When the amount of free memory on the system is somewhat low, gnome-shell
will sometimes fail to launch apps, reporting the error:
fork(): Cannot allocate memory
fork() is failing here because while cloning the process virtual address
space, Linux worries that the thread being forked may end up COWing the
entire address space of the parent process (gnome-shell, which is
memory-hungry), and there is not enough free memory to permit that to
happen. This check is somewhat irrelevant because we are only forking
to immediately exec(), which will discard the whole virtual address
space anyway.
This issue can be avoided by using a new optimized gspawn codepath in
the latest glib development version, which uses posix_spawn() internally.
For the optimized codepath to be used, we must not pass a child_setup
function, so the the file descriptor management is reimplemented here
using new glib API to pass fds to the child process. The old API will
continue to be used on older glib versions.
We must also change the spawn flags for this code path to be hit.
I checked that gnome-shell's open file descriptors are all CLOEXEC
so using G_SPAWN_LEAVE_DESCRIPTORS_OPEN should be safe.
This will result in more resilient app launching when memory is low,
since the optimized spawn path avoids cloning the virtual address
space of the parent process (gnome-shell) and avoids the irrelevant
memory overcommit check.
https://gitlab.gnome.org/GNOME/gnome-shell/merge_requests/132
While can_open_new_window() uses some elaborate heuristics to predict
whether an application can open multiple windows, open_new_window()
will always simply relaunch the application. This is often the best
we can do, but when an application provides a "new-window" action in
its .desktop file or on the bus, it is much more likely to work as
expected than blindly activating the app and hoping for a particular
behavior.
https://bugzilla.gnome.org/show_bug.cgi?id=756844
Remove any usage of MetaScreen, as it has been removed from libmutter
in the API version 3. The corresponding functionality has been moved
into three different places: MetaDisplay, MetaX11Display (for X11
specific functionality) and MetaWorkspaceManager.
https://bugzilla.gnome.org/show_bug.cgi?id=759538
Similar to what it's done in _shell_app_add_window(), rely on calling
shell_app_sync_running_state() when removing windows too, to avoid
transitioning to STOPPED while it's still in the STARTING state.
This makes the logic compliant with the startup notification spec and
prevents stopping an application too early if it has a splash screen
that has been closed before the application's main window is shown.
https://bugzilla.gnome.org/show_bug.cgi?id=787905
And adapt existing callers to the new API. This will allow us to
implement a way to launch applications on the discrete GPU for systems
where an "Optimus" system exists.
https://bugzilla.gnome.org/show_bug.cgi?id=773117
We used to take window visibility into account when comparing apps
until commit 1dfc38d078, following changes in the window switcher
due to auto-minimization. However auto-minimization was abolished
and the window switcher changes reverted, so it makes sense again
to sort apps without non-minimized windows last again.
https://bugzilla.gnome.org/show_bug.cgi?id=766238
The JS code could still be holding on to a reference to a window-backed app
after all windows have vanished. (For example, the dash queues an idle to
refetch apps and display them.) Avoid dying with an error message if we
attempt to activate or otherwise manipulate such a window.
https://bugzilla.gnome.org/show_bug.cgi?id=674799
When the last interesting window of an app-backed window is removed,
we'll transition it back to STOPPED, but we transition the state and
send out the signal before we clear the running state.
This means that any listeners to the state-changed signal might
encounter a window-backed app that has a running state, but no
windows. If they call, e.g. shell_app_get_name, while in this state,
they'll encounter an assertion fail.
Apps that are starting might have uninteresting windows like splash
screens pop up and then go away (like LibreOffice), even when
startup-notification hasn't completed yet. In those cases, we don't
want to transition the app back to stopped -- it should remain in
the running state.
Themes - namely the HighContrast one - may now request symbolic
icons rather than fullcolor ones. In order to have recoloring
work as expected in that case, we will need a theme node to pick
up colors from - using an StIcon instead of manually loading a
texture from the cache gives us that for free, so do that.
https://bugzilla.gnome.org/show_bug.cgi?id=740447
Using a separate property to show when the application is busy rather
than cramming it into the state property makes the code clearer. In most
places we only care if an app is running or not, not whether it is
actually busy.
https://bugzilla.gnome.org/show_bug.cgi?id=736492
This simplifies the code and fixes a race where an application could
call g_application_mark_busy() before the shell subscribed to change
notifications on the application's busy state.
https://bugzilla.gnome.org/show_bug.cgi?id=736492
If the application reports itself as single window (through
an explicit indication in the desktop file or some heuristics),
not show a "New window" item that doesn't actually open a new window.
https://bugzilla.gnome.org/show_bug.cgi?id=722554
So far we have assumed that whether or not a window is interesting
is static. In general this is the case, but as it is legal for the
underlying properties to change at any time, there are of course
offenders that actually do this (flash I'm looking at ya).
While we used the property to determine whether a window should be
tracked or not, the worst case was showing windows that should be
hidden or missing windows that should be shown.
However as we nowadays base an app's running state on the number of
interesting windows, we need to be more careful in order to avoid
ending up with running apps with no windows.
https://bugzilla.gnome.org/show_bug.cgi?id=723308
With the lastest ShellApp changes, an app is considered stopped
when the last "interesting" window is closed. However the app
may still track non-interesting windows, so if we unref the
running state on the state transition, we hit an assertion later-on
when trying to remove the non-interesting window.
Fix this by keeping the running state around until the last window
is closed.
https://bugzilla.gnome.org/show_bug.cgi?id=722840
An app should be considered running if it has at least one "interesting"
window, however the code considers an app running if it has at least
one tracked window. This was fine while we were only tracking interesting
windows, but since commit d21aa0d85f this is no longer the case.
So keep track of the number of interesting windows as well and use that
to determine the running state.
https://bugzilla.gnome.org/show_bug.cgi?id=722690
Using the new list_actions() API in Gio, add entries for static
actions specified in .desktop files in the right-click app menus,
in the dash, app well and search.
https://bugzilla.gnome.org/show_bug.cgi?id=669603
Filtering out "non-interesting" windows beforehand as we currently do
means that we may get properties that should be based on all windows,
like the last time the application was used, wrong.
Just track all windows and filter out non-interesting windows manually
in the one place we actually care about the difference.
https://bugzilla.gnome.org/show_bug.cgi?id=719824
Rather than scanning all apps for searching, use Ryan's new desktop
file index and the glib support APIs for app searching instead of our
own system.
https://bugzilla.gnome.org/show_bug.cgi?id=711631
Systemd-for-the-user-session would also do this, but that's a deeply
invasive change that I may not actually get to this cycle. This
change is tiny and non-invasive, but provides an important benefit:
You can actually reliably tell *which* applications are logging which
messages (assuming they're launched by the shell).
This actually complements a recent change in DBus:
See https://bugs.freedesktop.org/show_bug.cgi?id=68559
which does a similar thing for bus activated apps.
https://bugzilla.gnome.org/show_bug.cgi?id=711626
