Adding a child source to an already-attached parent source would
crash, because we were passing the parent's context when setting the
child's priority.
01ed78d525cf2f8769022e27cc2573ec7ba123b3 introduced assertion checks for
creating a main context, forking, and attempting to use the main context
from the child side of the fork.
Some code (such as gnome-keyring-daemon) daemonise after calling
GMainContext. That's probably still mostly safe since we still only
have one side of the fork touching the context afterwards.
This use case is still troubling, however, since if any worker threads
have been created at the time of the fork(), we could end up in the
classic situation of leaving some mutexes in a locked state when the
other threads disappear from the copy of the image that the child gets.
This will require some deeper thinking...
Some code using GLib (gnome-keyring-daemon, for example) assumes that
they can catch signals by masking them out in the main thread and
calling sigwait() from a worker.
The problem is that our new worker thread catches the signals before
sigwait() has a chance and the default action occurs (typically
resulting in program termination).
If we mask all the signals in our worker, then this can't happen.
And remove the 'joinable' argument from g_thread_new() and
g_thread_new_full().
Change the wording in the docs. Clarify expectations for
(deprecated) g_thread_create().
Add g_main_context_ref_thread_default(), which always returns a
reffed GMainContext, rather than sometimes returning a (non-reffed)
GMainContext, and sometimes returning NULL. This simplifies the
bookkeeping in any code that needs to keep a reference to the
thread-default context for a while.
https://bugzilla.gnome.org/show_bug.cgi?id=660994
Deprecate both g_thread_create functions and add
g_thread_new() and g_thread_new_full(). The new functions
expect a name for the thread.
Change GThreadPool, GMainContext and GDBus to create named threads.
https://bugzilla.gnome.org/show_bug.cgi?id=660635
This variable, which is the cause of much grief, exists for two reasons:
- ensuring the the wakeup pipe doesn't fill up
- preventing the first poll() after adding a source from waking up
immediately
The first point is no longer an issue with GWakeup.
The second point is addressed by using different logic: we only signal a
wakeup in the case that the context is currently acquired by a thread
that is not us.
As an added bonus, we can now implement g_main_context_wakeup() without
taking a lock.
https://bugzilla.gnome.org/show_bug.cgi?id=583511https://bugzilla.gnome.org/show_bug.cgi?id=320888
Since GMainLoop is now purely monotonic, it really doesn't make sense to
cache the wallclock time just for the sake of making a deprecated call
more efficient.
The first time this is called, this creates a GMainContext * and a
thread to run it. Future calls return the same. There are a lot of
places that we could use this in GLib.
G_THREADS_ENABLED still exists, but is always defined. It is still
possible to use libglib without threads, but gobject (and everything
above it) is now guaranteed to be using threads (as, in fact, it was
before, since it was accidentally impossible to compile with
--disable-threads).
https://bugzilla.gnome.org/show_bug.cgi?id=616754
Since there was nothing guaranteeing synchronization of the
assignments to checked and clockid, it would be possible for one
thread to set clockid = CLOCK_MONOTONIC, and for another thread to see
checked = TRUE but still clockid = CLOCK_REALTIME.
https://bugzilla.gnome.org/show_bug.cgi?id=655129
The Linux eventfd() call is basically tailor made for the main loop
wake up pipe - all we want is a threadsafe way to write to a file
descriptor, and wake up the context on the other end; we don't care
about the content at all.
The eventfd manual page basically explains the benefits:
Applications can use an eventfd file descriptor instead of a
pipe (see pipe(2)) in all cases where a pipe is used simply to
signal events. The kernel overhead of an eventfd file
descriptor is much lower than that of a pipe, and only one file
descriptor is required (versus the two required for a pipe).
When writing my multithreaded spawn test case I actually hit the 1024
file descriptor limit quickly, because we used 2 fds per main context.
This brings that down to 1.
https://bugzilla.gnome.org/show_bug.cgi?id=653140
Running gthread/tests/spawn-multithreaded in a loop, I very easily hit:
GLib-CRITICAL **: g_main_context_wakeup: assertion `g_atomic_int_get (&context->ref_count) > 0' failed
Testing the refcount still left a window where we would fall into the
assertion. Fix this by just locking the context.
Unify it more with the rest of the signal handling code. There's
no reason not to specify SA_RESTART and SA_NOCLDSTOP for flags
always, so just do it.
Remove unnecessary initialization, and push the internal API
towards just ensure_unix_signal_handler_installed_unlocked().
https://bugzilla.gnome.org/show_bug.cgi?id=651725
