You’ve maybe heard that with C and C++, it’s rarely a good idea to mix usage of the POSIX threading library, and fork(2) based concurrency. I’d heard this myself, but to be honest, I never quite understood why. The Linux manual page for fork(2) puts it in the following way:
A process shall be created with a single thread. If a multi-threaded process calls fork(), the new process shall contain a replica of the calling thread and its entire address space, possibly including the states of mutexes and other resources. Consequently, to avoid errors, the child process may only execute async-signal-safe operations until such time as one of the exec functions is called. Fork handlers may be established by means of the pthread_atfork() function in order to maintain application invariants across fork() calls.
But like so many things, it’s often easier to understand something when you have a concrete example.
So, for one of our projects, we’re using collectd for system monitoring, together with the exec plugin. Most the time, it would work perfectly well. But there turned out to be an interesting, if somewhat sporadic bug. We’d find that the child process would hang, before it had gotten round to executing the script we wanted to run.
Even weirder, when attached to the child with a debugger, and looked at the stack trace, it had hung inside of the name service switch subsystem, specifically getpwnam_r. This function can be used for mapping a username to a numeric user and group id, as collectd generally runs as root, and it’s generally unwise to run your plugins with super-powers intact.
Now, based on the immediate evidence, it looked awfully like there was a manifest bug in the C library. Now, you can imagine how confused I was by thisãƒ¼how could such a commonly used function contain such an obvious deadlock?
Now, as you might have guessed, it wasn’t a bug, it was what C library documentation (seemingly mockingly) calls undefined behavior.
What had happened, is that the name service switch code uses a global to protect its internal state. So, the thread running the exec plugin had forked whilst another thread was running with that mutex held, so when the mutex was unlocked in the parent, it never got unlocked in the child (because only the thread that calls fork(2) is copied to the child). So when collectd tried to lookup the user id to switch to, it called into getpwnam_r, and hung.
As it turned out, this had been a known issue for a while. Thankfully, like a lot of things, once you understand the problem correctly, the solution was simple, just move the username lookups to the parent process.
Anyway, I hope this is illuminating for you as it was for me.