes the case of ancillary .desktop file
entry points (e.g. gnome-control-center's various shortcut .desktop
files) correctly show System Settings, and not whatever the shortcut
is.
In the future I'd like to have a way to say "this .desktop file
is a shortcut, ignore me" or something.
https://bugzilla.gnome.org/show_bug.cgi?id=646689
We weren't specifying _ALLOW_REPLACEMENT for anything except
org.gnome.Shell, which created a race - if the exiting process
didn't exit fast enough, the replacing process would fail
to get the name.
https://bugzilla.gnome.org/show_bug.cgi?id=646257
As a side effect of (see bug 642221), we no longer put docks or
transient windows into the hash table mapping windows to apps. The
"focused application" code relied on at least transients being in
there.
Fix this by calling the public API to map a window to an app, which
will at least follow transients. Whether we also want further
matching here (e.g. with window grouping) is another issue, but that
can happen as a different bug.
https://bugzilla.gnome.org/show_bug.cgi?id=647082
If a caller sets an StLabel's text to what it already is (as, eg, the
clock menu does), do nothing. Unless the label is editable, in which
case, setting the text has a visible side effect (dropping the
selection), so we don't optimize that out.
https://bugzilla.gnome.org/show_bug.cgi?id=645648
Add an uninstall-hook to undo the effect of install-exec-hook, and add
an $(AM_V_GEN) to the rule that copies either gnome-shell-real or
gnome-shell-jhbuild to gnome-shell.
https://bugzilla.gnome.org/show_bug.cgi?id=646730
If we're unmapped (or destroyed) during a scroll, we want to clean
up the changes we've made to Clutter's event handling, remove our
signal handler, and emit ::scroll-stop.
https://bugzilla.gnome.org/show_bug.cgi?id=646825
Adding correct annotations to Gio.File.load_contents revealed that gjs
doesn't actually support array+length combinations. For 3.0 this would
be invasive to fix, so add a method to ShellGlobal which does what
we need.
https://bugzilla.gnome.org/show_bug.cgi?id=646333
Add a function that gets the current allocation of an actor
transformed into stage coordinates. This avoids a misfeature of
clutter_actor_get_transformed_size() where when a size request is
queued (even if it won't eventually change the size), the returned
value is the transformed size request rather than the last allocation.
https://bugzilla.gnome.org/show_bug.cgi?id=645744
settings.menu was removed in gnome-menus commit
b68bcd27f44ce2c494f6e3cd9695890b9c02af04; gnomecc.menu is the intended
replacement.
(On Red Hat Linux derived systems, settings.menu continues to exist)
https://bugzilla.gnome.org/show_bug.cgi?id=645063
Previously (because I suck) we were ignoring the return value of
RequestName, and so we'd totally ignore the fact that we failed
to acquire the DBus name.
Make this consistent by using meta_get_replace_current_wm() and
if we're in --replace, actually replace immediately.
https://bugzilla.gnome.org/show_bug.cgi?id=645593
The next draft of the CSS Backgrounds and Borders module will actually
define when the blur radius means. Fix our code to use that definition
(2 * standard deviation) rather than using the 1.9 * that we extracted
from what Mozilla was doing.
https://bugzilla.gnome.org/show_bug.cgi?id=632506
Some functions in StTextureCache enforce square ClutterTextures,
even in cases where the underlying CoglTexture has a different
width:height ratio.
Add padding in those cases to keep the resulting image from being
stretched.
https://bugzilla.gnome.org/show_bug.cgi?id=643866
If, for example, the stage is divided into multiple monitors, we
might want to constrain tooltips so they don't cross monitor boundaries.
Add a function to set a per-stage callback to constrain tooltips.
https://bugzilla.gnome.org/show_bug.cgi?id=645547
Instead of showing tooltips immediately on hover, wait until a timeout
after the last motion (timeout is given by the gtk-tooltip-timeout
GtkSetting.)
https://bugzilla.gnome.org/show_bug.cgi?id=642871
Use ClutterContainer functions for adding the tooltip instead of
calling clutter_actor_set_parent behind the stage's back, and do
it inside st_widget_show_tooltip (which is a normal method) instead
of overriding st_tooltip_show, which is a vfunc and it is called
internally by Clutter, therefore it is limited in what it can safely
do.
