PyObject_GetAttrString and tp_getattr do not agree

[petdance]

BPO	39620
Nosy	@loewis, @encukou, @serhiy-storchaka, @ammaraskar, @petdance

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2020-02-13.03:23:38.855>
labels = ['expert-C-API', 'type-feature']
title = 'PyObject_GetAttrString and tp_getattr do not agree'
updated_at = <Date 2021-06-22.14:42:14.566>
user = 'https://github.com/petdance'

bugs.python.org fields:

activity = <Date 2021-06-22.14:42:14.566>
actor = 'petr.viktorin'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['C API']
creation = <Date 2020-02-13.03:23:38.855>
creator = 'petdance'
dependencies = []
files = []
hgrepos = []
issue_num = 39620
keywords = []
message_count = 4.0
messages = ['361929', '361930', '361931', '361936']
nosy_count = 5.0
nosy_names = ['loewis', 'petr.viktorin', 'serhiy.storchaka', 'ammar2', 'petdance']
pr_nums = []
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue39620'
versions = []

[petdance]

PyObject_GetAttrString(PyObject *v, const char *name)

typedef PyObject *(*getattrfunc)(PyObject *, char *)

The outer PyObject_GetAttrString takes a const char *name, but then casts away the const when calling the underlying tp_getattr. This means that an underlying function would be free to modify or free() the char* passed in to it, which might be, for example, a string literal, which would be a Bad Thing.

The setattr function pair has the same problem.

The API doc at https://docs.python.org/3/c-api/typeobj.html says that the tp_getattr and tp_setattr slots are deprecated. If they're not going away soon, I would think this should be addressed.

Fixing this in the cPython code by making tp_getattr and tp_setattr take const char * pointers would be simple. I don't have any idea how much outside code it would affect.

[ammaraskar]

Note that there was an earlier attempt to make it const
af68c87

but this was reverted as part of 15e6274

[petdance]

Do you know why it was reverted? (Granted, it was 15 years ago...)

It looks like the original changeset is trying to address at least two different problems with non-const string literals. My ticket here is focusing only on getattrfunc and setattrfunc. The handling of all the kwlist is a separate issue, that I'm about to write up as a different ticket.

[serhiy-storchaka]

It was reverted because it is backward incompatible change. It breaks a code which assigns to the tp_getattr field without explicit type cast to getattrfunc.