|author||Will Thompson <email@example.com>||Wed Nov 22 11:43:55 2017 +0000|
|committer||Philip Withnall <firstname.lastname@example.org>||Wed Oct 03 11:30:19 2018 +0100|
gfileutils: make g_file_set_contents() always fsync() Previously, this function only called fsync() if @filename exists and is non-empty. This behaviour was introduced when the function was first written (6cff88ba18b3bc0d118308f109840cb163dcea03) and shortly afterwards (d20a188b1250ab3cf211d684429127d99378e886) respectively, with the latter justified as a performance optimisation. This meant that g_file_set_contents() does not provide the guarantee that developers assume it has, namely that after a call and a crash, @filename will either contain its previous contents or its new @contents. In practice, when it was previously non-existent or empty on a bog-standard ext4 filesystem, it would often contain NUL bytes matching the @length of @contents, requiring application developers to explicitly handle this third case. Given the documentation includes the word "atomic", we make this function provide the guarantee that was previously implied but untrue, and document it. If applications require higher performance at the cost of correctness, they can open-code the old behaviour, or we can add a new function to glib providing weaker guarantees. https://gitlab.gnome.org/GNOME/glib/issues/1302
GLib is the low-level core library that forms the basis for projects such as GTK+ and GNOME. It provides data structure handling for C, portability wrappers, and interfaces for such runtime functionality as an event loop, threads, dynamic loading, and an object system.
The official web site is: http://www.gtk.org/
Information about mailing lists can be found at http://www.gtk.org/mailing-lists.php
To subscribe, send mail to email@example.com with the subject “subscribe”.
See the file ‘INSTALL’
Bugs should be reported to the GNOME issue tracking system. (https://gitlab.gnome.org/GNOME/glib/issues/new). You will need to create an account for yourself.
In the bug report please include:
Information about your system. For instance:
And anything else you think is relevant.
How to reproduce the bug.
If you can reproduce it with one of the test programs that are built in the tests/ subdirectory, that will be most convenient. Otherwise, please include a short test program that exhibits the behavior. As a last resort, you can also provide a pointer to a larger piece of software that can be downloaded.
If the bug was a crash, the exact text that was printed out when the crash occured.
Further information such as stack traces may be useful, but is not necessary.
Patches should also be submitted as merge requests to gitlab.gnome.org. If the patch fixes an existing issue, please refer to the issue in your commit message with the following notation (for issue 123): Closes: #123
Otherwise, create a new merge request that introduces the change, filing a separate issue is not required.
GTask no longer imposes a fixed limit on the number of tasks that can be run_in_thread() simultaneously, since doing this inevitably results in deadlocks in some use cases. Instead, it now has a base number of threads that can be used “for free”, but will gradually add more threads to the pool if too much time passes without any tasks completing.
The exact behavior may continue to change in the future, and it‘s possible that some future version of GLib may not do any rate-limiting at all. As a result, you should no longer assume that GTask will rate-limit tasks itself (or, by extension, that calls to certain async gio methods will automatically be rate-limited for you). If you have a very large number of tasks to run, and don’t want them to all run at once, you should rate-limit them yourself.
g_test_run() no longer runs tests in exactly the order they are registered; instead, it groups them according to test suites (ie, path components) like the documentation always claimed it did. In some cases, this can result in a sub-optimal ordering of tests, relative to the old behavior. The fix is to change the test paths to properly group together the tests that should run together. (eg, if you want to run test_foo_simple(), test_bar_simple(), and test_foo_using_bar() in that order, they should have test paths like “/simple/foo”, “/simple/bar”, “/complex/foo-using-bar”, not “/foo/simple”, “/bar/simple”, “/foo/using-bar” (which would result in test_foo_using_bar() running before test_bar_simple()).
(The behavior actually changed in GLib 2.36, but it was not documented at the time, since we didn't realize it mattered.)
It is no longer necessary to call g_type_init(). If you are loading GLib as a dynamic module, you should be careful to avoid unloading it, then subsequently loading it again. This never really worked before, but it is now explicitly undefined behavior. Note that if g_type_init() was the only explicit use of a GObject API and you are using linker flags such as --no-add-needed, then you may have to artificially use some GObject call to keep the linker from optimizing away -lgobject. We recommend to use g_type_ensure (G_TYPE_OBJECT) for this purpose.
This release contains an incompatible change to the g_get_home_dir() function. Previously, this function would effectively ignore the HOME environment variable and always return the value from /etc/password. As of this version, the HOME variable is used if it is set and the value from /etc/passwd is only used as a fallback.
The ‘flowinfo’ and ‘scope_id’ fields of GInetSocketAddress (introduced in GLib 2.32) have been fixed to be in host byte order rather than network byte order. This is an incompatible change, but the previous behavior was clearly broken, so it seems unlikely that anyone was using it.
GIO now looks for thumbnails in XDG_CACHE_HOME, following a recent alignment of the thumbnail spec with the basedir spec.
The default values for GThreadPools max_unused_threads and max_idle_time settings have been changed to 2 and 15*1000, respectively.
It is no longer necessary to use g_thread_init() or to link against libgthread. libglib is now always thread-enabled. Custom thread system implementations are no longer supported (including errorcheck mutexes).
The thread and synchronisation APIs have been updated. GMutex and GCond can be statically allocated without explicit initialisation, as can new types GRWLock and GRecMutex. The GStatic_______ variants of these types have been deprecated. GPrivate can also be statically allocated and has a nicer API (deprecating GStaticPrivate). Finally, g_thread_create() has been replaced with a substantially simplified g_thread_new().
