$ ./python -m test test_multibytecodec -m test.test_multibytecodec.Test_IncrementalEncoder.test_subinterp -R 3:3
(...)
test_multibytecodec leaked [258, 258, 258] references, sum=774

I simplified the code. The following test leaks references:

    def test_subinterp(self):
        import _testcapi
        code = textwrap.dedent("""
            import _codecs_jp
            codec = _codecs_jp.getcodec('cp932')
            codec = None
        """)
        _testcapi.run_in_subinterp(code)

_codecs_jp.getcodec() is defined in Modules/cjkcodecs/cjkcodecs.h. Extract:

    cofunc = getmultibytecodec();
    ...
    codecobj = PyCapsule_New((void *)codec, PyMultibyteCodec_CAPSULE_NAME, NULL);
    if (codecobj == NULL)
        return NULL;

    r = PyObject_CallOneArg(cofunc, codecobj);
    Py_DECREF(codecobj);

getmultibytecodec() is the _multibytecodec.__create_codec() which is defined in Modules/cjkcodecs/multibytecodec.c. Simplified code:

    codec = PyCapsule_GetPointer(arg, PyMultibyteCodec_CAPSULE_NAME);
    _multibytecodec_state *state = _multibytecodec_get_state(module);
    self = PyObject_New(MultibyteCodecObject, state->multibytecodec_type);
    self->codec = codec;
    return (PyObject *)self;

I added the test which leaks in bpo-42846.

commit 07f2cee
Author: Victor Stinner <vstinner@python.org>
Date: Fri Jan 8 00:15:22 2021 +0100

bpo-42846: Convert CJK codec extensions to multiphase init (GH-24157)

Convert the 6 CJK codec extension modules (_codecs_cn, _codecs_hk,
_codecs_iso2022, _codecs_jp, _codecs_kr and _codecs_tw) to the
multiphase initialization API (PEP-489).

Remove getmultibytecodec() local cache: always import
_multibytecodec. It should be uncommon to get a codec. For example,
this function is only called once per CJK codec module.

Fix a reference leak in register_maps() error path.

I don't think that the leak is new. It's just that it wasn't seen previously.

By the way, I wrote an article "Leaks discovered by subinterpreters":
https://vstinner.github.io/subinterpreter-leaks.html

This leak may be new kind related to capsule, I'm not sure so far.

Thank you so much for taking the time to write these blog posts, Victor, and for explaining your fixes is such great detail. It is very helpful!

PR 24165 fix one reference leak in the getcodec() function of CJK codecs. But it doesn't fix all ref leaks of this issue.

The leak can be simplified to:

    @support.cpython_only
    def test_subinterp(self):
        import _testcapi
        code = textwrap.dedent("""
            import encodings
            import _codecs_jp
            encodings._cache['cp932'] = _codecs_jp.getcodec('cp932')
        """)
        res = _testcapi.run_in_subinterp(code)
        self.assertEqual(res, 0)

Links about ref leaks related to encodings:

• https://bugs.python.org/issue1635741#msg364833
• https://bugs.python.org/issue1635741#msg364968
• https://bugs.python.org/issue36854#msg357160
• master...phsilva:remove-codec-caches

See also bpo-42671 "Make the Python finalization more deterministic" but it seems like PR 23826 makes the leak worse (+2000 leaked references instead of +200) :-)

encodings._cache['cp932'] = _codecs_jp.getcodec('cp932')

• encodings._cache is kept alive by encodings.search_function.__globals__
• encodings.search_function function is kept alive by PyInterpreterState.codec_search_path list. The function by _PyCodec_Register() in encodings/init.py: codecs.register(search_function).

For example, unregistering the search function prevents the leak:

            import encodings
            import _codecs_jp
            encodings._cache['cp932'] = _codecs_jp.getcodec('cp932')

            import codecs
            codecs.unregister(encodings.search_function)

The PyInterpreterState.codec_search_path list is cleared at Python exit by interpreter_clear().

The weird part is that the _codecs_jp.getcodec('cp932') codec object *is* deleted. I checked and multibytecodec_dealloc() is called with the object stored in the encodings cache.

A _multibytecodec.MultibyteCodec instance (MultibyteCodecObject* structure in C) is a simple type: it only stores pointer to C functions and C strings. It doesn't contain any Python object. So I don't see how it could be part of a reference cycle by itself. Moreover, again, it is deleted.

Calling gc.collect() twice works around the issue, which sounds like a missing traverse function somewhere.

diff --git a/Python/pystate.c b/Python/pystate.c
index c791b23999..66bbe1bf7d 100644
--- a/Python/pystate.c
+++ b/Python/pystate.c
@@ -321,6 +321,7 @@ 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.

New changeset e542d41 by Victor Stinner in branch 'master':
bpo-42866: Fix refleak in CJK getcodec() (GH-24165)
e542d41

Don't you need to free memory using PyObject_GC_Del when you allocate using PyObject_GC_New?

Ref. https://docs.python.org/3/c-api/typeobj.html#Py_TPFLAGS_HAVE_GC

Don't you need to free memory using PyObject_GC_Del when you allocate using PyObject_GC_New?

My PR uses the generic tp->tp_free which PyObject_GC_Del().

typo: My PR uses the generic tp->tp_free which *is* PyObject_GC_Del().

Ah, thanks! I also found that info in the docs: https://docs.python.org/3/c-api/typeobj.html#c.PyTypeObject.tp_free

New changeset 11ef53a by Victor Stinner in branch 'master':
bpo-42866: Add traverse func to _multibytecodec.MultibyteCodec (GH-24166)
11ef53a

Ok, it's now fixed.

To make sure that a heap type can be deleted, it must by a GC type, has a traverse function, and it must track its instances. We should take this in account when we convert static types to heap types. In practice, deleting a type is mostly an issue when we check for reference leak and an instance is kept alive until the last GC collection, at the end of Py_EndInterpreter().