I'm investigating a memory leak in 3.10, and while looking at the offending commit I stumbled upon this: in ceval.c:PyEval_EvalCodeEx, there's this allocation:

    PyObject **kwargs = PyMem_Malloc(sizeof(PyObject *)*kwcount);

The problem is that this isn't ever freed. And `kwargs` isn't used anywhere in the function body. It seems to me that this is silently leaking memory.

(this is the context of our current investigation btw: https://github.com/MagicStack/asyncpg/issues/874#issuecomment-1009383262)

This is the questionable place in the code: https://github.com/python/cpython/blob/dce642f24418c58e67fa31a686575c980c31dd37/Python/ceval.c#L6131-L6166

See that `kwargs` is allocated but never freed or used.

I can confirm that removing the kwargs allocation prevents the leak described in the asyncpg issue from happening.

New changeset 607d8a838f29ad3c4c4e85b39f338dade5f9cafe by Yury Selivanov in branch 'main':
bpo-46347: Fix memory leak in PyEval_EvalCodeEx. (#30546)
https://github.com/python/cpython/commit/607d8a838f29ad3c4c4e85b39f338dade5f9cafe

Turns out this leak affects CYthon quite a lot:

https://github.com/cython/cython/blob/29ad96444b8b1a4f05a6ac2328fde01de4782691/Cython/Utility/ObjectHandling.c#L2139-L2155

This seems to imply that every function call using __Pyx_PyCFunction_FastCall is leaking memory in 3.10, which is quite bad. We may need an urgent release for this :(

The function will still leak "kwnames" and "default" if creating the "func" object fails.  Admittedly that would only happen in a low-memory condition which is a) rare and b) probably only happens just before the interpreter completely dies, so it's not worth addressing during today's mild emergency.

(Sorry--it'll leak "kwnames", "newargs", and "defaults".)

Apparently, this is causing adyncpg to leak megabytes in seconds

New changeset b1a94f1fab7c0aee0705483616a1b2c3f2713c00 by Miss Islington (bot) in branch '3.10':
bpo-46347: Fix memory leak in PyEval_EvalCodeEx. (GH-30546)
https://github.com/python/cpython/commit/b1a94f1fab7c0aee0705483616a1b2c3f2713c00

The memory leak has been fixed in 3.10 and main branches. Can this issue be closed now?

> The function will still leak "kwnames" and "default" if creating the "func" object fails.

Yeah, here's another PR to address that: https://github.com/python/cpython/pull/30551

New changeset 20b5791ce9b47195ce51cfd5acb223b1ca59cdf0 by Yury Selivanov in branch 'main':
bpo-46347: Fix PyEval_EvalCodeEx to correctly cleanup in error paths (#30551)
https://github.com/python/cpython/commit/20b5791ce9b47195ce51cfd5acb223b1ca59cdf0

New changeset 6f9ca53a6ac343a5663cc5c52546acf9a63b605a by Yury Selivanov in branch '3.10':
bpo-46347: Fix PyEval_EvalCodeEx to correctly cleanup in error paths (#30553)
https://github.com/python/cpython/commit/6f9ca53a6ac343a5663cc5c52546acf9a63b605a

New changeset be578e0c063dad1dbb273f86d5bc77e4e6f14583 by Yury Selivanov in branch 'main':
bpo-46347: Yet another fix in the erorr path of PyEval_EvalCodeEx (#30554)
https://github.com/python/cpython/commit/be578e0c063dad1dbb273f86d5bc77e4e6f14583

when was the regression introduced?

I wouldn't necessarily rush an urgent release out unless this was just introduced in the 3.10.1 patch release.  Could it wait for 3.10.2 already scheduled for four weeks from now?

The leak was introduced in 3.10a5 (https://github.com/python/cpython/commit/0332e569c12d3dc97171546c6dc10e42c27de34b)

> Could it wait for 3.10.2 already scheduled for four weeks from now?

I don't feel comfortable leaving a ton of Cython functions leaking memory constantly on many function calls. According to https://github.com/MagicStack/asyncpg/issues/874 asyncpg can leak quite a lot of memory depending on your usage pattern.

Even if this is the first time we heard about this problem, I think this is important enough to justify a extra release.