The g_once_init_enter()/_leave() functions have been replaced with macros that allow for a pointer to any gsize-sized object, not just a gsize*. The assertions to ensure that a pointer to a correctly-sized object is being used will not work with generic pointers (ie: (void*) and (gpointer) casts) which would have worked with the old version.
It is now mandatory to include glib.h instead of individual headers.
The -uninstalled variants of the pkg-config files have been dropped.
For a long time, gobject-2.0.pc mistakenly declared a public dependency on gthread-2.0.pc (when the dependency should have been private). This means that programs got away with calling g_thread_init() without explicitly listing gthread-2.0.pc among their dependencies.
gthread has now been removed as a gobject dependency, which will cause such programs to break.
The fix for this problem is either to declare an explicit dependency on gthread-2.0.pc (if you care about compatibility with older GLib versions) or to stop calling g_thread_init().
g_debug() output is no longer enabled by default. It can be enabled on a per-domain basis with the G_MESSAGES_DEBUG environment variable like G_MESSAGES_DEBUG=domain1,domain2 or G_MESSAGES_DEBUG=all
It is now allowed to call g_thread_init(NULL) multiple times, and to call glib functions before g_thread_init(NULL) is called (although the later is mainly a change in docs as this worked before too). See the GThread reference documentation for the details.
GObject now links to GThread and threads are enabled automatically when g_type_init() is called.
GObject no longer allows to call g_object_set() on construct-only properties while an object is being initialized. If this behavior is needed, setting a custom constructor that just chains up will re-enable this functionality.
GMappedFile on an empty file now returns NULL for the contents instead of returning an empty string. The documentation specifically states that code may not rely on nul-termination here so any breakage caused by this change is a bug in application code.
The functions g_snprintf() and g_vsnprintf() have been removed from the gprintf.h header, since they are already declared in glib.h. This doesn't break documented use of gprintf.h, but people have been known to include gprintf.h without including glib.h.
The Unicode support has been updated to Unicode 4.1. This adds several new members to the GUnicodeBreakType enumeration.
The support for Solaris threads has been retired. Solaris has provided POSIX threads for long enough now to have them available on every Solaris platform.
‘make check’ has been changed to validate translations by calling msgfmt with the -c option. As a result, it may fail on systems with older gettext implementations (GNU gettext < 0.14.1, or Solaris gettext). ‘make check’ will also fail on systems where the C compiler does not support ELF visibility attributes.
The GMemChunk API has been deprecated in favour of a new ‘slice allocator’. See the g_slice documentation for more details.
A new type, GInitiallyUnowned, has been introduced, which is intended to serve as a common implementation of the ‘floating reference’ concept that is e.g. used by GtkObject. Note that changing the inheritance hierarchy of a type can cause problems for language bindings and other code which needs to work closely with the type system. Therefore, switching to GInitiallyUnowned should be done carefully. g_object_compat_control() has been added to GLib 2.8.5 to help with the transition.
GLib 2.6 introduces the concept of ‘GLib filename encoding’, which is the on-disk encoding on Unix, but UTF-8 on Windows. All GLib functions returning or accepting pathnames have been changed to expect filenames in this encoding, and the common POSIX functions dealing with pathnames have been wrapped. These wrappers are declared in the header <glib/gstdio.h> which must be included explicitly; it is not included through <glib.h>.
On current (NT-based) Windows versions, where the on-disk file names are Unicode, these wrappers use the wide-character API in the C library. Thus applications can handle file names containing any Unicode characters through GLib's own API and its POSIX wrappers, not just file names restricted to characters in the system codepage.
To keep binary compatibility with applications compiled against older versions of GLib, the Windows DLL still provides entry points with the old semantics using the old names, and applications compiled against GLib 2.6 will actually use new names for the functions. This is transparent to the programmer.
When compiling against GLib 2.6, applications intended to be portable to Windows must take the UTF-8 file name encoding into consideration, and use the gstdio wrappers to access files whose names have been constructed from strings returned from GLib.
Likewise, g_get_user_name() and g_get_real_name() have been changed to return UTF-8 on Windows, while keeping the old semantics for applications compiled against older versions of GLib.
The GLib uses an ‘_’ prefix to indicate private symbols that must not be used by applications. On some platforms, symbols beginning with prefixes such as _g will be exported from the library, on others not. In no case can applications use these private symbols. In addition to that, GLib+ 2.6 makes several symbols private which were not in any installed header files and were never intended to be exported.
To reduce code size and improve efficiency, GLib, when compiled with the GNU toolchain, has separate internal and external entry points for exported functions. The internal names, which begin with IA__, may be seen when debugging a GLib program.
On Windows, GLib no longer opens a console window when printing warning messages if stdout or stderr are invalid, as they are in “Windows subsystem” (GUI) applications. Simply redirect stdout or stderr if you need to see them.
The child watch functionality tends to reveal a bug in many thread implementations (in particular the older LinuxThreads implementation on Linux) where it's not possible to call waitpid() for a child created in a different thread. For this reason, for maximum portability, you should structure your code to fork all child processes that you want to wait for from the main thread.
A problem was recently discovered with g_signal_connect_object(); it doesn't actually disconnect the signal handler once the object being connected to dies, just disables it. See the API docs for the function for further details and the correct workaround that will continue to work with future versions of GLib.