We want to be able to summarize the behavior of the shell's
performance in a series of "metrics", like the latency between
clicking on the Activities button and seeing a response.
This patch adds the ability to create a script under perf/
in a special format that automates a series of actions in the
shell, writing events to the performance log, then collects
statistics as the log as replayed and turns them into a set
of metrics.
The script is then executed by running as gnome-shell
--perf=<script>.
The 'core' script which is added here will be the primary
performance measurement script that we use for shell performance
regression testing; right now it has a couple of placeholder
metrics.
https://bugzilla.gnome.org/show_bug.cgi?id=618189
ShellPerfLog provides a way for different parts of the code to
record information for subsequent analysis and interactive
exploration. Events exist of a timestamp, an event ID, and
arguments to the event.
https://bugzilla.gnome.org/show_bug.cgi?id=618189
To support scheduling performance-measurement scripts that want to run
a number of actions in series, add shell_global_run_at_leisure() to run
a callback when all work is finished.
The initial implementation of this is not that accurate: we track
business in Tweener.js via new shell_global_begin_work(),
shell_global_end_work() functions, and we also handle the case
where the main loop is continually busy.
https://bugzilla.gnome.org/show_bug.cgi?id=618189
The API docs for ShellApp claimed it sorted by the last time the
user interacted with the app, but if one closed a window, then
we would fall back to comparing against a possibly much older
timestamp from another window. Fix this by just keeping a
user time per app.
Also clean up the comparison function to explicitly check the state
instead of deferring to the window list.
https://bugzilla.gnome.org/show_bug.cgi?id=618378
The ShellGlobal initialization performs several actions like connecting
to the X server, ensuring directories exist, etc., that are problematic
because we were creating the object even when running the binary for
introspection scanning. During compilation we may not even have X11
available in e.g. autobuilder type environments, and it's just a
bad idea to connect even if we do.
Avoid this by deferring creation of the ShellGlobal object
until the plugin is actually started.
Now that we're initializing things later, remove the connection to
screen changes, and initialize cached ShellGlobal state at the point
when the plugin is set. The root pixmap actor is now sized initially
on creation too. Instead of relying on screen-size-changed being
emitted on startup, explicitly invoke _relayout().
https://bugzilla.gnome.org/show_bug.cgi?id=618371
I have no idea why there existed code that if we saw e.g. min-width
without a width, we assigned min-width to ->width, thus effectively
treating it as a maximum.
Just delete that bit.
https://bugzilla.gnome.org/show_bug.cgi?id=618482
Starting with gtk-2.20.0 there is a gdk_screen_get_primary_monitor,
which supports querying the primary monitor from xrandr.
But due to a sorting bug and lack of heuristics in the fallback path,
it isn't really useable.
Those bugs are fixed in gtk-2.20.1, so use it when building with
gtk-2.20.1+.
https://bugzilla.gnome.org/show_bug.cgi?id=608647
Passing an explicit width in the wfh case or a height in the hfw case
messes up the request caching, and confuses actors that assume they
won't be called with an explicit width/height unless they're being
allocated along the other axis.
https://bugzilla.gnome.org/show_bug.cgi?id=618295
Adds the ability to create one or more zoom regions that show magnified or
enhanced views of the desktop. The magnifier provides options for:
* magnification factor,
* four mouse tracking modes common to screen magnifiers,
* positioning the magnified view in one of four screen location, or full screen,
* crosshairs to accentuate the position of the mouse,
* user preferences persistence via GConf (schemas in
.../data/gnome-shell.schemas).
* a DBus API to allow other processes to drive the magnifier as a service.
https://bugzilla.gnome.org/show_bug.cgi?id=595507
Add StContainer, which implements the ClutterContainer interface based
on the container methods in st-private and make the existing containers
subclass it.
https://bugzilla.gnome.org/show_bug.cgi?id=613907
To replace all calls to deprecated code, GTK+ 2.20 is required - add
some basic compatibility code, so that is still possible to build the
shell with GTK+ 2.18 when not using -DGSEAL_ENABLE.
https://bugzilla.gnome.org/show_bug.cgi?id=618258
With the transition to GTK+ 3.0, direct access to struct members
will no longer be possible.
This bumps the required minimum version of GTK+ to 2.20.
https://bugzilla.gnome.org/show_bug.cgi?id=618258
The StTable code only supports height-for-width. When called in
width-for-height sizing mode, instead of treating the -1 flag
value of 'for_width' as a real width, and requesting all the
children at 1 pixel wide, use the natural width of the table
as the width for determing the height.
Since we can't rewrap in width-for-height mode, we then report
the natural width also as the minimum width of the table.
https://bugzilla.gnome.org/show_bug.cgi?id=618104
readlink() on /proc/<pid>/exe can have results like:
/usr/bin/gnome-keyring-daemon.#prelink#.5DFZsF (deleted)
/usr/bin/gnome-session (deleted)
To find gnome-session in a more robust way, read /proc/<pid>/cmdline
instead.
https://bugzilla.gnome.org/show_bug.cgi?id=616706
The design calls for raising all windows for a given app in
certain circumstances; implement this. The new _focus method
raises all windows for the app if it's running.
We further change the _activate method (which a lot of the shell
UI calls now) to invoke _focus for the running case, which means
that e.g. the application well will now raise all app windows.
https://bugzilla.gnome.org/show_bug.cgi?id=616051
When creating the directory to store user data, XDG_DATA_HOME is
not guaranteed to exist. Also, the standard mandates permissions
of 0700 for the directory.
https://bugzilla.gnome.org/show_bug.cgi?id=617555
As StEntry handles hover differently depending on whether it is
activated or not, the generic hover support in StWidget is
insufficient. Nevertheless it makes sense to set the hover status
using StWidget methods instead of setting the pseudo class directly.
While the extension system already uses an XDG location (XDG_CONFIG_HOME),
other components use the deprecated $HOME/.gnome2 directory.
Replace both with XDG_DATA_HOME - the existing data (app usage stats,
looking glass history and extensions) is not migrated to the new location.
https://bugzilla.gnome.org/show_bug.cgi?id=617555
The (optional) spread radius allows to make the shadow bigger without
enlarging the blur value. Mozilla supports this parameter for the
-moz-box-shadow property.
https://bugzilla.gnome.org/show_bug.cgi?id=613832
To avoid visual duplication with the application menu, eliminate
the window menu button. (The window menu can still be accessed by
right-clicking on the title bar, or Alt-right-clicking anywhere on the
window.)
