[issue38962] Reference leaks in subinterpreters

[STINNER Victor]

STINNER Victor <vstinner at python.org> added the comment:

"""
5 tests failed again:
    test__xxsubinterpreters test_atexit test_capi test_httpservers
    test_threading
"""

I merged Pablo's fix for test_httpservers.

With PR 17457 + PR 17453 (and the test_httpservers fix), the 5 tests don't leak anymore.

I am fine with PR 17457 and PR 17453, I added some comments.

I would prefer to avoid a revert (PR 17455) if possible ;-)

--

Except of PR 17453 which was my fault in my refactoring work, the problem is the work on subinterpreters doesn't introduce regressions but only reveal existing issues.

Example with Python compiled in the debug mode (master branch):

$ ./python -X showrefcount -c pass
[18564 refs, 6496 blocks]

Python "leaks" 18564 references at exit.

My work is to better isolate subinterpreters from the main interpreter, and so the missing clean up code is revealed by reference leak buildbot checks.

----------

_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue38962>
_______________________________________