bpo-40071: Fix refleak in _functools module.

[phsilva]

https://bugs.python.org/issue40071

[vstinner]

I merged the PR to bring back Refleaks buildbots to green: the change fix https://bugs.python.org/issue40071

PR 19151 introduced a regression. If a second instance of _functools module is created and then destroyed, _functools_free() will clear kwd_mark. Later, if the first module instance is still used, it will likely crash since kwd_mark is now NULL :-(

C extension modules should not use static global variables, but use a module state.

@phsilva: Do you want to attempt to write a change to add a module state to _functools to store this variable? Maybe @corona10 can help you, or write if you cannot.

[phsilva]

@phsilva: Do you want to attempt to write a change to add a module state to _functools to store this variable? Maybe @corona10 can help you, or write if you cannot.

Yes, I do. WIll send a PR and ask @corona10 for help if needed.

I suspect this is a little trickier because the module uses kwd_mark in functions that do not receive PyModule*, but will take a further look. Thanks!

[vstinner]

I suspect this is a little trickier because the module uses kwd_mark in functions that do not receive PyModule*, but will take a further look. Thanks!

It's used in lru_cache_make_key() which is called indirectly by:

static PyObject *
lru_cache_new(PyTypeObject *type, PyObject *args, PyObject *kw)

Hum, maybe this change depends on PEP 573 implementation... which is not written yet! The PEP was just accepted.

[vstinner]

A workaround is to disable _functools_free() for now: never clear kwd_mark. And only reenable _functools_free() once lru_cache_new() will be able to access the module state.