test_basic_multiple_interpreters_deleted_no_reset() of test_imp/test_import leaks references

[vstinner]

The test_basic_multiple_interpreters_deleted_no_reset() test of test_import (previously in test_imp) leaks references but only when test_import is run with -jN option:

The test leaks since it was added by the commit bb0cf8f at Feb 27.

Leak with -j1 option:

$ ./python -m test -j1 -R 3:3 test_import -m test_basic_multiple_interpreters_deleted_no_reset 
test_import leaked [10, 10, 10] references, sum=30

No leak without the -j1 option, magic!

$ ./python -m test -R 3:3 test_import -m test_basic_multiple_interpreters_deleted_no_reset 
(...)
Tests result: SUCCESS

cc @ericsnowcurrently

[vstinner]

The bug was not spotted by buildbots since the test fail when run with -jN option, but pass when regrtest main process re-runs the test in the same process:

./python -E  ./Tools/scripts/run_tests.py -j 1 -u all -W --slowest --fail-env-changed --timeout=11700 -R 3:3 -u-cpu -j40 
== CPython 3.13.0a0 (heads/main:bd1b6228d1, May 26 2023, 02:34:01) [GCC 13.1.1 20230511 (Red Hat 13.1.1-2)]
== Linux-6.2.14-300.fc38.aarch64-aarch64-with-glibc2.37 little-endian
== Python build: debug
== cwd: /home/buildbot/buildarea/3.x.cstratak-fedora-stable-aarch64.refleak/build/build/test_python_3001460æ
== CPU count: 40
== encodings: locale=UTF-8, FS=utf-8
Using random seed 3314398
0:00:00 load avg: 3.41 Run tests in parallel using 40 child processes (timeout: 3 hour 15 min, worker timeout: 3 hour 20 min)
(...)
0:00:50 load avg: 21.90 [150/449/1] test_import failed (reference leak) (32.7 sec) -- running: test_decimal (50.6 sec), test_lzma (36.9 sec), test_cmd_line_script (40.5 sec), test_xmlrpc (49.2 sec), test_selectors (46.9 sec), test.test_asyncio.test_subprocess (43.5 sec), test_weakref (33.9 sec), test_traceback (36.1 sec), test_venv (34.9 sec), test_buffer (42.8 sec), test_pdb (44.9 sec), test_socket (35.1 sec), test.test_asyncio.test_events (49.8 sec), test_support (40.4 sec), test.test_asyncio.test_futures (47.1 sec), test_multiprocessing_forkserver (34.2 sec), test_unicodedata (49.0 sec), test_tarfile (50.3 sec), test_imaplib (48.0 sec), test_cppext (50.3 sec), test_gdb (40.4 sec), test_exceptions (47.1 sec)
beginning 6 repetitions
123456
......
test_import leaked [78, 80, 77] references, sum=235
test_import leaked [67, 67, 66] memory blocks, sum=200
(...)

1 test failed:
    test_import

(...)
0:33:44 load avg: 0.08 Re-running test_import in verbose mode
(...)
OK (skipped=3)

== Tests result: FAILURE then SUCCESS ==

433 tests OK.

1 re-run test:
    test_import

regrtest should always spawn a subprocess to re-run tests in a clean environment.

[sunmy2019]

It is to be fixed by #104796

when run with -jN option

He let tests skipped if -R is in sys.args (with -j, -R will not be passed to the actual test). You can see my PR for details and ongoing discussions.