I don't know if it is pertinent to this at all, but I raised https://bugs.python.org/issue44449 in which the faulthandler module can lead to a segfault inside Py_TRASHCAN_SAFE_BEGIN. Would that be avoided if frameobject.c was changed to use Py_TRASHCAN_BEGIN / END?

Duncan.

On Tue, 2021-08-17 at 12:00 +0200, Łukasz Langa wrote:

Hi everybody,
I'd like to revive this thread as I feel like we have to do something here but some consensus is needed first.

To recap, the current state of things is as follows:
- in March 2000 (d724b23420f) Christian Tismer contributed the "trashcan" patch that added Py_TRASHCAN_SAFE_BEGIN and Py_TRASHCAN_SAFE_END macros which allow destroying nested objects non-recursively.
- in May 2019 (GH-11841 of BPO-35983) Antoine Pitrou merged a change by Jeroen Demeyer which made Py_TRASHCAN_SAFE_BEGIN/END (unintentionally?) backwards incompatible; this was released in Python 3.8.0.
- by the way, GH-11841 introduced a new pair of macros (because they have different signatures) called simply Py_TRASHCAN_BEGIN and Py_TRASHCAN_END.
- by that time there was already a follow-up PR open (GH-12607) to improve backwards compatibility of the macros, as well as introduce tests for them; this was never merged.
- in Feb 2020 (0fa4f43db08) Victor Stinner removed the trashcan mechanism from the limited C API (note: not ABI, those are macros) since it accesses fields of structs not exposed in the limited C API; this was released in Python 3.9.0.
- in May 2020 Irit noticed that the backwards incompatibility (BPO-40608) causes segfaults for C API code that worked fine with Python 3.7. Using the new macros requires code changes but doesn't crash.

Now, there are a couple of things we can do here:
Option 1: Finish GH-12607 to fix the old macros, keeping in mind this will restore compatibility lost with Python 3.8 - 3.10 only for users of 3.11+
Option 2: Review and merge GH-20104 that reverts the macro changes that make old client code segfault -- unclear what else this needs and again, that would only fix it for users of 3.11+
Option 3: Abandon GH-12607 and GH-20104, instead declaring the old macros deprecated for 3.11 and remove them in 3.13

I personally agree with Irit, voting +1 for Option 3 since the old macros were soft-deprecated already by introducing new macros in 3.8, and more importantly made incompatible with pre-3.8 usage.

Let's talk on how to proceed.

- Ł

--

Duncan Grisby <duncan@grisby.org>