Get a strong reference atomically for list item instead of 2 operations.

Original description of the problem from 91.md:

Vulnerability #91

Title: Racy list item borrow causes UAF

Category: Memory Safety Violations

Tags: write,race,env,dos

CWEs: CWE-416, CWE-367

CVSS: CVSS:4.0/AV:L/AC:L/AT:P/PR:L/UI:N/VC:L/VI:L/VA:H/SC:N/SI:N/SA:N

Severity: Medium (5.8)

Location: cpython/Modules/_pickle.c:3213:3214 in function batch_list_exact

Description

batch_list_exact reads list elements using PyList_GET_ITEM and only then increments the reference count (Py_INCREF) without holding the list lock or using a safe strong-ref getter (cpython/Modules/_pickle.c:3213-3214, also cpython/Modules/_pickle.c:3197-3198). In free-threaded mode, _pickle runs without the GIL (cpython/Modules/_pickle.c:8242), so concurrent list mutation is possible while pickling. A mutator thread can remove/replace the same element and decref it to zero (cpython/Objects/listobject.c:1145-1156) between size check/access in batch_list_exact (cpython/Modules/_pickle.c:3212-3214), leading to Py_INCREF on freed memory (use-after-free write).

Trigger Conditions

Pre-conditions:

• CPython is built/run in free-threaded mode (no global interpreter lock).
• The C accelerator _pickle is used (it declares no-GIL operation at cpython/Modules/_pickle.c:8242).
• Target object is an exact list and protocol is greater than 0, so save_list uses batch_list_exact (cpython/Modules/_pickle.c:3266-3269).
• The same list object is shared across threads.

Data flow:

1. Thread A calls pickle.dumps(shared_list, protocol=5); save_list dispatches to batch_list_exact (cpython/Modules/_pickle.c:3266-3269).
2. In batch_list_exact, Thread A evaluates loop/size and then fetches a borrowed element via PyList_GET_ITEM (cpython/Modules/_pickle.c:3212-3213).
3. Concurrently, Thread B mutates that list index (e.g., del shared_list[i] or shared_list[i] = other), and list code decrefs the old element (cpython/Objects/listobject.c:1145-1156), potentially freeing it.
4. Thread A performs Py_INCREF(item) on the stale pointer (cpython/Modules/_pickle.c:3214), triggering UAF memory corruption/crash.

Impact

This can cause native memory corruption in the interpreter (use-after-free with refcount write), typically leading to process crash (denial of service) and potentially enabling arbitrary code execution in worst case under favorable heap/layout conditions. Exploitability is constrained by requiring a free-threaded build and a concurrent mutation race on the same list object, but the code path is reachable from normal Python APIs (pickle.dumps) and does not enforce safety against such concurrent access.

Remediation

• In batch_list_exact() (cpython/Modules/_pickle.c), replace both unsafe borrowed-ref reads (PyList_GET_ITEM + Py_INCREF) with a strong-reference API (PyList_GetItemRef / internal equivalent) so element lifetime is acquired atomically for free-threaded builds.
• Remove dependence on repeatedly reading live list size inside the loop; instead iterate against a stable expected length captured in save_list() and passed into batch_list_exact().
• Add explicit mutation detection/error handling in the exact-list fast path (similar to dict/set batching): if size/index validity changes during traversal, raise a deterministic RuntimeError (e.g., “list changed size during iteration”) instead of continuing.
• Add/extend free-threaded regression tests for concurrent pickle.dumps() + list mutation to ensure no UAF/crash, and verify behavior is clean exception-only under race.

Reproduction

1. Build a free-threaded CPython (the no-GIL configuration) with debug aids enabled if possible (ASAN/UBSAN build, or at least --with-pydebug).
2. In that interpreter, create a test scenario with a shared exact list used by two concurrent threads:
   • one thread repeatedly pickles the same list with protocol > 0 (so _pickle takes the fast exact-list path),
   • the other thread continuously mutates that same list (delete/replace items, especially around low indexes).
3. Make the list elements objects with short lifetimes (frequent replacement with fresh temporary objects) and run both threads in tight loops for many iterations to maximize race frequency.
4. Repeat the run a few times if it does not trigger immediately; race timing is non-deterministic. Increasing core count, reducing sleeps, and extending runtime should increase hit rate.
5. Confirm the issue by observing any of the following:
   • interpreter crash/segfault during pickling,
   • debug build fatal error related to refcount/object validity,
   • sanitizer report showing a use-after-free or invalid memory access with frames near _pickle list batching / Py_INCREF on a list item.

Code Context

            item = PyList_GET_ITEM(obj, total);
            Py_INCREF(item);

• Issue: 22 free-threading race conditions #149816