(In reply to comment #0)
> From what I've been taught by Lennart, and what PulseAudio does, is that
> calling dbus_bus_get_private in several threads in parallel is correct. However
> doing so results in a segfault. 

What version of dbus are you using? There is known multi-threading breakage on Unix in all releases older than 1.5.10, a long-standing bug present since 2006 (Bug #43744).

Ok, thanks for the reply.
 
I'm using 1.4.18. Do you know if there is a fix available/planned for the 1.4 series?

(In reply to comment #2)
> Do you know if there is a fix available/planned for the 1.4 series?

Maybe... I can backport the changes to 1.4.x (and I need to do that soon anyway, for an embedded distro with a 1.4.x fork), but they're more intrusive than I'd really like to apply to a supposed stable branch. If they were less intrusive I'd have developed against 1.4.x to start with.

If you could try your test case with 1.5.10 (e.g. from Debian experimental), that would be very useful. I notice you have a Canonical e-mail address, so you're probably using Ubuntu? If so, you might not be able to use 1.5.10 installed system-wide without merging Ubuntu's Upstart-integration changes (probably not at all trivial), but it should still be possible to build 1.5.10 locally and link your test case against the libdbus-1.so.3 that it produces.

I'm hoping to get a 1.6.x new-stable branch out sometime soon, but there's still a backlog of changes that should get into in 1.6.x, which need code review (I can't review most of them myself, because I wrote them).

Hi,

It seems the problem is present in 1.5.10 as well. Perhaps you can reproduce by running the test case on your installation? It should be really simple to compile ( gcc dbus-test.c $(pkg-config --libs --cflags dbus-1) ).

I did the following (in Ubuntu 12.04, currently in beta) :

* sudo apt-get builddep dbus
* Downloaded 1.5.10, configure, make
* export LD_LIBRARY_PATH=$(pwd)/dbus/.libs
* ran the test case, which fails sometimes with SIGSEGV and sometimes with glibc "double free or corruption" errors.

Now that I look at your test case... you're not telling libdbus to use mutexes, so it won't be thread-safe at all. You need something like this:

    if (!dbus_threads_init_default ())
        {
            fprintf (stderr, "Unable to initialize threads (out of memory?)");
            exit (1);
        }

(or replace with your favourite way to handle fatal errors).

Annoyingly, libdbus can't be thread-safe-by-default without incompatible changes. Mutex setup requires memory to be allocated, at least on some platforms, so it can fail (on OOM), and we need to handle that somehow. libdbus policy is to not crash on OOM, so we can't do that - but some functions that are implemented with mutexes have no way to report OOM, so we can't initialize threads lazily and report OOM if it fails.

>     while (!quit) {

This is not reliable in a multi-threaded environment either, FWIW. Use an atomic variable or a mutex if you need two threads to communicate.

Indeed we seem to miss dbus_threads_init_default in both the test case and in PulseAudio. I will fix that, and report back in case that does not resolve our crashes.

Thanks for the advice!