OK, but that's not the problem I see. The parent calls fork(), creates a child that then runs the atfork() handlers *before* returning from the os.fork() call (which is the expected behaviour). At least one of those atfork() handlers is the one registered by the logging module (Lib/logging/__init__.py) and that one attempts to acquire a lock. So in my case the child never gets to the point of calling exec(), which is why I didn't notice the bug in the script.