I'm curious as to the results you get running "python -m test.__main__" and "python Lib/test/__main__.py" as well. I suspect both will fail.

Looking at the forking.py code, I think multiprocessing makes some assumptions that are flat out invalid in the presence of the -m switch - there are no guarantees that you can reverse engineer the name of an arbitrary module from the __file__ attribute. (This actually ties in with the current thread on python-ideas about making the real name of the main module readily available - then the parent process could just pass that through the pipe as a "main_name" value and this problem would go away)

In the short term, however, I think multiprocessing just needs to either special case the situation where "main_name" actually *is* __main__ or just drop the assertion across the board.