gh-108240: Add _PyCapsule_SetTraverse() internal function

[vstinner]

The _socket extension uses PyCapsule_SetTraverse() to visit and clear the socket type in the garbage collector. So the _socket.socket type can be cleared in some corner cases when it wasn't possible before.

• Issue: _socket extension C capsule cannot be visited by the GC and so _socket.socket type may stay alive longer than expected #108240

---

📚 Documentation preview 📚: https://cpython-previews--108339.org.readthedocs.build/

[vstinner]

@erlend-aasland: Here is a more complete solution to add traverse/clear functions to the socket C API capsule.

[Eclips4]

IMO, I like this solution more than #108241, because #108241 looks like a solution of corner case.

[vstinner]

Maybe this function should not be added to the limited C API.

[vstinner]

test_ssl fails on macOS and altered the environment when re-run in verbose mode:

ERROR: test_preauth_data_to_tls_client (test.test_ssl.TestPreHandshakeClose.test_preauth_data_to_tls_client)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/Users/runner/work/cpython/cpython/Lib/test/test_ssl.py", line 4802, in test_preauth_data_to_tls_client
    client.connect(server.listener.getsockname())
ConnectionResetError: [Errno 54] Connection reset by peer
(...)
Warning -- threading_cleanup() failed to cleanup 0 threads (count: 0, dangling: 2)
Warning -- Dangling thread: <SingleConnectionTestServerThread(preauth_data_to_tls_client, stopped 123145440923648)>
Warning -- Dangling thread: <_MainThread(MainThread, started 4610459136)>

test.pythoninfo:

ssl.OPENSSL_VERSION: OpenSSL 3.0.10 1 Aug 2023
ssl.OPENSSL_VERSION_INFO: (3, 0, 0, 10, 0)

[erlend-aasland]

@erlend-aasland: Here is a more complete solution to add traverse/clear functions to the socket C API capsule.

As I mentioned to you earlier, I had thoughts of a similar API, and I think this is a better solution to the problem than #108241. I think we should get more eyes on the API, though. If we are to make it public right away, we should be sure to get it right the first time ;) Perhaps create a topic on Discourse?

[vstinner]

I removed the new function from the limited C API.

[vstinner]

Since this function is only used for a single extension yet, the _socket extension, I changed my PR to only add the function to the internal C API (Py_BUILD_CORE). This function can be made public later if needed.

[vstinner]

@Eclips4 @erlend-aasland: I chose the middle ground, start by making the API internal.

I tested manually that the change fix issue #108240 leak.

It's not worth it to backport the change it's a corner case to manually unload a module from sys.modules.

[erlend-aasland]

You could have used the internal APIs here:

Suggested change

	if (!PyObject_GC_IsTracked(op)) {
	PyObject_GC_Track(op);
	assert(_PyObject_IS_GC(op));
	if (!_PyObject_GC_IS_TRACKED(op)) {
	_PyObject_GC_TRACK(op);
	}

OTOH, _PyCapsule_SetTraverse is probably not part of hot code, so I guess the public APIs are fine.

[vstinner]

I used _PyObject_GC_IS_TRACKED() in my second PR, but without assert(_PyObject_IS_GC(op)) which looks overkill.

[erlend-aasland]

else is unneeded here.

[vstinner]

Currently, the callback can be NULL. I added a check in my second PR to reject NULL callbacks.

[erlend-aasland]

@Eclips4 @erlend-aasland: I chose the middle ground, start by making the API internal.

I agree that keeping it internal makes sense for now. I do think a public API like this is needed, though.

It's not worth it to backport the change it's a corner case to manually unload a module from sys.modules.

I agree that we should not backport this PR.

[vstinner]

@erlend-aasland: I wrote PR #108417 to address your review.