Also, instead of positioning the tooltip with clutter_actor_set_position,
modify the anchor point when the associated widget moves, so that
only a redraw is queued.
https://bugzilla.gnome.org/show_bug.cgi?id=635100
Inside the Shell, all the UI (including chrome, the overview, and
the actual windows) is not a child of the stage but of a special
ClutterGroup, which is cloned inside the magnifier.
Add function for setting this special actor so that actors added by
St are visible in the magnifier. Nothing yet uses this, but the
tooltip will soon.
https://bugzilla.gnome.org/show_bug.cgi?id=635100
As of commit 34ce17c4b3, search results use large icons, or thumbnails
when available. To keep the amount of upscaling for the latter as small
as possible, request a large thumbnail size.
https://bugzilla.gnome.org/show_bug.cgi?id=645493
If you have XFixes 5 (and corresponding xserver support) then we
add barriers on the panel and in the message tray corner so that
its easy to reach the corners even when there are monitors to the
sides of the primary monitor.
https://bugzilla.gnome.org/show_bug.cgi?id=622655
If the pointer moves on or off the stage while another process has a
grab, we will lose track of it. One example of this is that if you use
a popup menu from a message tray trayicon, the tray will stay up after
the menu goes away, because the shell never saw the pointer leave it.
Add a new method shell_global_sync_pointer() that causes clutter to
recheck what actor is under the pointer and generate leave/enter
events if appropriate.
Of course, we can't actually tell for sure when another process has a
grab, so we need a heuristic of when to call this. Currently we call
it from Chrome._windowsRestacked(), which is not really the right
thing at all, but does fix the menu-from-trayicon case...
https://bugzilla.gnome.org/show_bug.cgi?id=630842
If the user clicks on the title of a trayicon's SummaryItem, forward
that click to the trayicon. Also adjust
gnome_shell_plugin_xevent_filter() so that if the trayicon takes a
grab as a result of this, we don't hide the message tray.
https://bugzilla.gnome.org/show_bug.cgi?id=630842
StButton was mistakenly considering any Space/Enter KEY_RELEASE to be
a click, when in fact it should only count as a click if it also got
the corresponding KEY_PRESS as well. This meant that when typing in a
chat notification, any Space/Enter keypress would dismiss the
notification, since the StEntry would take the PRESS event but ignore
the RELEASE, allowing it to propagate to the notification itself,
which would treat it as a click.
https://bugzilla.gnome.org/show_bug.cgi?id=645243
Alt+F2 restart was failing after a rebuild when running from the
source tree because it would try to restart
".libs/lt-gnome-shell-real", which didn't exist yet. Fix this by using
"libtool --mode=execute" at build time to regenerate that file.
https://bugzilla.gnome.org/show_bug.cgi?id=645390
In commit 9bd22dc0, I introduced an API to load an arbitrary
.desktop file, not necessarily from the menu path. It turns
out this function was broken because it created ShellApp instances
that were *different* from ones that were cached normally.
As far as I can tell, we didn't initially use it. Then later
Util.spawnDesktop was created which used this function.
Remove this broken function and all callers; if we're loading
.desktop files from *outside* the menu path, we can look at
readding.
This patch also kills off Util.spawnDesktop in favor of callers
talking to ShellAppSystem directly, now that the latter reports
errors.
https://bugzilla.gnome.org/show_bug.cgi?id=644402
Move the "system notification error" handling out of
util.js, and add it to ShellGlobal so we can start
calling it from across the codebase better (including
C).
https://bugzilla.gnome.org/show_bug.cgi?id=644402
When jhbuilding, we use a jhbuilt gobject-introspection, so the
default typelib path is the jhbuild prefix, not /usr. So if we are
using NetworkManager from packages, we need to adjust GI_TYPELIB_PATH
to include it.
https://bugzilla.gnome.org/show_bug.cgi?id=621707
Adds an implementation of nm-applet in javascript. Uses the new
introspection from NetworkManager, and temporarily requires
nm-applet to be running for the secret service.