This is implemented using the new Mutter facility for GConf key
redirection; we add separate key for 'button_layout' with a default
that omits the menu button.
https://bugzilla.gnome.org/show_bug.cgi?id=591390
With the changes from bug 615586, Mutter now expects a separate start()
method which is called after construction. Move our initialization from
constructed() to start() so it can access the MetaScreen.
https://bugzilla.gnome.org/show_bug.cgi?id=615592
This patch combines several high level changes which are conceptually
independent but in practice rather intertwined.
* Add a "state" property to ShellApp which reflects whether it's
stopped, starting, or started. This will allow us to later clean
up all the callers that are using ".get_windows().length > 0" as
a proxy for this property
* Replace shell_app_launch with shell_app_activate and shell_app_open_new_window
A lot of code was calling .launch, but it's signficantly clearer
if we call this ".open_new_window()", and later if we gain the ability
to call into an application's menu, we can implement this correctly rather
than trying to update all .launch callers.
* Because ShellApp now has a "starting" state, rebase panel.js on top of
this so that when we get a startup-notification sequence for an app
and transition it to starting, it becomes the focus app, and panel.js
cleanly just tracks the focus app, rather than bouncing between SN
sequences. This removes display of non-app startup sequences, which
I consider an acceptable action in light of the committed changes
to startup-notification and GTK+.
https://bugzilla.gnome.org/show_bug.cgi?id=614755
We can't use gdk_display_get_pointer/gdk_window_get_pointer from gjs
when XKB is active. We already had a wrapper that did the
get-modifier-state part of that, but some places also need the
get-pointer-location part of it. So update our wrapper to return both,
and update js code to use it.
https://bugzilla.gnome.org/show_bug.cgi?id=613428
Also, remove a lot of cruft from genericDisplay.js leftover from
previous St-ifications, and remove the pre-gtk-2.16 hacks from the
status tray in panel.js (which are much less needed with the
nearly-all-black panel anyway).
https://bugzilla.gnome.org/show_bug.cgi?id=614516
The actual changes to shell-menu.[ch] are pretty minimal; most of the
changes there are just style/spacing/indentation.
Also, removed shell_menu_append_separator() since it wasn't needed;
the separators would already have been behaving as intended just
because they were non-reactive.
https://bugzilla.gnome.org/show_bug.cgi?id=614516
Most subclasses override get_preferred_width/height, but if they don't
(eg, StDrawingArea), then make sure they still take CSS-specified
sizes into account.
https://bugzilla.gnome.org/show_bug.cgi?id=614516
StThemeNode holds a reference to its parent, but we never released
that reference. This could cause us to hold onto whole chains
of theme nodes with rather dire memory usage implications.
Also move the other g_object_unref into _dispose.
https://bugzilla.gnome.org/show_bug.cgi?id=614660
Currently shell_global_get_primary_monitor just returns the first screen,
in the list as primary.
This is not always correct as the first screen reported by mutter isn't always,
the first one listed by RANDR.
Use gdk_screen_* to query the monitor information and add a heuristic to prefer
LVDS displays (similar like in done for gnome-panel) to prefer the laptop's
internal screen over external displays.
https://bugzilla.gnome.org/show_bug.cgi?id=608647
Ok, there admittedly wasn't a strong rationale for having it in a
later, and by performing this immediately we reduce race conditions
for our focus_app versus startup_notification handling.
https://bugzilla.gnome.org/show_bug.cgi?id=612833
Paint/pick all children, regardless of whether or not they lie within
the content_box. The previous behavior was that a child that was 99%
outside the box would be fully visible, but a child that was 100%
outside the box would be fully hidden. This is somewhat odd, and
doesn't match the behavior of the other St container classes, and at
any rate, the use of clutter_actor_get_allocation_box() for this
optimization was incorrect since it doesn't take into account
transformations (anchor point, rotation, etc) that might cause the
child to be drawn within the content_box anyway.
(For scrolled StBoxLayouts, drawing is still clipped to the
content_box, as before, but will now properly take transformations
into account as well.)
https://bugzilla.gnome.org/show_bug.cgi?id=614047
When the window's top-left corner is the same as the monitor's top-left
corner the check in shell_global_get_focus_monitor fails to detect that.
Use <= rather than < to fix that.
https://bugzilla.gnome.org/show_bug.cgi?id=613944
The idea behind this move is that we have a lot more control over
rendering if StWidget isn't a big pile of actors, and things are
more efficient.
https://bugzilla.gnome.org/show_bug.cgi?id=607500
The hover rewrite added a freeze/thaw_notify to
st_clickable_leave_event() (to match the one already in
st_clickable_enter_event()), which broke code in two places that
assumed "pressed" would still be TRUE when "hover" changed to FALSE.
Fix that by exposing the "held" property as well.
A pointer does not equal to an int on x86_64
(which results into pointer - pointer not being an int either),
fix that by casting the resulting value to an int.
Breakage was introduced by commit 909b5ec43c
If track-hover is set, update the hover property automatically, and
the "hover" pseudo class to match, as StClickable used to do. (Remove
the corresponding code in StClickable). Tweak the tooltip handling to
use track-hover, which also makes it slightly more reliable in the
presence of reactive children, etc.
Since style_class and pseudo_class are space-separated lists of names,
add new methods to add and remove individual names rather than just
re-setting the entire name.
Update existing code to use the new pseudo-class methods where
appropriate. In some cases, this may result in actors having multiple
pseudoclasses where previously they only had one at a time, but there
don't seem to be any visible differences.
(There are some places that could usefully use the new style_class
methods as well, but this patch doesn't change them.)
Also, update test-theme.c to test the new methods.
https://bugzilla.gnome.org/show_bug.cgi?id=604943
The way we were loading data into a CoglTexture, then pulling it out
and manipulating it on the CPU, then loading it back into a texture
was a bit lame.
Clean things up a bit here by loading directly into the CPU, doing
the fading, then creating a texture.
Also cache the faded data in StTextureCache.
https://bugzilla.gnome.org/show_bug.cgi?id=612759
If we don't do this, then boolean/int/list keys will seem to sort of
work (but defaulting to false/0/[] instead of the correct schema
defaults), but string keys will return null, which will usually cause
exceptions or crashes.
https://bugzilla.gnome.org/show_bug.cgi?id=611214
The relationship between adjustments and scrollbars and
scrollable widgets was much more complex than it needed to be.
