In order to use the new posix_spawn gspawn codepath, for more robust
app launching when available memory is low, we need to meet some
conditions.
child_setup needs to be NULL for this optimization to work, so drop
the internal child_setup that is used here. Replace it with a lightweight
wrapper binary (gio-launch-desktop) that sets GIO_LAUNCHED_DESKTOP_FILE_PID
before executing the app.
Adjust PATH for gio tests so that it can execute the new binary from the
build directory.
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.
In this case we are simply calling fork() in order to quickly call exec(),
which will throw away the entirity of the duplicated VM, so we should
look for ways to avoid the overcommit check.
The well known solution to this is to use clone(CLONE_VM) or vfork(), which
completely avoids creating a new memory address space for the child.
However, that comes with a bunch of caveats and complications:
https://gist.github.com/nicowilliams/a8a07b0fc75df05f684c23c18d7db234https://ewontfix.com/7/
In 2016, glibc's posix_spawn() was rewritten to use this approach
while also resolving the concerns.
https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=9ff72da471a509a8c19791efe469f47fa6977410
I experimented with a similar approach in glib, but it was not practical
because glibc has several items of important internal knowledge (such as
knowing which signals should be given special treatment because they are
NPTL implementation details) that are not cleanly exposed elsewhere.
Instead, this patch adapts the gspawn code to use posix_spawn() where
possible, which will reap the benefits of that implementation.
The posix_spawn API is more limited than the gspawn API though,
partly due to natural limitations of using CLONE_VM, so the posix_spawn
path is added as a separate codepath which is only executed when the
conditions are right. Callers such as gnome-shell will have to be modified
to meet these conditions, such as not having a child_setup function.
In addition to allowing for the gnome-shell "Cannot allocate memory"
failure to be avoided, this should result in a general speedup in this
area, because fork()'s behaviour of cloning the entire VM space
has a cost which is now avoided. posix_spawn() has also recently
been optimized on OpenSolaris as the most performant way to spawn
a child process.
Add a new process spawning function variant which allows the caller
to pass specific file descriptors for stdin, stdout and stderr.
It is otherwise identical to g_spawn_async_with_pipes.
Allow the same fd to be passed in multiple parameters. To make this
workable, the child process logic that closes the fd after the first time
it has been dup2'ed needed tweaking; we now just set those fds to be
closed upon exec using the CLOEXEC flag. Add a test for this case.
This will be used by gnome-shell to avoid performing equivalent
dup2 actions in a child_setup function. Dropping use of child_setup will
enable use of an upcoming optimized process spawning codepath.
Previously, calling:
g_dbus_is_supported_address ("some-imaginary-transport:", NULL)
correctly returned FALSE; but calling:
g_dbus_is_supported_address ("some-imaginary-transport:", &error)
crashed with:
GLib-GIO:ERROR:../gio/gdbusaddress.c:434:g_dbus_is_supported_address:
assertion failed: (ret || (!ret && (error == NULL || *error != NULL)))
This was because, if the address component did not start with a known
transport, no error was set. Fix this, reusing an error string used by
the corresponding else branch in g_dbus_address_connect(), and adjust
the test to pass both NULL and non-NULL GError **s to this function in
every test case. This case:
g_assert (!g_dbus_is_supported_address ("some-imaginary-transport:foo=bar;unix:path=/this/is/valid", NULL));
would have caught this bug with a non-NULL GError **.
The way things were before: a FreedesktopNotification struct is
allocated before the dbus call, and this same struct is possibly re-used
for other dbus calls. If the server becomes unavailable, the callback
will be invoked after the call times out, which leaves a long time where
other dbus calls can happen, re-using the same FreedesktopNotification
as user data. When the first call times out, the callback is invoked,
and the user data is freed. Subsequent calls that used the same user
data will time out later on, and try to free a pointer that was already
freed, hence segfaults.
This bug can be reproduced in Cinnamon 3.6.7, as mentioned in:
<https://github.com/linuxmint/Cinnamon/issues/7491>
This commit fixes that by always allocating a new
FreedesktopNotification before invoking dbus_call(), ensuring that the
callback always have a valid user data.