Features a renewed interface, with each device controllable through
a switch, which if toggled off disconnects, and if toggled on
connects to the most recently used valid connection. More esoteric
features like creation of ad-hoc networks have been moved to the
control center panel.
https://bugzilla.gnome.org/show_bug.cgi?id=621707
gdm_user_get_real_name() checks for an empty real
name and automatically falls back to username if
real name is NULL. It doesn't automatically fall
back to username if real name is empty, however.
This commit makes it fall back for both cases.
https://bugzilla.gnome.org/show_bug.cgi?id=644765
When we compare the boxes for two actors, they may appear to overlap
by a small amount because of floating-point imprecision. Allow for
up to 0.1 pixel overlap when determining what children are in the
focus direction from the currently focused actor.
https://bugzilla.gnome.org/show_bug.cgi?id=644134
If a process does not have any "interesting" windows, then it can't be
considered a running app. (Previously we were calling
get_app_for_window() before ruling out non-interesting windows, which
ended up calling _shell_app_new_for_window(), which would add the
window to the ShellApp directly, bypassing the is_interesting check.)
https://bugzilla.gnome.org/show_bug.cgi?id=642221
This is another workaround for the lack of gjs supporting array
arguments, this time wrapping tp_connection_upgrade_contacts to
add new features to the connection's self contact.
https://bugzilla.gnome.org/show_bug.cgi?id=642793
Monkey-patch Date.prototype.toLocaleFormat() with a version that uses
g_date_time_format() since the Spidermonkey built-in can't handle
format strings with Unicode characters.
https://bugzilla.gnome.org/show_bug.cgi?id=643350
clutter_init() fails under normal circumstances like
being unable to open a display connection, so it shouldn't
be handled with g_error() producing a core dump.
Clutter consistently produces an error message when
clutter_init() fails, so we don't need to print out any
error message.
https://bugzilla.gnome.org/show_bug.cgi?id=643910
The number instructions in a shader is limited to 64 on r300 hardware,
the fade shader in StScrollViewFade was ending up using 97 instructions
which is way over the limit.
So refactor the shader to use less instructions by precomputing as many
values as possible outside of the conditionals. The resulting shader
ends up using 34 instructions which is well within the hardware limits.
https://bugzilla.gnome.org/show_bug.cgi?id=644589
GThemedIcon expects the first name to be the most specific, and
will thus prefer it to later ones. We thus need to order the names
from the longer to the shorter.
https://bugzilla.gnome.org/show_bug.cgi?id=621707
* Run gnome-shell-perf-helper during performance tests
* Use MUTTER_WM_CLASS_FILTER to omit all other windows
* Add new Scripting methods: createTestWindow,
waitTestWindows, destroyTestWindows
* Create a single 640x480 test window for testing overview
animation performance.
https://bugzilla.gnome.org/show_bug.cgi?id=644265
We need to connect to the NameOwnerChanged signal before we execute the
DConf binary. Refactor things so that we connect to the signal when
we first get the bus object. Split the code for waiting for a D-Bus
name into a wait_for_dbus_name() function for latter reuse.
https://bugzilla.gnome.org/show_bug.cgi?id=644265
When navigating from a non-immediate descendant of a container, we
were attempting to use clutter_actor_get_transformed_position() to get
the exact position of that actor relative to the container, but this
did not really make sense, since we would be using the position of
the intermediate container when navigating back.
https://bugzilla.gnome.org/show_bug.cgi?id=644134
g_themed_icon_new_with_default_fallbacks() does not do what we want
with symbolic icons; if the user's icon theme is not "gnome", then it
will end up preferring a non-symbolic icon from the higher-level theme
over a symbolic icon from gnome-icon-theme-symbolic.