StScrollView: Have the scroll view own a pair of adjustments,
set them on the child on add(), remove unnecessary
change notification signal connections.
StBoxLayout: Remove auto-create of adjustments, just take the
adjustments from the scrollbars and set them on the scrollable
child. Notify for hadjustment/vadjustment properties.
StScrollBar: Notify adjustment property.
StScrollable: Document how adjustment setting works.
https://bugzilla.gnome.org/show_bug.cgi?id=611740
* Add missing chain-up for dispose and finalize methods
* ShellGenericContainer needs to destroy its children in dispose()
* Fix variable naming and excess casts in st_label_dispose()
https://bugzilla.gnome.org/show_bug.cgi?id=612511
StScrollable: Document how to set adjustments
StBoxLayout: Make sure that we always have upper >= lower + page_size,
so that clamping works properly. Set the page_increment to be slightly
less than the page_size so there is some overlap, as is customary.
StScrollView: Remove unnecessary fabs() calls, rewrite expressions
for additional clarity.
https://bugzilla.gnome.org/show_bug.cgi?id=611740
- Fix existing typos and spacing problems
- Get preferred height, not current height, of shadows
- Let shadows overflow don't clamp them when we have too little space
- Remove a now-unecessary stray MAX()
- Fix up scrollview visibility for the pathological case of no child
- Disconnect from adjustments on remove()
- Don't unset the adjustments on the child on remove(), since they
already existed or were autocreated on add()
(We should what we are doing and set the adjustments of the
scrollbars on the child rather than setting the adjustments of
the child, so we match GTK+'s scrolllable interface, but this
at least makes it consistent instead of a weird mix.)
https://bugzilla.gnome.org/show_bug.cgi?id=611740
StScrollable: Document how size negotation now works between the
parent and scrollable child.
StBoxLayout: Adapt to the new contract for how size negotiation
works; in particular, handle being allocated less than the
minimum size when scrolled and treat the minimum size as the
size of the scrolled area in instead of the natural size.
StScrollView: Substantially rewrite with fixes including:
- Implement new size negotation contract; this allows us
to determine scrollbar visibility without having to
connect to the adjustment.
- Implement all ALWAYS along with the existing NEVER/AUTO
- When hiding and showing scrollbars and shadows, don't
hide and show widgets, just turn on and off including them
in pick and paint. This avoids queueing relayouts.
- Cleanups for the code for connecting to adjustments,
for changing policy, and for turning on and off shadows.
scroll-view-sizing.js: New test case for StScrollView, allowing
resizing the scroll view interactively, changing the scrollbar
policies and turning shadows on and off.
https://bugzilla.gnome.org/show_bug.cgi?id=611740
When an StScrollView is allocated, allocating the child would
cause the adjustment values to change, which would result in
the scrollbars queueing a relayout, which isn't allowed during
allocation.
To avoid this, instead of queueing a relayout when the adjustment
changes:
- When we have a valid allocation already, just go ahead
and reallocate the children.
- Otherwise do nothing immediately and wait until we get allocated
Because the 'needs_allocation' flag in ClutterActor isn't exposed,
this requires some slightly ugly code to shadow that state locally.
https://bugzilla.gnome.org/show_bug.cgi?id=611944
Having StDrawingArea use ClutterCairoTexture causes circularity
problems with sizing - StDrawingArea wants to use its allocation for
the size of the texture, but ClutterTexture wants to use the size of
the texture to determine the requited size.
Avoid this by making StDrawingArea directly use Cairo and CoglTexture;
while doing this, the API is changed a bit for simplicity and to
match our use case:
- Instead of clutter_cairo_texture_create(), we have
st_drawing_area_get_context() to retrieve an already created
context. This can only be called in the ::repaint signal.
- The ::redraw signal is changed to ::repaint so we can have
st_drawing_area_queue_repaint() that doesn't collide with
clutter_actor_queue_redraw()
- ::repaint is now emitted lazily when painting the actor rather
than synchronously at various different points.
https://bugzilla.gnome.org/show_bug.cgi?id=611750
- Specify a minimum version of clutter-1.2.0
- Switch clutter branch in the moduleset to master
- Replace deprecated cogl_texture/material_unref() with
cogl_handle_unref()
- Use cogl_clip_push_rectangle() rather than cogl_clip_push()
- Replace cogl_check_extension() with strstr - should be
accurate enough.
https://bugzilla.gnome.org/show_bug.cgi?id=610679
It's wrong to do anything that requires looking up a widget's style
before you add the widget to the stage, since its final style may
depend on properties inherited from its parents.
st_widget_get_theme_node() used to emit a warning in this case, but
many would-be contributors apparently didn't notice. Help them out.
https://bugzilla.gnome.org/show_bug.cgi?id=610279
StButton animated the background for button transitions; since these aren't
presently part of the shell design, simply remove them. We can readd
these in the future.
StTooltip should probably have :vertical and :horizontal pseudo classes
to make the arrow work but it should still function.
https://bugzilla.gnome.org/show_bug.cgi?id=607500
Rather than having ShellTextureCache know about the type of each
item it's caching, this lays the foundation for simply caching
arbitrary string -> CoglHandle.
https://bugzilla.gnome.org/show_bug.cgi?id=607500
For StWidget we want the ability to load raw CoglHandle references
rather than having a big pile of actors backing StWidget.
Several bits of StWidget expect to be able to synchronously load
textures; this is a crutch to support that until we can cleanly
make this asynchronous.
https://bugzilla.gnome.org/show_bug.cgi?id=607500
Brute force merge these two by essentially replacing St.TextureCache
with a (renamed) Shell.TextureCache.
One function was added for convenience, namely "st_texture_cache_load_file_simple".
St.TextureCache had a function to load a texture from a filename, and it
returned NULL on error but only half the callers actually checked this. This
function is better.
https://bugzilla.gnome.org/show_bug.cgi?id=607500
StScrollBar: Be robust against being disposed multiple times,
which can happen, and in fact, normally happens when destroying
the parent.
StScrollView: Implement remove() for the hscroll and vscroll members,
and just destroy them in dispose() and let them be removed.
unparent the shadows, instead of just unref'ing them directly.
https://bugzilla.gnome.org/show_bug.cgi?id=611203
When starting oocalc or ooimpress from oowriter's menu get_app_for_window,
fails to recognize the newly started up as such, which result into the appMenu
still showing "Openoffice.org Writer" as app name.
