I'm skeptical about the implementation strategy chosen, in particular about the notion of a reversed semaphore, which I have never heard of. Is there prior art for such a construct?
Traditionally, rw locks provide certain fairness guarantees - why doesn't this implementation? To put it more bluntly: why is that so extravagant?

As for the style of the patch itself:
- please provide it as a unified or context diff. Making it relative to some released 2.5 version is fine.
- please drop the redundant pass statements
- documentation patches should be made against Doc/lib/libthreading.tex. If you cannot write LaTeX, just adding sections in plain text is fine.
- tests should be added to Lib/test/test_threading.py