[issue13817] deadlock in subprocess while running several threads using Popen

[Charles-François Natali]

Charles-François Natali <neologix at free.fr> added the comment:

And here's the test.
To sum up:
- reinit_tls.diff makes sure to call PyThread_ReInitTLS() at the beginning of PyOS_AfterFork(), so that the TLS API is usable within PyOS_AfterFork() (e.g. in _after_fork()). It would be applied to 3.2 and default, even though they are not affected (at least on platform withs a native TLS implementation)
- reinit_tls_test.diff would be applied to 2.7, 3.2 and default: it should not trigger a failure on any of the active branches (2.7, 3.2 and default), it's just to prevent potential future regressions.
Notes regarding the test:
- I replaced subprocess with bare fork(), since it's way cheaper than spawning an entire subprocess
- I reduced the number of threads compared to Christoph's reproducer, because I feel uncomfortable forking up to 1024 processes in parallel (even though in practise they won't be that much, because child processes will terminate before the other threads call fork())
- even with 2.7.2, it won't deadlock 100%, since the race window is really short (but having a failure on one of the buildbots once would be enough to detect the problem, should it ever resurface)

----------
versions: +Python 2.7
Added file: http://bugs.python.org/file24348/reinit_tls_test.diff

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue13817>
_______________________________________