Fix this by trying to window itself first before using the group for finding the app.
https://bugzilla.gnome.org/show_bug.cgi?id=611288
Since shell-slicer.h includes st.h, and that includes
<st/*.h>, we need to add -I $(srcdir) to the command line when
running g-ir-scanner to generate Shell-1.0.gir.
For srcdir != builddir, in the builddir, the st/ subdirectory doesn't
exist, so we can't generate st.h there. Just switch to building st.h
directly in the current directory. (The other option would be to
create a st/ subdirectory in the builddir if necessary; might be a
little cleaner, but this works for now and gets things distchecking.)
Add a 'vshadow' property to StScrollView, which, when turned on,
overlays gradient shadows on the top and bottom of the StScrollView.
Turn this on for the StScrollView used for the app browser.
https://bugzilla.gnome.org/show_bug.cgi?id=609604
The type we don't currently handle is _ALWAYS, my original use
of g_return_if_fail was wrong.
Also the default should be AUTOMATIC in the properties, not ALWAYS.
Finally the test case was still using the incorrect St.ScrollPolicy.
https://bugzilla.gnome.org/show_bug.cgi?id=609015
Previously we were hacking out the vertical scrollbar, this patch
allow us to sanely say in which directions we want to scroll.
Note this patch intentionally changes St to depend on GTK+ in the
API. I believe this is a correct long term change where we should
view St as co-evolving with GTK+ rather than replacing or paralleling.
https://bugzilla.gnome.org/show_bug.cgi?id=609015
st_theme_node_adjust_preferred_width/height now limit the content area
of an actor to the max, if given. (The requested width/height may be
larger to make room for borders, etc.)
https://bugzilla.gnome.org/show_bug.cgi?id=606755
In StBin, StBoxLayout, and StTable, if a child has a potential
allocation that is larger than its preferred size, we give it its
preferred size instead. However, the corresponding
get_preferred_height/width methods were not making the same
assumption, which meant that if we had more width than the widget
wanted, we would allocate it its preferred width, but with the height
that corresponded to the larger width.
Fix this by defining new helpers _st_actor_get_preferred_width() and
_st_actor_get_preferred_height() and using them everywhere. Also, make
StBin and StTable use _st_allocate_fill() rather than having
nearly-identical duplicate copies of the code.
https://bugzilla.gnome.org/show_bug.cgi?id=609848
The forward/backward steppers are always allocated a square region at the
scroll bar's ends. Change the allocation to be based on the steppers' size
requests instead.
https://bugzilla.gnome.org/show_bug.cgi?id=609401
Make the framerate, file extension and gstreamer pipeline used by the
screencast recorder configureable using gconf.
This patch does not change the defaults, it justs provides a way for
the user to override them.
https://bugzilla.gnome.org/show_bug.cgi?id=608995
Fixes drawing of the overview in the case where there is no root pixmap
(eg, --xephyr mode).
(And the workaround for not drawing an unintialized ClutterTexture can
be removed now, since the texture will always have been initialized.)
https://bugzilla.gnome.org/show_bug.cgi?id=609339
To comply with C89, structure initializers should have
only constant values.
(Not a thorough check for this throughout the codebase, just
StWidget is fixed up in this commit.)
https://bugzilla.gnome.org/show_bug.cgi?id=608746
As of 2a78adc5e3, gobject no longer allows for setting construct-only
properties via g_object_set() during object initialization; set the
properties in a custom constructor instead.
https://bugzilla.gnome.org/show_bug.cgi?id=608802
In some situations we might need to look up an application from
a process identifier, such as the notification system where we
will determine application from the message sender.
By calling clutter_actor_get_allocation() in st_widget_recompute_style()
to determine whether to redraw gradients, we triggered a complete
reallocation of the stage for each gradient.
As gradients are processed in st_widget_real_style_changed() anyway, the
additional checks in st_widget_recompute_style() are redundant and can
be removed altogether.
https://bugzilla.gnome.org/show_bug.cgi?id=608847
On style changes from gradient to solid backgrounds, the new background
must be drawn unconditionally, not depending on whether old and new
background color differ.
https://bugzilla.gnome.org/show_bug.cgi?id=608914
The current way we keep track of skip-paint children is more
difficult than it needs to be, and can lead to subtle bugs.
(For one, the skip-paint state of a child is remembered
when it is removed then added back again, which is completely
unexpected.)
Instead of using weak references to track children, just remove
items from the skip-paint list by overriding the remove() virtual
function of the ClutterContainer interface.
The 'skip_paint' hash table is then destroyed in finalize rather than
dispose since it doesn't hold references to memory any more but just
passively tracks an attribute of the children that are currently in
the container.
https://bugzilla.gnome.org/show_bug.cgi?id=608848
When moving a widget with a gradient, its allocation changes
continuously, resulting in constant redraws.
Checking for actual size changes before the operation avoids
unnecessary redraws.
https://bugzilla.gnome.org/show_bug.cgi?id=608715
The screen recording wasn't working because of two bad interactions
with Cogl:
- Buffered primitives weren't being flushed out
- Cogl changes the pixel store values away from their default values
Thanks for Jon Nettleton for tracking down the source of the
problems with the recorder.
https://bugzilla.gnome.org/show_bug.cgi?id=598390
StButton has an internal animation for the border-image actor, then
it connects to the "completed" signal passing itself as data. However,
if the button is destroyed, nothing prevents the animation's completed
signal from then causing a reference to freed memory.
Holding a reference to the button is the most straightforward fix here.
https://bugzilla.gnome.org/show_bug.cgi?id=607825
Added signal 'screen-size-changed' to ShellGlobal.
Connect to this signal in main.js and run the _relayout() method.
If Overview or calendar are visible when this signal emit, they will be hiding.
https://bugzilla.gnome.org/show_bug.cgi?id=584526
Replace ngettext with dngettext and set the correct translation
domain, so gnome-shell's domain is searched for translations
rather than mutter's.
https://bugzilla.gnome.org/show_bug.cgi?id=597882
Some theme authors have stated interest in radial gradient backgrounds.
