Jonas Dreßler [Sat, 24 Oct 2020 18:25:49 +0000 (20:25 +0200)]
Make sure to allocate all children in allocate vfuncs
Clutter expects actors overriding the allocate vfunc to allocate all
mapped children of the actor, otherwise bad things happen.
So make sure we actually allocate all our visible children in our custom
allocation functions, and since we don't want to give them a real
allocation, just pass them an empty ClutterActorBox.
It would be nice if we had a way to hide children during the allocation
process where no relayout is queued like gtk allows with
gtk_widget_set_child_visible(), then we could avoid those weird empty
ClutterActorBoxes.
Jonas Dreßler [Sat, 24 Oct 2020 20:24:35 +0000 (22:24 +0200)]
windowManager: Check whether metaWindow is NULL after destroy animation
When logging out or terminating gnome-shell, mutter will unmanage all
open windows, triggering the window-close animation in gnome-shell and
very quickly after that emitting "kill-window-effects". That means we'll
call _destroyWindowDone() to cleanup our animation data, but at this
point the MetaWindow of the window is already gone, so we get an error
that get_meta_window() returns NULL.
Fix that by checking whether get_meta_window() returned NULL and if it
did, don't access the window.
Jordan Petridis [Fri, 20 Nov 2020 13:54:13 +0000 (14:54 +0100)]
ci: Replace only/except with rules
only/except keywords where deperecated in favor of rules.
Since we started using GNOME/gnome-shell!1492 it introduced
a second pipeline being run for each commit.
Detached pipelines are the only way to access CI_MERGE_REQUEST_*
variables, and if we disable normal pipelines you will need to
create wip/spam MRs in order to run the tests.
This reworked rules makes it so, the normal pipeline needs manual
interaction to be started, and the detached/MR pipleines is always
run.
Jonas Ådahl [Fri, 20 Nov 2020 16:34:11 +0000 (17:34 +0100)]
ci: Generate junit report if project doesn't enable the registry
The image registry must be enabled for the CI to work, as the container
image from the upstream registry is placed (as a reference, more or
less) in the local registry.
Recently, ci-templates got a bit more helpful if this happened, and will
fail up front and generate a JUnit report with a message describing the
issue.
Jonas Ådahl [Fri, 20 Nov 2020 10:32:12 +0000 (11:32 +0100)]
ci: Bump ci-templates to include registry path fix
ci-templates had a bug where it failed to find the ready-built container
images on the upstream repository due to invalid assumptions about the
relationship between upstream repository and image registry. This
resulted in the container image to be rebuilt, wasting valuable CPU
cycles and storage space. Bump the included ci-templates template to
include the fix for this.
See https://gitlab.freedesktop.org/freedesktop/ci-templates/-/issues/22.
Jonas Ådahl [Fri, 6 Nov 2020 17:59:19 +0000 (18:59 +0100)]
ci: Use ci-templates for building image
This rebuilds the automaticaly whenever the image tag changes. Whenever
something in the image needs to change, alter the installation script
and change the tag to the current date.
This removes the -s (strict) argument from js68, as it doesn't exist in
js78.
Jonas Ådahl [Fri, 6 Nov 2020 17:07:07 +0000 (18:07 +0100)]
ci: Wait for 'check-commit-log' before continuing
The check-commit-log is quick, and to get a result early is helpful as
one can then more quickly check for failures via the report provided via
the JUnit report.
Jonas Ådahl [Fri, 6 Nov 2020 16:59:28 +0000 (17:59 +0100)]
ci/check-commit-log: Reverse the merge request URL check
Instead of making sure there is a reference to a bug or merge request,
make sure there isn't. The reason for this is that marge-bot will always
append a merge request URL in the end of the commit message.
Florian Müllner [Mon, 9 Nov 2020 19:24:34 +0000 (20:24 +0100)]
ci: Enforce non-legacy style for new code
We have been transitioning to the new coding style for a while now,
and there has been reasonable progress. So hopefully it is not too
intrusive at this point to enforce non-legacy style for all lines
that are changed in a merge request. If it turns out to be still too
annoying, we can always reconsider and turn off the additional job.
Florian Müllner [Sat, 7 Nov 2020 06:11:09 +0000 (07:11 +0100)]
ci: Use node-js script for running eslint
This is unnecessary hard in shell when compared to a proper programming
language. It becomes even easier with a node-js script, as that gives us
access to the underlying ESLint module rather than just the CLI interface.
Besides that, node-js has the added benefit that we don't need to add
more dependencies to the CI image.
If the adapt to size occurs after the grid page has been changed the
page is set to zero. This patch just keep the current page with the same
value that it has before.
appDisplay: Only show add/remove favorite when shouldShowApp() is true
`shouldShowApp()` is called in `_addFavorite`, so adding a favorite when
this isn't true won't work. Also, it seems when this is false, favorites
that do exist won't be shown anyway.
Carlos Garnacho [Wed, 28 Oct 2020 19:29:55 +0000 (20:29 +0100)]
magnifier: Avoid cursor visibility changes on startup
When initializing the shell, we create the magnifier, and (normally)
let it disabled. This still toggles cursor visibility on, which is
not right since there's other considerations to take during
initialization.
