Backport: Fix crash in getcurrent()/greenlet during early Py_FinalizeEx (3.2.6)

[nbouvrette]

Summary

Backports all crash fixes from PR #499 (targeting main for 3.3.3) to the maint/3.2 branch for a 3.2.6 release, providing stability fixes for Python 3.9 users.

While regression-testing greenlet 3.2.5 (the backport of PR #495) against Python 3.9.7 under uWSGI, we discovered three additional crash paths during Py_FinalizeEx that were not fixed by PR #495:

1. clear_deleteme_list() allocator crash: The vector copy used PythonAllocator (PyMem_Malloc), which can SIGSEGV during early Py_FinalizeEx when Python's allocator pools are partially torn down. Replaced with std::swap (zero-allocation) and switched deleteme_t to std::allocator.

2. ThreadState memory corruption: ThreadState objects were allocated via PyObject_Malloc, placing them in pymalloc pools that can be disrupted during finalization. Switched to std::malloc/std::free.

3. getcurrent() crash on invalidated type objects: _Py_IsFinalizing() is only set after call_py_exitfuncs completes inside Py_FinalizeEx, so atexit handlers could still call greenlet.getcurrent() when type objects had been invalidated. An atexit handler now sets a shutdown flag before any other cleanup runs.

4. Exception preservation: clear_deleteme_list() now preserves pending exceptions around its cleanup loop.

All new code is guarded by #if !GREENLET_PY311 — zero impact on Python 3.11+.

Additional changes

• Py_IsFinalizing() compat shim: Maps the public API name (available since Python 3.13) to _Py_IsFinalizing() on older versions. All call sites now use the standard name; remove the two-line shim when < 3.13 is dropped.
• Windows USS tolerance: Adds a 512 KB tolerance to the USS memory test on Windows to prevent flaky CI failures from OS-level measurement noise.
• Additional shutdown tests: Ports the new test_interpreter_shutdown.py tests covering atexit-registered greenlet construction and cross-thread cleanup.

Test plan

• Existing test suite passes
• New test_interpreter_shutdown.py tests pass
• Manual uWSGI stress test: 0 segfaults in 600 requests with worker recycling on Python 3.9.7 (greenlet reports 3.2.6.dev0)
• CI on all supported Python versions (3.9–3.14)

Relationship to PR #499

This is a manual port (not cherry-pick) of PR #499's crash fixes, adapted to the maint/3.2 code structure which differs from main in several ways (no free-threading support, setup.py vs pyproject.toml, different allocator layout, etc.). The crash fix logic is identical.

Made with Cursor