The w3c has some draft:
http://dev.w3.org/csswg/css3-images/#radial-gradients
As this is rather complex, we add only some very basic support, which
extends our syntax for linear gradients:
background-gradient-direction: [vertical|horizontal|radial]
Gradients are centered circles, whose size is determined by the closest
side.
https://bugzilla.gnome.org/show_bug.cgi?id=604945
Some changes to the way we handle CSS gradients:
* draw without padding, thus interpreting gradients as part of the
background rather than as content
* clip to (rounded) border area
* draw the border along the gradient instead of trying to align the
gradient layer with the background/border layer
* use the border_image actor instead of the background_image one
https://bugzilla.gnome.org/show_bug.cgi?id=606257
We were holding on to UsageData structure pointers in previously_running,
but those can be purged by the unused app cleanup. Instead just hold
onto dup'd copies of the application id strings.
Add support for a new -st-shadow property, which is based loosely
on the CSS3 box-shadow property:
http://www.css3.info/preview/box-shadow/
It defers from the specification as follows:
* no multiple shadows
* the optional color argument may be placed anywhere
* the shape is not determined by the widget's bounding box,
but by the background-image property
https://bugzilla.gnome.org/show_bug.cgi?id=603691
Change "./src/gnome-shell --create-extension" to use "gnome-open"
when opening the newly created .js file, so that it is launched
with the user's preferred text editor, instead of hardcoding gedit.
In case of duplicate infos structures with the same id, the
info structures we get from looking up the id in app_id_to_info
aren't necessarily the same as those we used to match, so we
can't rely on matching to implicitly initialize info->casefolded_name.
Since the name collation key isn't used in matching results,
just in sorting, init it on-demand in the sorting which is also more
efficient.
Move CSS handling of StLabel and StButton for their underlying
ClutterText objects into st_private, and implement support for
the underline and strikethrough St text-decoration properties.
Overline isn't implemented for lack of a corresponding Pango
attribute, and blink, well...
https://bugzilla.gnome.org/show_bug.cgi?id=599661
Consumer documentation will live at http://live.gnome.org/GnomeShell/Extensions
In terms of implementation; basically we load extensions from the well-known
directories. Add a GConf key to disable extensions by uuid. There is a new
option --create-extension for the gnome-shell script which takes a bit of
interactive input, sets up some sample files, and launches gedit.
No extensions UI in this patch; that will come later.
https://bugzilla.gnome.org/show_bug.cgi?id=599661
The high level goal is to separate the concern of searching for
things with display of those things; for example in newer mockups,
applications are displayed exactly the same as they look in the
AppWell.
Another goal was optimizing for speed; for example,
application search was pushed mostly down into C, and we avoid
lowercasing and normalizing every item over and over.
https://bugzilla.gnome.org/show_bug.cgi?id=603523
Ideally we'd be able to override _paint, but given that we can't
at the moment, this method gives a way to implement containers
which don't happen to paint all of their children.
https://bugzilla.gnome.org/show_bug.cgi?id=603522
The synchronous close causes us to block in fsync() which has extremely
poor interactivity implications on ext3.
Also, the 5 second timeout was an accidental commit from debugging, 5
minutes is fine.
https://bugzilla.gnome.org/show_bug.cgi?id=602456
StClickable replaces ShellButtonBox. Reduce the number of
button-like things by deleting button.js.
To do so, add CSS style for the actitivies button.
https://bugzilla.gnome.org/show_bug.cgi?id=602131
Now a StBin, and add hover/active style properties. Also, add the
event to the CLICKED signal. Otherwise a straightforward namespace
transformation.
https://bugzilla.gnome.org/show_bug.cgi?id=602131
It's nicer to have ShellDrawingArea as a St widget so it can
participate more cleanly in CSS styling, such as queuing a redraw
automatically on style changes, and allowing subclasses to use
CSS styling.
https://bugzilla.gnome.org/show_bug.cgi?id=602131
Rather than having gradients be individually implemented by higher
level JS widgets, move basic gradient functionality into StWidget.
There is prior art in WebKit for CSS gradients:
http://webkit.org/blog/175/introducing-css-gradients/
However, implementing this would be quite a lot of work; all we
need in the Shell design at the moment is basic horizontal/vertical
linear gradients. So, the syntax now supported is:
background-gradient-type: [vertical|horizontal]
background-gradient-start: color;
background-gradient-end: color;
https://bugzilla.gnome.org/show_bug.cgi?id=602131
An earlier commit was overzealous in removing (out) annotations;
introspection supports (out) for integral types just fine, we
only need to remove them for (out) types where the caller needs
to allocate a boxed type.
https://bugzilla.gnome.org/show_bug.cgi?id=602131
In addition to the Makefile changes, we also change uid_t to gulong in
the public API (which matches how it was already represented in the
gobject properties).
https://bugzilla.gnome.org/show_bug.cgi?id=601458
In a variety of places we're using boxes as data-modeling displays,
and in doing so we often want to either remove the children or
explictly destroy them.
Now ideally Gjs would support callbacks, and this would make using
the for_each functions possible, but even then these functions
are more efficient and shorter to type, at least.
https://bugzilla.gnome.org/show_bug.cgi?id=600734
If the space we're allocated is too small for our border + padding
constraints, don't give negative allocations to callers. Squash
to zero.
It isn't really useful for callers to get negative content sizes,
and certainly breaks most allocation code.
https://bugzilla.gnome.org/show_bug.cgi?id=600734
StTheme CSS supports different border widths for different sides. Implement
it for StWidget by drawing the border internally. However, we don't support
a nonzero corner-radius with nonuniform borders.
https://bugzilla.gnome.org/show_bug.cgi?id=599442
Previously shell_app_remove_window assumed that it was being
passed a window in its list; rather than having callers check
whether a window is interesting and only if so removing it
from the app, just ignore removal of windows we aren't interested
in, like how we ignore addition of windows we already have.
https://bugzilla.gnome.org/show_bug.cgi?id=598502
The behavior in respect to borders matches CSS - the properties set the size of
the content exclusive of the borders (CSS3 box-sizing property - not implemented
here - changes this).
min-width/min-height correspond very closely to the CSS meanings.
width/height are a little different from the CSS meanings - the CSS meaning is
"exactly this size unless overridden by min/max-width/height" - but within the
realm of our layout algorithm, making them control natural size is pretty
close.
This way we can force elements to have a fixed natural or minimum size.
https://bugzilla.gnome.org/show_bug.cgi?id=598651
For the purposes of determining which application is focused, don't
skip "uninteresting" windows. The old get_focused_window code
was used for usage tracking, but here we want reliable application
association.
