I fixed the memory leak that showed up on windows when using the thread library. The leak indeed was related to tss code (oposite to what I believed before). This also already has been fixed by Peter Dimov in the HEAD branch. However I took a different approach, by having replaced the leaking mutex by a CRITICAL_SECTION whose memory is static, thus is not subject to heap tracking. I did this because in Peters code I see a small chance that TlsFree could get into trouble. The on_process_exit could be called multiple times. E.g. from userland code before end of main _and_ from the tss callbacks. (Of course this could be easily cured by use of another once call.) Also I think it is usefull to be able to set a breakpoint on attached_thread_count to check if all threads (native and boost) have already died when it comes to shutdown. This can be used as an additional debugging aid. (The current BOOST_ASSERT does not work well in this place, and I will try to replace it by something more sensible. Ideas welcome! ) The difference between Bills code and my fix also is somewhat smaller, so I hope giving less chances to regressions. What still is left is the other leaks reported when really using tss. I will try to tackle these with a similar approach. Roland