gh-133009: fix UAF in xml.etree.ElementTree.Element.__deepcopy__

Lib/test/test_xml_etree.py

@@ -2960,6 +2960,24 @@ def element_factory(x, y):
         del b
         gc_collect()
 
+    def test_deepcopy(self):
+        # Prevent crashes when __deepcopy__() mutates the element being copied.
+        # See https://github.com/python/cpython/issues/133009.
+        class X(ET.Element):
+            def __deepcopy__(self, memo):
+                root.clear()
+                return self
+
+        root = ET.Element('a')
+        evil = X('x')
+        root.extend([evil, ET.Element('y')])
+        if is_python_implementation():
+            self.assertRaises(RuntimeError, copy.deepcopy, root)
+        else:
+            c = copy.deepcopy(root)
+            # In the C implementation, we can still copy the evil element.
+            self.assertListEqual(list(c), [evil])
+
 
 class MutationDeleteElementPath(str):
     def __new__(cls, elem, *args):

Misc/NEWS.d/next/Library/2025-04-26-15-50-12.gh-issue-133009.etBuz5.rst

@@ -0,0 +1,3 @@
+:mod:`xml.etree.ElementTree`: Fix a crash in :meth:`Element.__deepcopy__
+<object.__deepcopy__>` when the element is concurrently mutated.
+Patch by Bénédikt Tran.

Modules/_elementtree.c

@@ -813,6 +813,8 @@ _elementtree_Element___deepcopy___impl(ElementObject *self, PyObject *memo)
 
     PyTypeObject *tp = Py_TYPE(self);
     elementtreestate *st = get_elementtree_state_by_type(tp);
+    // The deepcopy() helper takes care of incrementing the refcount
+    // of the object to copy so to avoid use-after-frees.
     tag = deepcopy(st, self->tag, memo);
     if (!tag)
         return NULL;
@@ -850,8 +852,7 @@ _elementtree_Element___deepcopy___impl(ElementObject *self, PyObject *memo)
         if (element_resize(element, self->extra->length) < 0)
             goto error;
 
-        // TODO(picnixz): check for an evil child's __deepcopy__ on 'self'
-        for (i = 0; i < self->extra->length; i++) {
+        for (i = 0; self->extra && i < self->extra->length; i++) {
             PyObject* child = deepcopy(st, self->extra->children[i], memo);
             if (!child || !Element_Check(st, child)) {
                 if (child) {
@@ -865,7 +866,12 @@ _elementtree_Element___deepcopy___impl(ElementObject *self, PyObject *memo)
         }
 
         assert(!element->extra->length);
-        element->extra->length = self->extra->length;
+        /*
+         * The original 'self->extra' may be gone at this point if deepcopy()
+         * has side-effects. However, 'i' is the number of copied items that
+         * we were able to successfully copy.
+         */
+        element->extra->length = i;
     }
 
     /* add object to memo dictionary (so deepcopy won't visit it again) */
@@ -899,6 +905,8 @@ deepcopy(elementtreestate *st, PyObject *object, PyObject *memo)
 
     if (Py_REFCNT(object) == 1) {
         if (PyDict_CheckExact(object)) {
+            // Exact dictionaries do not execute arbitrary code as it's
+            // impossible to override their __iter__() method.
             PyObject *key, *value;
             Py_ssize_t pos = 0;
             int simple = 1;
@@ -908,13 +916,20 @@ deepcopy(elementtreestate *st, PyObject *object, PyObject *memo)
                     break;
                 }
             }
-            if (simple)
+            if (simple) {
+                // This does not call object.__copy__(), so it's still safe.
                 return PyDict_Copy(object);
+            }
             /* Fall through to general case */
         }
         else if (Element_CheckExact(st, object)) {
-            return _elementtree_Element___deepcopy___impl(
-                (ElementObject *)object, memo);
+            // The __deepcopy__() call may call arbitrary code even if the
+            // object to copy is a built-in XML element (one of its children
+            // may mutate 'object' in its own __deepcopy__() implementation).
+            ElementObject *tmp = (ElementObject *)Py_NewRef(object);
+            PyObject *res = _elementtree_Element___deepcopy___impl(tmp, memo);
+            Py_DECREF(tmp);
+            return res;
         }
     }
 
@@ -925,8 +940,10 @@ deepcopy(elementtreestate *st, PyObject *object, PyObject *memo)
         return NULL;
     }
 
-    PyObject *args[2] = {object, memo};
-    return PyObject_Vectorcall(st->deepcopy_obj, args, 2, NULL);
+    PyObject *args[2] = {Py_NewRef(object), memo};
+    PyObject *res = PyObject_Vectorcall(st->deepcopy_obj, args, 2, NULL);
+    Py_DECREF(args[0]);
+    return res;
 }