If the shell requests a symbolic icon, and there is only a
non-symbolic icon available, that should be considered a programmer
error, just like requesting a non-existent icon name. So change the
code to only look up "-symbolic" names when drawing an
ST_ICON_SYMBOLIC icon.
https://bugzilla.gnome.org/show_bug.cgi?id=644142
* calendar-server subdirectory of build directory needs to
be created.
* Generate shell-enum-types.c/.h in build directory rather than
in the possibly-read-only source directory.
Build gnome-shell as a binary linked against libmutter-wm, instead of
a module to be loaded by libmutter-wm. Move the majority of
initialization-type stuff from gnome_shell_plugin_start() into main().
We still build libgnome-shell as a shared library, so that the linker
doesn't discard all the methods that are never called from C.
https://bugzilla.gnome.org/show_bug.cgi?id=641724
Add Ctrl-Alt-Tab support to ViewTab, and fix the Applications pane to
scroll to track the keyboard focus.
The Windows pane can be switched to, but navigation within the pane is
not yet implemented.
https://bugzilla.gnome.org/show_bug.cgi?id=618887
PopupMenuManager was pretending that it knew nothing about the menu's
sourceActors, while also trying to handle keynav between them. This
was a big mess, and resulted in bugs in navigation between panel menus
and the Activities button, and it totally gets in the way when trying
to add keynav to the dash (whose menu sources are arranged vertically
rather than horizontally).
Fix this up by moving the panel-specific parts to PanelMenuButton
instead.
https://bugzilla.gnome.org/show_bug.cgi?id=641253
At times, RTL locales require different CSS, so always create theme
nodes with :ltr/:rtl pseudo classes depending on the widget's text
direction, to provide an easy way to handle those cases without
requiring a separate stylesheet.
https://bugzilla.gnome.org/show_bug.cgi?id=643835
Instead of defining rules manually, use the included make support,
which means Makefiles are less prone to API changes, support
verbosity correctly (instead of inconsistently hardcoding --quiet),
are cleaner to read, and emit GISCAN/GICOMP instead of GEN (which
is nicer).
https://bugzilla.gnome.org/show_bug.cgi?id=643371
The current gnome-shell.in script has a huge amount of
unnecessary complexity for the installed, normal case. Fix
this by adding a configure option (defaulting to false) that
installs a simple, obvious wrapper script around mutter.
We do change the gnome-shell build setup to pass this option
by default for jhbuild.
https://bugzilla.gnome.org/show_bug.cgi?id=642084
Unfortunately the evolution-data-server client-side libraries seem to
block the calling thread. This is a major problem as we must never
ever block the main thread (doing so causes animations to flicker
etc.). In the worst case, this problem causes login to hang (without
falling back to fall-back mode) and in the best case it slows down
login until a network connection is acquired.
Additionally, in order to sanely use these evolution-data-server
libraries, GConf has to be involved and GConf is not thread-safe. So
it's not really feasible just moving the code to a separate
thread. Therefore, move all calendar IO out of process and use a
simple (and private) D-Bus interface for the shell to communicate with
the out-of-process helper.
For simplification, remove existing in-process code since internal
interfaces have been slightly revised. This means that the shell is no
longer using any native code for drawing the calendar dropdown.
https://bugzilla.gnome.org/show_bug.cgi?id=641396
Signed-off-by: David Zeuthen <davidz@redhat.com>
See commit f2158218bef0c51 in mutter. Basically, we need
to grab org.freedesktop.Notifications before anything else
in the session gets started.
Note: I intentionally removed the Util.killall bits. I believe that
for notification-daemon at least, if we specify
DBUS_NAME_FLAG_REPLACE_EXISTING, we'll take over the name. Not sure
about notify-osd; if that's still a problem, then what we need to do
is add killing (and possibly respawning) of notify-osd to
"gnome-shell --replace", and not have it embedded randomly in a JS file.
https://bugzilla.gnome.org/show_bug.cgi?id=642666
