Fix fcntl module to accept 'unsigned long' type commands for ioctl(2).

[koobs]

BPO	25026
Nosy	@larryhastings, @ceronman, @vadmium, @serhiy-storchaka, @koobs, @corona10

^{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 2015-09-08.06:35:17.934>
labels = ['interpreter-core', 'easy', 'type-bug']
title = "(FreeBSD/OSX) Fix fcntl module to accept 'unsigned long' type commands for ioctl(2)."
updated_at = <Date 2019-08-25.10:01:51.693>
user = 'https://github.com/koobs'

bugs.python.org fields:

activity = <Date 2019-08-25.10:01:51.693>
actor = 'corona10'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2015-09-08.06:35:17.934>
creator = 'koobs'
dependencies = []
files = []
hgrepos = []
issue_num = 25026
keywords = ['easy']
message_count = 10.0
messages = ['250163', '250164', '250166', '250170', '250173', '250174', '250178', '250322', '296062', '350447']
nosy_count = 6.0
nosy_names = ['larry', 'ceronman', 'martin.panter', 'serhiy.storchaka', 'koobs', 'corona10']
pr_nums = []
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue25026'
versions = ['Python 2.7', 'Python 3.4', 'Python 3.5', 'Python 3.6']

Linked PRs

• gh-69214: Fix fcntl.ioctl() request type #119498
• [3.13] gh-69214: Fix fcntl.ioctl() request type (GH-119498) #119504
• [3.12] gh-69214: Fix fcntl.ioctl() request type (#119498) #119505
• [3.12] Revert "[3.12] gh-69214: Fix fcntl.ioctl() request type (#119498) (#1… #119905
• [3.13] Revert "[3.13] gh-69214: Fix fcntl.ioctl() request type (GH-119498) (… #119906

[koobs]

In my current attempt to create a FreeBSD port for python35, I've come across a patch rejection for the fcntlmodule.c for a local port patch we've been carrying since Python 2.6:

https://svnweb.freebsd.org/ports/head/lang/python27/files/patch-Modules__fcntlmodule.c?revision=391238&view=markup

The original commit log for this change is:

====================================

Fix fcntl module to accept 'unsigned long' type commands for ioctl(2).

Although POSIX says the type is 'int', all BSD variants (including Mac OS X)
have been using 'unsigned long' type for very long time and its use predates
the standard long enough. For certain commands (e.g., TIOCSWINSZ, FIONBIO),
the Python value may get sign-extended on 64-bit platforms (by implicit type
promotion) and it causes annoying warnings from kernel such as this:

WARNING pid 24509 (python2.6): ioctl sign-extension ioctl ffffffff8004667e

====================================

I'm not sure how this should be fixed upstream, nor clear on how to re-patch it given recent changes to fcntlmodule.c

[serhiy-storchaka]

You need just replace unsigned_int with unsigned_long in Clinic declaration for fcntl.ioctl in Modules/fcntlmodule.c and regenerate Clinic code (make clinic).

[koobs]

Thanks for the insight Serhiy. A few questions ..

Is clinic code updated based on *.c declarations at build time? If not when/how is the best place/method to run this for our ports/package builds?

Does your suggestion to switch unsigned_int to unsigned_long imply that the following lines are no longer necessary?

• if (PyArg_ParseTuple(args, "O&Iw#|i:ioctl",
  + if (PyArg_ParseTuple(args, "O&kw#|i:ioctl",

How do we get something like this fixed upstream for FreeBSD/OSX?

If left un-patched, what is the impact on the user/system? Just a warning on console?

[serhiy-storchaka]

No, the clinic code is not updated at build time. Your should either update clinic code just after patching Modules/fcntlmodule.c (Python 3 is needed), or update clinic code at patch creation time and include changes of generated clinic files in the patch.

Generated clinic files contain a code for argument parsing (PyArg_ParseTuple...).

I don't know if there is easy way to make this change conditionally for FreeBSD/OSX. The only way that I know is writing custom converters.

Perhaps the original commit log is outdated. What was the type of code at the time of this log? Now it is unsigned int and this doesn't make sign extension when casted to unsigned long. While the code parameter is in the range 0..0xffffffff, unpatched code shouldn't have any visible effects.

[koobs]

@serhiy

If by "type of code at the time of commit" you mean upstream python code, msg250163 contains a link to what bits we replace in fcntlmodule.c and that "I -> k" has always been the same.

[serhiy-storchaka]

Originally the type of the code variable in fcntl_ioctl() was int. In cbad1f5cabb1 it was changed to unsigned int. I guess that the log message that you cited was written before this.

[vadmium]

If necessary, perhaps we could unconditionally change the Python-level argument to unsigned_long, and then add a conditional bit in the C code to convert it to int if appropriate.

But I wonder if most of the problem is fixed by bpo-1471 (linked from the commit Serhiy identified). As I see it, your patch should now only be needed to support “code” values outside of the range of unsigned_int, e.g. that require > 32 bits.

[serhiy-storchaka]

The question is: are ioctl codes outside of the unsigned int range used on BSD family or Mac OS X?

[vadmium]

Maybe bpo-16124 is related; it mentions 64-bit values, but it sounds like an obscure use case.

[corona10]

Can I take a look at this issue?
Is there anything should I care about except update clinic?
Thanks!

[emaste]

bpo-16124 is #60328

The question is: are ioctl codes outside of the unsigned int range used on BSD family or Mac OS X?

I don't have an authoritative answer for other BSDs or OS X, but for FreeBSD they are not. In addition, we have hidden the warning under a diagnostic option in freebsd/freebsd-src@a90fb6c