Signed-off-by: Arnaud Rebillout <elboulangero@gmail.com>
A slightly modified patch originally written by Morten Welinder
<terra@gnome.org> to make the error codes returned by g_spawn_*()
functions more specific when on Windows. They are already this specific
on Linux.
Add a unit test for the ENOENT case.
https://gitlab.gnome.org/GNOME/glib/issues/303
Signed-off-by: Philip Withnall <withnall@endlessm.com>
The error number was saved after some g_debug() and g_free() calls, in
various places, which meant it could have been overwritten since the
error we care about happened.
https://gitlab.gnome.org/GNOME/glib/issues/303
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Use a GOnce to make sure we only warn about notifications not being
supported on Windows once, rather than on each attempted notification.
Signed-off-by: Philip Withnall <withnall@endlessm.com>
Helps: https://gitlab.gnome.org/GNOME/glib/issues/1234
This allows test calls to produce output with \n
line separators on Windows, instead of \r\n.
Reduces the number of ifdefs, since all checks
can be done against one template on all platforms.
Test that the API behaves as expected, especially when we get to
saturation.
Additionally, check that both the function and the macro versions of the
API behave identically.
If we're using GCC we can use __extension__ to inline the grefcount and
gatomicrefcount API, and avoid the function call.
These macros are only enabled if G_DISABLE_CHECKS is defined, as they
remove critical warnings when the reference counters achieve saturation.
We have a common pattern for reference counting in GLib, but we always
implement it with ad hoc code. This is a good chance at trying to
standardise the implementation and make it public, so that other code
using GLib can take advantage of shared behaviour and semantics.
Instead of simply taking an integer variable, we should create type
aliases, to immediately distinguish the reference counting semantics of
the code; we can handle mixing atomic reference counting with a
non-atomic type (and vice versa) by using differently signed values for
the atomic and non-atomic cases.
The gatomicrefcount type is modelled on the Linux kernel refcount_t
type; the grefcount type is added to let single-threaded code bases to
avoid paying the price of atomic memory barriers on reference counting
operations.
On an Android build, API 22, at least, I got a:
> warning: "UNIX_PATH_MAX" redefined
We were currently defining it as:
> #define UNIX_PATH_MAX sizeof (((struct sockaddr_un *) 0)->sun_path)
Whereas Android's headers define this variable of sockaddr_un as:
> char sun_path[UNIX_PATH_MAX];
So by definition, we will still get the right result in the end by just
using the original value of UNIX_PATH_MAX.
C_IN macro was added years ago in bcbaf1bef0, using same value as the
internal code of Android with the reasonning that "some parts of the API
used by the resolver objects is not public in the Android NDK (yet)".
Well since then things are changed, since it is definitely available (at
least on the API 22 of Android which I am using) in the public header
arpa/nameser_compat.h.
Let's just add a #ifndef to handle both cases when you build with an
older or recent API.
Let people know that Visual Studio 2015 and later provide a /utf-8
option to eliminate the need to set the System's non-Unicode locale and
the subsequent reboot.
Also let people know that it is recommended to have GIT for Windows
installed so that some dependencies' sources can be downloaded and built
along with GLib et al when needed.
Please see MR !65 for a discussion on the reasoning why GIT for Windows
is recommended.
Meson has the ability to classify tests according to "suites", a list of
tags. This is especially useful when we want to run specific sets of
tests — e.g. only GLib's tests — instead of the whole test suite. It
also allows us to classify special tests, like "slow" ones, so that we
can only run them when needed.
FIONREAD ioctl on Linux reports the size of payload on UDP sockets.
However, other systems usually add internal header size to the reported
size, which vary between different operating systems and socket types.
To make it work on more systems, we should follow what we do on Windows
instead of using this unreliable FIONREAD ioctl.
This fixes socket test on FreeBSD.
libelf, just like libc, is not a single project. It is an interface
which can be implemented independently by different operating systems.
Therefore, we cannot expect all systems to provide a .pc file, and we
should fallback to cc.find_library and cc.has_function like what we
already do in autotools build.
Non-glibc gettext implementation seems to decide the language from
LC_MESSAGES environment variable instead of LC_MESSAGES locale, so
we should set both environment variable and locale when running tests
which need translation from specific languages.
