Using my PR 7827, I saw that test_refleaks_in_decompressor___init__() of test_lzma leaks 100 handles on Windows.

Extract of the code:

    @support.refcount_test
    def test_refleaks_in_decompressor___init__(self):
        gettotalrefcount = support.get_attribute(sys, 'gettotalrefcount')
        lzd = LZMADecompressor()
        refs_before = gettotalrefcount()
        for i in range(100):
            lzd.__init__()
        self.assertAlmostEqual(gettotalrefcount() - refs_before, 0, delta=10)

The test comes from bpo-31787: commit d019bc8.

In _lzmamodule.c

self->lock = PyThread_allocate_lock();

This seems leak some resource. But I'm not sure this lock contains refcnt on Windows.

Have a look at bpo-23224 and PR 7822.

This seems leak some resource. But I'm not sure this lock contains refcnt on Windows.

PR 7827 checks for leaks of Windows handles, this issue is about Windows handles, not Python reference leaks. But a Windows lock may use a Windows handle internally, I don't know :-)

It seems like test_bz2 has the same issue, but I'm not sure:
#7827 (comment)

I four times tried to update this issue, and only the forth one was successful. Other attempts was failed because "Edit Error: someone else has edited this issue (nosy, components, versions)." This is an extreme case of race condition in real life.

New changeset 9b7cf75 by Victor Stinner in branch 'master':
bpo-33916: Fix bz2 and lzma init when called twice (GH-7843)
9b7cf75

New changeset efc6bf6 by Miss Islington (bot) in branch '3.7':
bpo-33916: Fix bz2 and lzma init when called twice (GH-7843)
efc6bf6

New changeset 1094891 by Victor Stinner in branch '3.6':
bpo-33916: Fix bz2 and lzma init when called twice (GH-7843) (GH-7872)
1094891

It seems like Python 2.7 is inaffected by the bug:

static int
BZ2File_init(BZ2FileObject *self, PyObject *args, PyObject *kwargs)
{
    ...
#ifdef WITH_THREAD
    if (!self->lock) {
        self->lock = PyThread_allocate_lock();
    }

    if (!self->lock) {
        PyErr_SetString(PyExc_MemoryError, "unable to allocate lock");
        goto error;
    }
#endif
    ...
}

I fixed the leak in Python 3.6, 3.7 and master. I close the issue.