Coverage.py has a trace function written in C, for speed. It uses the Python C API, which is notoriously tricky to get right because you have to manage reference counts yourself.
I've made some significant changes to the trace function recently, to add plugin support to the C tracer. Adding tests for badly behaved plugins, I managed to crash Python. Not a traceback, a for-real crash in CPython.
Naturally, this means something is wrong in my C extension. Poring over the code, I couldn't see anything amiss. I'd long been intrigued by the idea of David Malcolm's CPyTracer, a plugin to gcc that performs static path analysis to find mistakes in Python C extensions, so I decided to give it a try.
The best instructions are on A. Jesse Jiryu Davis' blog: Analyzing Python C Extensions With CPyChecker. I installed Fedora as suggested, and got the compiler running without much trouble (I just typed "yum" every time I wanted to type "apt-get").
The simple way to run the checker worked fine:
CC=~/gcc-python-plugin/gcc-with-cpychecker python setup.py build
This generates very nice HTML reports (like this) in two different styles that walk you through a path through your code that leads to a bad outcome. Well, supposedly a bad outcome. I found as Jesse did that there are false positives.
With the default settings, the checker only considers 256 paths through a function then stops, to avoid combinatorial explosions. But my functions had many more paths than that.
I increased the memory size of my Fedora Vagrantfile, then told CPyChecker to push on to examine a quarter million paths:
CFLAGS="--maxtrans 250000" python setup.py build
This found a few issues, but did not resolve the crash I'm experiencing. Next step: rebuild CPython --with-pydebug.
BTW, Stefan Behnel has rewritten my extension in Cython, and I really should seriously consider switching over, so that this kind of thing doesn't happen any more.