Hi Eric, I'm sorry but I had to revert your recent work. It introduced regressions in at least in test_io and test_multiprocessing_spawn on Windows and FreeBSD. I simply don't have the bandwidth to investigate this regression yet, whereas we really need the CI to remain stable to catch other regressions (the master branch received multiple new features recently, and there are some other regressions by that way ;-)).

test_multiprocessing_spawn is *very* hard to reproduce on FreeBSD, but I can reliably reproduce it. It just takes time. The issue is a crash producing a coredump. I consider that the bug is important enough to justify a revert.

The revert is an opportunity to seat down and track the root cause rather than urging to push a "temporary" quickfix. This bug seems to be very deep in the Python internals: thread state during Python shutdown. So it will take time to fully understand it and fix it (or redesign your recent changes, I don't know).

Tell me if you need my help to reproduce the bug. The bug has been seen on FreeBSD but also Windows:

* test_multiprocessing_spawn started to produce coredumps on FreeBSD: https://bugs.python.org/issue36114
* test_multiprocessing_spawn started to fail randomly on Windows: https://bugs.python.org/issue36116
* test_io started to fail randomly on Windows: https://bugs.python.org/issue36177

-- The Night's Watch