gh-151295: Fix use-after-free in bytes.join()/bytearray.join() via re-entrant __buffe...

diff --git a/Lib/test/test_bytes.py b/Lib/test/test_bytes.py
index e0e8dd4eccfb1b27332dedaaf9efd9de6901cd51..e211c3d15a4ed2018413fe6a5090eb850389d911 100644 (file)
--- a/Lib/test/test_bytes.py
+++ b/Lib/test/test_bytes.py
@@ -645,6 +645,32 @@ class BaseBytesTest:
         with self.assertRaises(TypeError):
             dot_join([memoryview(b"ab"), "cd", b"ef"])
 
+    def test_join_concurrent_buffer_mutation(self):
+        # __buffer__() can release the GIL, letting another thread concurrently
+        # mutate the joined sequence (simulated here by mutating in __buffer__).
+        # See: https://github.com/python/cpython/issues/151295
+        def make_seq(mutate):
+            # Item is only referenced from the list slot, so mutate() frees it.
+            class Item:
+                def __buffer__(self, flags):
+                    mutate(seq)
+                    return memoryview(b'x')
+            seq = [b'a', Item(), b'c']
+            return seq
+
+        for sep in (self.type2test(b''), self.type2test(b'::')):
+            with self.subTest(sep=sep):
+                # Changing the list length is reported as a RuntimeError.
+                seq = make_seq(lambda seq: seq.clear())
+                self.assertRaises(RuntimeError, sep.join, seq)
+
+                # The list length is unchanged, so the size-change recheck
+                # cannot fire: only keeping the item alive avoids the crash.
+                def replace(seq):
+                    seq[1] = b'z'
+                seq = make_seq(replace)
+                self.assertEqual(sep.join(seq), sep.join([b'a', b'x', b'c']))
+
     def test_count(self):
         b = self.type2test(b'mississippi')
         i = 105

diff --git a/Misc/NEWS.d/next/Library/2026-06-11-00-00-00.gh-issue-151295.NQYUzW.rst b/Misc/NEWS.d/next/Library/2026-06-11-00-00-00.gh-issue-151295.NQYUzW.rst
new file mode 100644 (file)
index 0000000..e9012f0
--- /dev/null
+++ b/Misc/NEWS.d/next/Library/2026-06-11-00-00-00.gh-issue-151295.NQYUzW.rst
@@ -0,0 +1,4 @@
+Fixed a crash (use-after-free) in :meth:`bytes.join` and
+:meth:`bytearray.join` that could occur if an item's
+:meth:`~object.__buffer__` concurrently mutates the sequence being joined.
+The mutation is now reported as a :exc:`RuntimeError` instead.

diff --git a/Objects/stringlib/join.h b/Objects/stringlib/join.h
index de6bd83ffe4c8bc127674f294a53fef74bed86b7..deebfeadc0f4fd87e4ddefc68418b6cd32063eae 100644 (file)
--- a/Objects/stringlib/join.h
+++ b/Objects/stringlib/join.h
@@ -68,13 +68,18 @@ STRINGLIB(bytes_join)(PyObject *sep, PyObject *iterable)
             buffers[i].len = PyBytes_GET_SIZE(item);
         }
         else {
+            /* item is only borrowed; its __buffer__() may run Python that
+               drops the sequence's last reference to it. */
+            Py_INCREF(item);
             if (PyObject_GetBuffer(item, &buffers[i], PyBUF_SIMPLE) != 0) {
+                Py_DECREF(item);
                 PyErr_Format(PyExc_TypeError,
                              "sequence item %zd: expected a bytes-like object, "
                              "%.80s found",
                              i, Py_TYPE(item)->tp_name);
                 goto error;
             }
+            Py_DECREF(item);
             /* If the backing objects are mutable, then dropping the GIL
              * opens up race conditions where another thread tries to modify
              * the object which we hold a buffer on it. Such code has data