Sadly, the following test leaks:

    def test_leak(self):
        code = textwrap.dedent(r"""
            import os
            import _thread

            lock = _thread.allocate_lock()

            def _after_fork():
                pass

            os.register_at_fork(after_in_child=_after_fork)
        """)
        test.support.run_in_subinterp(code)

Output:

$ ./python -m test test_threading -R 3:3 -m test_leak
0:00:00 load avg: 0.77 Run tests sequentially
0:00:00 load avg: 0.77 [1/1] test_threading
beginning 6 repetitions
123456
......
test_threading leaked [56, 56, 56] references, sum=168
test_threading leaked [21, 21, 21] memory blocks, sum=63
test_threading failed

== Tests result: FAILURE ==

1 test failed:
    test_threading

Total duration: 724 ms
Tests result: FAILURE

I will investigate later why it leaks.

Workaround for the leak:

diff --git a/Python/pystate.c b/Python/pystate.c
index ead25b08d7..66ad5d6514 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -311,6 +311,7 @@ interpreter_clear(PyInterpreterState *interp, PyThreadState *tstate)
 
     /* Last garbage collection on this interpreter */
     _PyGC_CollectNoFail(tstate);
+    _PyGC_CollectNoFail(tstate);
     _PyGC_Fini(tstate);
 
     /* We don't clear sysdict and builtins until the end of this function.

Example of leaking tests:

./python -m test test_threading -R 3:3 -m test_threads_join -m test_threads_join_2 

These tests run import threading in a subinterpreter, and Lib/threading.py calls os.register_at_fork() which keeps the threading module dictionary alive longer than expected: I created https://bugs.python.org/issue42671 to try to fix this issue.

I fixed the leak by adding a traverse function to the lock type (and convert it to a GC type).