Only do this after actual changes to the magnifier state, so
initialization is left unperturbed.
It turns out, StViewport's pick implementation was broken, since
Clutter stopped reading clipping from the framebuffer. Rejoice with
bugfixing by removing code!
Jonas Dreßler [Sat, 24 Oct 2020 18:51:56 +0000 (20:51 +0200)]
windowPreview: Handle case where window already is part of layout
We sometimes add dialogs multiple times to the WindowPreview, for
example for modal dialogs we receive both the "window-added" and the
"window-entered-monitor" signal, which means we call
WindowPreview.addDialog() twice.
We handle that fine already in the WindowPreviewLayout and return NULL
in case the window already was added, so simply handle that NULL return
value and bail out of WindowPreview.addDialog() in this case.
Harry Míchal [Sat, 24 Oct 2020 15:07:10 +0000 (15:07 +0000)]
mpris: Switch text in title and body
Most media players have a media player that shows the title of a song
and artists in the song. In those media players the title is
highlighted (bold text) and the list of artists is under the title.
Shell does the exact opposite in the player in the notification area.
Example media players: Spotify, Rhythmbox, GNOME Music
Jonas Dreßler [Fri, 23 Oct 2020 20:05:13 +0000 (22:05 +0200)]
tests/background-size: Remove ClutterActor::paint signal handler usage
The background-size interactive test includes a test-case which uses the
ClutterActor::paint signal to hijack the painting of the actor. We're
removing that signal from Clutter, so remove that signal handler and
paint the outer border unconditionally to mark the boundaries of the
widget (the outer border actually never worked to force the cairo/cogl
path). To fix the test, change it to not paint a border on the inner
container in the useCairo=false case, which should make sure the cogl
path gets used.
Also remove the useless useCairo argument from addTestLine().
appDisplay: Let icon labels be multiline when hovered
When hovered, remove constraints from the icon labels that
limit the number of lines. Do that inside a saved easing
state so that the allocation can ease the label.
iconGrid: Allocate preferred sizes if bigger than child size
IconGridLayout uses the biggest minimum size to allocate its
children. Next commit will make app icons with long names show
not ellipsize on hover, and it is important that the icon itself
is able to follow that.
Use preferred size if it's bigger than the minimum size.
Florian Müllner [Wed, 21 Oct 2020 01:29:18 +0000 (03:29 +0200)]
audioDeviceSelection: Center-align icons
When the icon area gets allocated additional space, we want the
icons centered rather than left-aligned. This may happen in locales
with long-ish translations for the title or buttons when only a
subset of possible icons is shown.
Jonas Dreßler [Tue, 20 Oct 2020 23:02:37 +0000 (01:02 +0200)]
dnd: Get transformed size for scaling before unparenting drag actor
Commit de610a13f1ad1e7e34d4b9a81df58d4da3693059 in mutter made it
impossible to access the actors last allocation after unmapping it,
this broke the scale-up/down animation when starting a drag.
Fix that animation again by saving the actors transformed size before
unparenting (and therefore unmapping) the actor instead of afterwards.
Jonas Ådahl [Tue, 13 Oct 2020 13:51:22 +0000 (15:51 +0200)]
Adapt to CoglFramebuffers turning into GObjects
Requires some more explicit type casting as things turned into 'void *'
less. Also memory management changes, as we can't use CoglObject
functions anymore for these objects.
Ray Strode [Thu, 8 Oct 2020 13:36:30 +0000 (09:36 -0400)]
screenShield: Use identity operator when matching for motion events
The screen shield watches for motion events to know to display
the pointer when the user wiggles their mouse.
It checks for motion events by looking at the event type and
seeing if it is of type `Clutter.EventType.MOTION`. To do this
comparison it uses the equality operator (==). Using the equality
operator isn't considered best practice, because it can returns true
when comparing disparate types, if those types happen to be equivalent
after coersion.
From a code resiliance point of view, it's better to use the
identity operator (===), which requires both sides of the comparison
to be of the same type.
As a policy, any legacy code that gets changed or moved should be
switched away from the equality operator to the identity operator, if
appropriate.
This commit makes that change as prep work for a fix to that part of
the code.
environment: Handle @content for property transitions
When using `Actor.ease_property` if the property starts with '@' and the
duration of the transition is zero (which may happen if the actor is not
mapped even if a non-zero duration was passed to `ease_property`), the
impl will try getting the actual target object where the property should
be set.
This works fine for most cases but it currently throws an error when
passing '@content.*' properties. Fix this by handling '@content' as
a property of `actor.content` (used by MetaBackgroundActor when
showing the overview).
Ray Strode [Tue, 6 Oct 2020 14:29:00 +0000 (10:29 -0400)]
panel: Disconnect destroy signal handler when needed
The panel corners try to match their style to the buttons closest
to them. In order to make sure the corner styles stay in sync with
their neighboring buttons, they connect to the style-changed signals
of the buttons.
In order to make sure the style-changed signal handler isn't leaked,
it gets disconnected when the button is destroyed.