Also convert a .text= to .set_text that was missed with the last
patch.
https://bugzilla.gnome.org/show_bug.cgi?id=599206
The two parts were mapping windows to applications, and
recording application usage statistics. The latter part
(now called ShellAppUsage) is much more naturally built on top of
the former (now called ShellWindowTracker).
ShellWindowTracker retains the startup-notification handling.
ShellWindowTracker also gains a focus-app property, which is
what most things in the shell UI are interested in (instead of
window focus).
ShellAppSystem moves to exporting ShellApp from more of its
public API, rather than ShellAppInfo. ShellAppSystem also
ensures that ShellApp instances are unique by holding
a hash on the ids.
ShellApp's private API is split off into a shell-app-private.h,
so shell-app.h can be included in shell-app-system.h.
Favorites handling is removed from ShellAppSystem, now inside
appFavorites.js.
Port all of the JavaScript for these changes.
https://bugzilla.gnome.org/show_bug.cgi?id=598646
The window lists were not being resorted when user-time changed, and
the app list was mistakenly "penalizing" apps for having *any*
minimized windows, rather than for having *only* minimized windows.
https://bugzilla.gnome.org/show_bug.cgi?id=598389
Previously, we had ShellAppInfo, which contains fundamental
information about an application, and methods on ShellAppMonitor
to retrieve "live" information like the window list.
AppIcon ended up being used as the "App" class which was painful
for various reasons; among them that we need to handle window
list changes, and some consumers weren't ready for that.
Clean things up a bit by introducing a new ShellApp class in C,
which currently wraps a ShellAppInfo.
AppIcon then is more like the display actor for a ShellApp. Notably,
the ".windows" property moves out of it. The altTab code which
won't handle dynamic changes instead is changed to maintain a
cached version.
ShellAppMonitor gains some more methods related to ShellApp now.
In the future, we might consider changing ShellApp to be a GInterface,
which could be implemented by ShellDesktopFileApp, ShellWindowApp.
Then we could axe ShellAppInfo from the "public" API and it would
return to being an internal loss mitigation layer for GMenu.
https://bugzilla.gnome.org/show_bug.cgi?id=598227
ClutterGroup calls _destroy, but most of St was just calling _unparent.
This caused problems because the DESTROY signal was not emitted
for child elements after destroying a toplevel. Also, in a GC'd
binding it would cause unpredictable lifetime of children.
Some St widgets simply didn't have _dispose at all; implement it.
Note because of the usage of the background_image in StButton,
we can't cleanly destroy it inside the StWidget.
https://bugzilla.gnome.org/show_bug.cgi?id=597845
When we get a ClutterModifierType from Clutter, it might contain
bits not in the enumeration. See bug 59771 for a similar problem
with GdkModifierType.
Add a wrapper Shell.get_event_state() around clutter_event_get_state()
to mask these bits out and only return approved bits.
https://bugzilla.gnome.org/show_bug.cgi?id=597735
gdk_display_get_pointer() sometimes returns values for the mask that
aren't part of the GdkModifierType enumeration, which gjs doesn't like
(bug 597292). Work around that by adding a C wrapper that strips out
the extra flags.
https://bugzilla.gnome.org/show_bug.cgi?id=597559
Add StIMText, which is a drop-in replacement for ClutterIMText but
uses GtkIMContext instead of ClutterIMContext.
StIMText doesn't have preedit support (would need ClutterText
changes), so isn't going to be useful for complicated input methods,
but is good enough to get dead keys and similar working.
entry.js: Simple test case of StEntry
gnome-shell.modules: Remove clutter-imcontext module
https://bugzilla.gnome.org/show_bug.cgi?id=597471
To work around a problem where libcroco < 0.6.2 can't handle
functions starting with 'r' or 'u', preconvert 'rgba' to 'RGBA'
when parsing stylesheets and then check for rgba()
case-insensitively.
(libcroco is uniformly case-sensitive, though the CSS spec requires
that ASCII should be handled case-insensitively.)
https://bugzilla.gnome.org/show_bug.cgi?id=597054
Before we hardcoded popdowns to only button 1 before. But we need
to actually pop down on the release of the activating button.
(Once the button is released, if the we don't pop-down the menu,
then subsequently we let the user use any button.)
https://bugzilla.gnome.org/show_bug.cgi?id=596371
round() is a C99 addition, so causes portability problems:
different C library versions require different #defines to
enable it. So simply avoid using it.
Property enumeration names should correspond exactly to the property names;
in particular the ACTIVE vs :checked disparity was confusing reading the
code.
http://bugzilla.moblin.org/show_bug.cgi?id=6504
Convert the StTable code from StStylable to StThemeNode. The
:row-spacing and :col-spacing GObject properties are converted
into spacing-rows and spacing-columns style properties.
A new interactive test is added for StTable.
https://bugzilla.gnome.org/show_bug.cgi?id=596811
Remove the StTable specific methods to add actors:
st_table_add_actor()
st_table_add_actor_with_properties()
Since they shadow the generic ClutterContainer add_actor() method,
and patch in our add() convenience function as we do for
StBoxLayout.
https://bugzilla.gnome.org/show_bug.cgi?id=596811
Remove the StBoxLayout:spacing GObject property, and instead make
BoxLayout look up the spacing from the CSS style. This makes it
consistent with padding and will allow the use of units. (The
removal of the GObject property entirely instead of making it an
override is consistent with how we handle color, font, padding, etc.)
https://bugzilla.gnome.org/show_bug.cgi?id=596803
Add clutter-text properties to allow getting access to the underlying
ClutterText actor. This corresponds to the get_clutter_text() methods.
The PROP_LABEL and PROP_ENTRY enum values are renamed to PROP_TEXT to
match the names of the properties that they correspond to, and the
properties of StEntry are reordered into alphabetical order.
Based on a patch from Colin Walters
https://bugzilla.gnome.org/show_bug.cgi?id=591245http://bugzilla.moblin.org/show_bug.cgi?id=6313
StBoxLayout: Make consistent that the area scrolled and clipped
to is the content area (excluding borders and padding.) Translate
back appropriately when chaining up so that the parent background
is drawn at the right place and picking on the box (if it's reactive)
picks at the right place on the screen.
clip-to-allocation is removed from StScrollView since it's just
not right - if the child has any non-moving elements, like headers or
borders, it will need to set a narrower clip. And even if the entire
child scrolls, we want to clip to an arrow that excludes the scrollbars.
https://bugzilla.gnome.org/show_bug.cgi?id=595997
When a StBoxLayout is allocated a size less than its natural size,
think "shrink" needs to be divided among the children that have
a smaller minimum size than natural size.
