I would like to describe an issue in the _elementtree module, and then
propose a fix which would prevent this type of bug everywhere in the
codebase.

The issue exists in _elementtree_Element_get_impl in
/Modules/_elementtree.c. Here is the code:

    static PyObject *
    _elementtree_Element_get_impl(ElementObject *self, PyObject *key,
                                  PyObject *default_value)
    {
        PyObject* value;

        if (...)
            value = default_value;
        else {
            value = PyDict_GetItem(self->extra->attrib, key);
            ...
        }
        ...
    }

We look up "key" in the dictionary, that is, in self->extra->attrib. This is
done with the call to PyDict_GetItem(self->extra->attrib, key).

We need to hash the "key" object in order to find it in the dictionary. This
sometimes requires calling the key's __hash__ function, i.e., it might call on
to python code.

What happens if the python code gets the dictionary (self->extra->attrib) freed?
Then PyDict_GetItem will use it after it has been freed.

I've attached a proof-of-concept script which causes a use-after-free through
_elementtree_Element_get_impl due to this issue.

We could fix this by calling Py_INCREF on self->extra->attrib before calling
PyDict_GetItem. But grepping for "PyDict_GetItem.*\->" shows that there are many
places in the codebase where this situation occurs. Some of these may not have
much impact, but some of them likely will.

All these bugs could be fixed at once by changing PyDict_GetItem to call
Py_INCREF on the dictionary before it hashes the key.

Here's a PoC for the _elementtree module.

--- begin script ---

import _elementtree as et

state = {
    "tag": "tag",
    "_children": [],
    "attrib": "attr",
    "tail": "tail",
    "text": "text",
}

class X:
    def __hash__(self):
        e.__setstate__(state) # this frees e->extra->attrib
        return 13

e = et.Element("elem", {1: 2})
e.get(X())

--- end script ---

Running it (64-bits Python 3.5.2, --with-pydebug) causes a use-after-free with
control of the program counter:

(gdb) r ./poc13.py
Starting program: /home/xx/Python-3.5.2/python ./poc13.py

Program received signal SIGSEGV, Segmentation fault.
0x00000000004939af in PyDict_GetItem (op=0x7ffff6d5b1a8, key=0x7ffff6d528e0) at Objects/dictobject.c:1079
1079 ep = (mp->ma_keys->dk_lookup)(mp, key, hash, &value_addr);
(gdb) p mp->ma_keys->dk_lookup
$8 = (dict_lookup_func) 0x7b7b7b7b7b7b7b7b

Verified problem. Added patch to keep reference to dict in PyDict_GetItem.

Added test to patch

This is a bug in elementtree - the caller should own a reference to the dict during the entire PyDict_GetItem call.

We won't add the refcount change you propose because it's not free and this is a hot function.

The test in your patch doesn't fail for me (I tried on version 3.11 on a Mac). Are you still seeing the problem on version >= 3.9, or perhaps it was fixed in eaementtree by now?

Concur with Irit. We should patch the caller of PyDict_GetItem, not PyDict_GetItem.

The issue is no longer reproduced by the original test because of the cache for dict key tables. But it is not gone, and can be reproduced with modified test.

There may be many similar bugs in the Python core end extensions. Adding incref/decref in PyDict_GetItem and similar C API functions could fix many of them, but perhaps not all, and it would hit performance. I suppose modt of uses of PyDict_GetItem are safe.

New changeset d15cdb2 by Serhiy Storchaka in branch 'main':
bpo-27946: Fix possible crash in ElementTree.Element (GH-29915)
d15cdb2

New changeset beb8342 by Miss Islington (bot) in branch '3.10':
bpo-27946: Fix possible crash in ElementTree.Element (GH-29915)
beb8342

New changeset 52a9a71 by Miss Islington (bot) in branch '3.9':
bpo-27946: Fix possible crash in ElementTree.Element (GH-29915)
52a9a71