halting problem :: GDK and threads

this article is meant as a guideline and an explanation for application developers still using the deprecated gdk_threads_* family of functions because of reasons, as well as application developers still using GTK+ 2.x.

newly written code should not use this API at all.

a programmer has a problem. “I know,” he said, “I will use threads!” now has he problems two. — old Romulan proverb

we all know that using GTK+ and threads is a fan favourite question; I routinely answered it on IRC, mailing lists, Stack Overflow, and even in person multiple times. some time ago, we actually fixed the documentation to provide an authoritative answer as to how you should do your long-running, blocking operations on a thread, and then update the UI from within the main thread at various synchronization points.

sadly, both the API and the documentation that came before the tenet above became well-known, were lacking. the wording was ambiguous, and the examples were far from clear in showing the idiomatic way of initializing and using GTK+ in a multi-threaded application.

if I asked ten random developers using GTK+ for their applications what is the correct way of initializing the old threading support, I’d probably get the answer “call gdk_threads_init() before you call gtk_init()” from at least half of them. that answer is wrong, but most likely they never noticed it because they never ported their application to other platforms.

the correct answer for the question of how to initialize the old thread safety support in GTK+ is actually this idiomatic code snippet:

int
main (int argc, char *argv[])
{
  // initialize GDK thread support
  gdk_threads_init ();

  // acquire the GDK lock
  gdk_threads_enter ();

  // initialize GTK+
  gtk_init (&argc, &argv);

  // build your UI ...

  // start the main loop
  gtk_main ();

  // tear down your UI...

  // release the GDK lock
  gdk_threads_leave ();

  // exit
  return 0;
}

as you can see, we acquire the GDK lock before calling the GTK+ API, like the documentation says:

As always, you must also surround any calls to GTK+ not made within a signal handler with a gdk_threads_enter() / gdk_threads_leave() pair.

why is this needed? because the gtk_main() call will try to release the GDK lock before entering the main loop, and re-acquire it before returning the control flow to you. this is needed to avoid deadlocks from within GDK itself, since the GDK frame processing is going to acquire the GDK lock whenever it’s needed inside its own critical sections.

if you don’t acquire the lock, gtk_main() will try to release an unlocked mutex, and if you carefully read the g_mutex_unlock() documentation you will notice that doing that results in undefined behaviour. what does undefined behaviour means? in GLib and GTK+, we use the term in the same sense as the ISO C standard uses it.

in this specific instance, if you’re on Linux, undefined behaviour does not mean much: by default, the GNU libc implementation of pthreads is permissive, so it will simply ignore the double unlock. if you’re running on an operating system from the *BSD family, however, your application will uncerimoniously abort. hence, why so far very few people actually noticed this.

starting from GLib 2.42 (the next stable release), the GMutex implementation on Linux has been switched from a pure pthread wrapper to be futex-based. given that we don’t pay any penalty for it, we decided to ensure consistent behaviour, and proper usage of the API. this means that GLib 2.42 will abort if you try to clear an uninitialized or locked GMutex, as well as if you try to unlock an already unlocked one. this ensures that your code will actually be portable, and you’ll catch errors much more quickly.

this also means that non-idiomatic GTK+ code will start breaking on Linux as well, just like it would on non-Linux platforms.

since our documentation was not really good enough for people to use, and since we could not enforce proper idiomatic code at the toolkit level, GDK will try to compensate.


it’s important to not that the fix in GDK does not absolve you from fixing your code: you are doing something wrong. it will allow, though, existing GTK+ 2.x/early GTK+ 3.x code calling gdk_threads_init() in the wrong way to continue working even in the face of undefined behaviour. take this as a chance to rework your code not to use the GDK API to mark critical sections, and instead use the proper approach of worker threads notifying the UI through idle and timeout callbacks executed from within the main thread — an approach that does not require calling gdk_threads_init() at all.

gtk development threads

Older posts

  1. , I put the notes of the GSK talk I gave at GUADEC 2014 online; I believe there should be a video coming soon as well. the notes are available on this very website.

Page 1 / 4