This is done by preferentially shrinking the children that are most
expanded from their minimum size and then increasing that set of
children until we've found enough total shrink.
A new method is used of allocating children at integral sizes - instead
of rounding the per-child extra amount to an integer (which causes
cumulative round-off errors), compute the position as we go along in
floats and round individually for each child widget.
Extend the box-layout test to include of a test of a box being set
to various widths, starting quite narrow.
http://bugzilla.moblin.org/show_bug.cgi?id=6311https://bugzilla.gnome.org/show_bug.cgi?id=595995
If the actor isn't in a stage, then setting up the adjustment
based on the actor's size (which we can't compute) and the
size of the default stage (which isn't relevant), doesn't make
sense. Just use arbitrary default values.
The adjustments will be updated to reasonable values when first
the box is first allocated.
It's not entirely clear to me why we ever want to compute the
adjustment settings this way; perhaps we should always use
default values.
http://bugzilla.moblin.org/show_bug.cgi?id=6307https://bugzilla.gnome.org/show_bug.cgi?id=595996
The CSS specification says that the background extends to the
edge of the border (settable in CSS3 with border-clip), make
BigRectangle match this by computing an "effective border color"
as 'border OVER background'.
(If we don't want this behavior - e.g., to be able to use the
transparent borders as margins, then alternatively transparent
border handling would have to be fixed in st-widget.c, since
prior to this transparent and translucent borders were handled
differently.)
https://bugzilla.gnome.org/show_bug.cgi?id=595993
The current CSS3 border-image is close to a superset of what we were
doing for -hippo-background-image. Woot! rename StThemeImage to
StBorderImage and change parsing to look for:
border-image: <url> <number>...
Rather than
-st-background-image: <url> <length>...
percentanges for the border sizes are not currently supported, neither
are the keywords for handling of the middle part. We always do 'stretch'
for now.
https://bugzilla.gnome.org/show_bug.cgi?id=595990
Use BigRectangle to draw the border and background if there's
a border width or border radius and no border image. (Only
uniform borders are supported for now with some deviations
from the CSS model noted in the comments.)
The background color and image parameters are removed from
StWidget's draw_background() method since they were not used
for StButton (the only current user) and the encapsulation
break that they presented caused some minor problems.
Add a test case for borders, and also use borders to style
the buttons in the 'inline-style' test case.
https://bugzilla.gnome.org/show_bug.cgi?id=595993
Rather than repeating the computation of borders in many different
widget subclasses, add helper functions:
st_theme_node_adjust_for_height()
st_theme_node_adjust_preferred_width()
st_theme_node_adjust_for_width()
st_theme_node_adjust_preferred_height()
st_theme_node_get_content_box()
That are used in get_preferred_width()/get_preferred_height() and
allocate() methods to consistently apply the necessary adjustments.
This allows removing the StPadding type.
Queueing a relayout when the borders/padding change is moved from
st_widget_real_style_changed() to the invoking code to allow access
to the old StThemeNode for comparison. (Should this be added as
a parameter to the signal?)
Borders are included in the geometry adjustments, but borders
are not yet drawn.
https://bugzilla.gnome.org/show_bug.cgi?id=595993
Add support for parsing and caching the border-radius property.
Different radii for the 4 corners are supported; elliptical corners
are not supported.
https://bugzilla.gnome.org/show_bug.cgi?id=595993
Add support for passing an inline-style string when creating a
StThemeNode.
Hook this up to a new 'style' property of StWidget.
Add a test case that demonstrates using this to update font sizes
on the fly.
https://bugzilla.gnome.org/show_bug.cgi?id=595991
ShellTheme replaces both StStyle and ccss_stylesheet_t.
The interface StStylable is replaced by usage of ShellThemeNode.
A concrete node class allows some significant optimizations of property
inheritance that would have been much more difficult to achieve with
the highly abstract pair of StStylable and ccss_node_t.
Some operations that were previously on StStylable (like the
::style-changed signal) are directly on NtkWidget.
Custom properties are no longer registered as param-specs; instead you
call directly into shell theme node to look up a length or color:
shell_theme_node_get_length (theme_node, "border-spacing", FALSE, &spacing);
The dependency on libccss is dropped, while preserving all existing
functionality and adding proper parsing and inheritance of font properties
and proper inheritance for the 'color' property.
Some more javascript tests for CSS functionality are added; workarounds for
a CSS bug where *.some-class was needed instead of .some-class are removed.
https://bugzilla.gnome.org/show_bug.cgi?id=595990
To each .c and .h file, add:
/* -*- mode: C; c-file-style: "gnu"; indent-tabs-mode: nil; -*- */
'gnu' is the default anyways for Emacs, but indent-tabs-mode is not,
so this sets things up to correspond to the policy of no-tabs.
http://bugzilla.moblin.org/show_bug.cgi?id=6467
Import:
HippoCanvasTheme => StTheme
HippoCanvasThemeImage => StThemeImage
HippoCanvasStyle => StThemeNode
StThemeContext is a new class managing the theme for a stage and
global properties like resolution.
test-theme.c is a newly written test program to do verification of the
style matching and property handling rules.
Various changes are made in the import:
- Comprehensive reindentation
- guint32 pixels replaced with ClutterColor
- General pseudo-class support added
- Old-fashioned (non-bordered) background image support added, though
with no support for repeat, etc.
- Bug fixes for problems revealed by test program
https://bugzilla.gnome.org/show_bug.cgi?id=595990
Remove several stale C files that we are no longer using; now
that we have a distcheck hook to catch non-distributed files, these
would otherwise prevent distchecking.
https://bugzilla.gnome.org/show_bug.cgi?id=595988
Add GObject Introspection annotations to methods where needed, in
particular adding (transfer none) to return values that don't transfer
ownership.
st_texture_cache_get_actor() and st_texture_cache_get_texture()
are annotated as (transfer none) since they return a newly
created *floating* texture.
https://bugzilla.gnome.org/show_bug.cgi?id=591245
Import the core MxWidget/MxBin and their dependencies; we use the
namespace "St" (Shell Toolkit) because it is the same length as Mx
so enabling easy sharing of code, but makes it clear that this is
a friendly fork and not a literal import.