Unfortunately, the destroy signal handler connection itself gets leaked!
This commit ensures the destroy signal handler gets disconnected any
time the neighboring button is re-determined.
Florian Müllner [Mon, 5 Oct 2020 13:24:01 +0000 (15:24 +0200)]
extensionUtils: Adjust version check
With the new versioning scheme, the previously-minor version gets
shifted up to major, and unstable releases are marked by non-numeric
"versions" rather than uneven numbers. Reflect that in the extension
version check.
Florian Müllner [Sat, 3 Oct 2020 08:12:13 +0000 (10:12 +0200)]
workspacesView: Work around Clutter optimization
Clutter nowadays omits reallocations when only the stage position
changed, that is when the allocation relative to the parent changed.
As a result (apart from better performance of course), workspaces
in the overview may now end up with an outdated "actual geometry"
in case the overview moved to a new primary monitor (of equal size
as the previous one).
Work around that by emitting a signal from the overview on allocation
changes, and use that to update the cached geometry.
We can revert that change once workspaces become part of the regular
overview hierarchy.
appDisplay: Protect against source icon destruction
When dragging icons out of a folder dialog, there is a very peculiar
combination of steps that may break GNOME Shell:
1. Open an app folder dialog
2. Start dragging an icon to outside the grid
3. Wait until the popdown animation starts
4. Before it finishes, drop the icon
5. See the warnings / crash
That's caused by the source icon being destroyed after the delayed
move timer starts, and before it finishes.
Protect against the source icon being destroyed before the delayed
move timeout triggers by connecting to the 'destroy' signal and
removing the timeout on the callback. Use a single field, called
'_delayedMoveData', to store all data related to delayed moves.
appDisplay: Remove icon from folder when dragging out
When dragging an icon outside of a folder dialog, there's a small delay
before the dialog pops down. If the icon is dropped during this delay,
the drag is cancelled, and the icon continues to be in the folder.
However, this behavior turned out to be problematic, and it was a common
point of failure that throwing icons outside folders wouldn't work.
Remove the icon from the folder when dragging it to outside the dialog.
At the end of BaseAppView._clearAnimateLater(), the '_grid' actor's opacity is
set to 255. As it turns out, _clearAnimateLater() is called, among others, by
vfunc_unmap(). However, unmapping is part of the destruction process, and at
the time it is called, '_grid' is already destroying, which makes GJS complain
about accessing an invalid object.
Don't change opacity on BaseAppView._clearAnimateLater(), and instead move it
to the couple of places outside vfunc_unmap() that call it.
appDisplay: Don't emit 'apps-changed' when deleting folder
To delete a folder, FolderView needs to reset all keys under that particular
folder's GSettings path. That generates 5 'changed' signals, all of which
end up calling AppDisplay._redisplay(), which is costly.
If you try and drop an icon that's in the same page, but before the
drop target, it'll be one position ahead of where it should be -
because we just removed one icon before the target position.
Adjust the final position of the to-be-created folder.
appDisplay: Lighten folder dialog background when dragging out
As per design feedback, make the app folder dialog background lighter when
hovering it with an icon. This gives the visual feedback to show that that
region is a target.
Use the new DIALOG_SHADE_NORMAL variable in the other places where its color
was hardcoded.
This new public API moves items without removing and readding
them, which allows us to avoid some tricky behavior. Noticeably,
following the original design described at 3555550d5, the new
IconGridLayout.moveItem() method does not call `layout_changed`.
This is done by IconGrid itself, queueing a relayout.
appDisplay: Factor out linear position calculation
When adding an item to the app grid, the item is added to
a sorted array. This is calculated by adding all visible
items in pages before the one being modified. Future commits
will need this to move items without reparenting them, so
factor this code into a separate function.
Florian Müllner [Mon, 31 Aug 2020 19:11:18 +0000 (21:11 +0200)]
workspacesView: Always clip to allocation
We couldn't clip workspaces views during the overview transition
when we used the "porthole" approach, but as view's allocation
now always matches the expected visible area, we can just apply
the clip unconditionally.
Florian Müllner [Fri, 2 Oct 2020 15:10:42 +0000 (17:10 +0200)]
dash: Restore icon's expected DND behavior
Commit ff3d32dd186fa added a custom DashIcon subclass that disables
all DND methods from ancestors, including canceling the context
menu timeout and emitting the overview's item-drag-begin signal.
All we want is opting out of the parent's scale-and-fade behavior,
so override those methods instead.
Currently the width of the calendar column is solely determined
by the calendar, while other elements are ellipsized as necessary.
While that is the desired behavior for the events-, world clocks-
and weather sections, we don't want to cut off the date in the
header. However switching to bold text made that more likely in
non-English locales or when using large text, so explicitly take
it into account for the width negotiation.
authPrompt: set value of beginRequestType to 'DONT_PROVIDE_USERNAME'
Commit 809f820c introduct vmware SSO service to gnome-shell, the value
beginRequestType should be set to 'DONT_PROVIDE_USERNAME', otherwise
vmware sso will fail. This patch will fix it.
projects / gnome-shell.git / log