Ok, I think I've managed to dig to the core issue.  It is actually the same issue as https://bugs.python.org/issue11768 (which was wrongly closed as fixed, apparently :-)).

Py_AddPendingCall() calls PyThread_acquire_lock() to try and take the pending calls lock.  Unfortunately, PyThread_acquire_lock() is not reentrant in the case where semaphores are not used (e.g. on OS X).  We can probably fix that first issue by calling pthread_mutex_trylock() instead of pthread_mutex_lock().

There is a second more fundamental issue, though, which is that PyThread_acquire_lock() calls into functions that are not async-signal-safe (see http://man7.org/linux/man-pages/man7/signal-safety.7.html for a list).  So, depending on the particular OS and libc implementation, PyThread_acquire_lock() can fail in mysterious ways (including hang the process) when called from a signal handler.

So perhaps the ultimate fix would be to remove the OS-based locking in Py_AddPendingCall and use a busy spinwait...  The performance implications may be bad, though.