Based on a patch by Colin Walters <walters@verbum.org>
https://bugzilla.gnome.org/show_bug.cgi?id=591245
Fix panel, app switcher, and looking glass to limit themselves to the
primary monitor, and run dialog to limit itself to the monitor
containing the currently-focused window.
The overview is also limited to the primary monitor now (with the
other monitors being blacked out), although the workspaces within the
overview are shaped like the full "screen" (the bounding box of all
monitors). To be fixed later.
https://bugzilla.gnome.org/show_bug.cgi?id=593060
Before, if the texture cache received a request to load say
the themed icon for an application multiple times (as could happen
since we have multiple application displays), it would often create
a thread for each one and in fact, load the pixbuf multiple times.
Avoid this by keeping track of outstanding requests.
https://bugzilla.gnome.org/show_bug.cgi?id=596121
This fixes a regression where we weren't using the correct event
timestamps, because for both of these we were sending an XClientMessage
to ourself.
https://bugzilla.gnome.org/show_bug.cgi?id=596262
A previous patch fixed a leak when loading items which shouldn't
be cached, but we also had a leak if two requests for the same
item were outstanding. In that case we load the pixbuf twice,
but should discard subsequent loads when we notice we've already
cached it.
https://bugzilla.gnome.org/show_bug.cgi?id=595321
The menu is needed by the app switcher as well as the overview, so
make it slightly more generic and move the code to appIcon. Also add
support for drawing the menu either to the right of or below the icon.
https://bugzilla.gnome.org/show_bug.cgi?id=590563
When the user click+hold+release over the icon, the effect we want
is for the menu to stick around.
Also, allow the user to mouse over the actual windows and select
them directly. If the user mouses over a window, reflect that in
the menu.
https://bugzilla.gnome.org/show_bug.cgi?id=594699
Callers will generally expect _popup and _popdown to be a no-op if
the menu is already in that state; make it so.
Also change the 'popdown' signal to be 'cancelled'; this is
clearer and allows us to avoid having activate also call popdown.
https://bugzilla.gnome.org/show_bug.cgi?id=594699
We have compatibility code which detects from the window title what
an application is. However, the code didn't handle the case where
we discovered by title, but didn't have the expected .desktop file
installed.
When the user click+hold+release over the icon, the effect we want
is for the menu to stick around.
Also, allow the user to mouse over the actual windows and select
them directly. If the user mouses over a window, reflect that in
the menu.
If start_shell() threw an exception before, we'd overwrite it with
an exception in the finally() clause. Handle this and just print a message
and let the exception propagate.
Before Clutter gained accessors for event information, we had
shell_global_ functions. Now that Clutter has them, use them and
delete the ShellGlobal code.
http://bugzilla.gnome.org/show_bug.cgi?id=594561
When we have multiple windows for an application, implement the following
behavior:
* On click + immediate release, go to the most recently used
* On click, hold for 0.6s, pop up a menu with windows, filtering
the window list to just those windows.
Mouse over on the window list highlights the moused-over window.
Implement this by splitting well item into InactiveWellItem
and RunningWellItem, sharing a base class BaseWellItem.
The application menu code wants to do a popup after a given timeout
while holding. We can implement that by adding a function to
manually break the grab held by the button box.
Freeze+thaw around the hover and pressed property notification on leave
since handlers may want to depend on the pressed state on a hover
transition.
The windows we considered for both the app monitor and the overview
workspaces were the same, but the code was duplicated once in C, once
in Javascript.
On OpenSolaris /usr/bin/python is 2.4; use AM_PATH_PYTHON to find
a newer Python. (The PYTHON environment variable can also be set
before running configure to override the search.)
http://bugzilla.gnome.org/show_bug.cgi?id=578196
Add .AUTOPARALELL which is my GNU-make fix for projects to specify
that the build is parallel-safe, and to automatically parallelize.
Add a missing dependency on built sources, and specify --libtool
to be safe.
Only mouse button 1 is supposed to activate button controls; other
mouse buttons should do nothing unless there is a context menu.
Checking the click count is important, since double-clicks will
otherwise look like unpaired button presses.
http://bugzilla.gnome.org/show_bug.cgi?id=593504
There's seldom a good justification for connecting to signals on
yourself rather than using the default handler slots in the class.
But in particular using the default handler slots means that
an application can connect to ::button-press-event and get in
before the default handling, to implement a button that does
something on press.
http://bugzilla.gnome.org/show_bug.cgi?id=593503
Add an 'active' property to ShellButtonBox. This allows ShellButtonBox
to be used as a "toggle button". It's up the application to connect
it to the ::activate signal; there's no default handling of this.
(It's seldom that the only time you want to toggle a toggle button
through the user interface, so you need some connection to the backend
data store in any case. Removing the default handling all-together
prevents weird interactions.)
When we have built-in styling for ShellButtonBox the 'active' state
would be one of the elements that would be affect the styling.
http://bugzilla.gnome.org/show_bug.cgi?id=593502
shell-global.[ch]: Add shell_global_display_is_grabbed() that
uses the newly added meta_display_get_grab_op() to check
for existing grabs.
shell-status-menu.[ch]: Add shell_status_menu_is_active() to
check if the menu is popped up. Check for active grabs before
popping the menu up. Use gtk_menu_popdown() rather than
gtk_widget_hide(). Remove an excess gtk_widget_show() and
some excess casts.
panel.js: Check whether the status menu is popped up after button
release, and if it's not popped up, unhighlight the button.
Reported by Nuno Donato
http://bugzilla.gnome.org/show_bug.cgi?id=593362
gnome-shell.in: Remove the code to replace gnome-panel by attaching
to it with GDB; this was always problematical (required gdb, debug
symbols, finding the pid of gnome-panel, etc.)
gnome-shell-build-setup.sh: Require 2.26 to be in place before building
the shell; remove gdb from the list of required packages.
http://bugzilla.gnome.org/show_bug.cgi?id=593325
This is a Box subclass which adds several signals useful for implementing
"button like" behavior, such as hover and pressed states, as well as
click activation on release.
Instead of starting Xephyr automatically, require --xephyr to be
passed explicitly.
This makes the operation easier to understand and has the benefit
of allowing running in Xephyr mode when some other window manager
(like gnome-shell!) is running. We also want to emphasize that
Xephyr is a development tool, and not a good preview of the
user-interface.
http://bugzilla.gnome.org/show_bug.cgi?id=592881
In both, using our allocation directly for the child is wrong; we
should create a new allocation that's our width and height.
In ShellDrawingArea, also need to chain up to parent.
