| msg207686 - (view) |
Author: STINNER Victor (vstinner) *  |
Date: 2014-01-08 14:33 |
| $ ./python -c 'import resource; resource.prlimit(-3, 11, "\udbff\udfff")' Erreur de segmentation (core dumped) The problem is a generic problem with PyArg_Parse functions and "(O)" format. With this format, the caller does not hold a reference to the object nor the tuple. If arbitrary Python code is executed before the object is used, the object pointer becomes a dangling pointer. resource.prlimit() uses: if (!PyArg_ParseTuple(args, _Py_PARSE_PID "i|(OO):prlimit", &pid, &resource, &curobj, &maxobj)) return NULL; In this issue, it's worse: the string is casted to a sequence, and each string character becomes a temporary substring of 1 character. The problem is that PyArg_ParseTuple() nor resource_prlimit() hold the reference, and so the curobj and maxobj are dangling pointer. Options: - raise an error if the second parameter is not a tuple: implement the check in prlimit() or i PyArg_ParseTuple()? - hold a reference to the sequence, to curobj and to maxobj instead of using borrowed references |
|
|
| msg207692 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) * |
Date: 2014-01-08 16:20 |
| Thank you for good example, Victor. See for early discussion. As for options: - I afraid we can't raise an error if the second parameter is not a tuple right now. Rather we should first emit deprecation warning, and raise an error only several releases later. - We can't turn borrowed references into non-borrowed references, because it will cause reference leaks in existing code. So what we should to do: * Convert all codes in the stdlib to not use "(...)" in PyArg_ParseTuple(). This was mainly done in . Perhaps resource.prlimit() was added after this. * Deprecate this dangerous feature. Early is better. And emit a warning to all core developers. |
|
|
| msg208272 - (view) |
Author: STINNER Victor (vstinner) * |
Date: 2014-01-16 11:06 |
| There is another option: modify the function to use argument clinic and implement something in argument clinic to hold a reference on borrowed references "O" and hold a reference on "(...)" sequence. Holding a reference on borrowed references "O" would make Python more safer against a whole class of bugs. |
|
|
| msg282897 - (view) |
Author: Martin Panter (martin.panter) * |
Date: 2016-12-11 03:14 |
| Revision 4bac47eb444c fixed setrlimit(). Perhaps those changes can just be applied again to prlimit(). I’m not an Arg Clinic expert, but isn’t one of its purposes to imitate native Python function signatures? Since argument unpacking was dropped from Python 2 to 3, I doubt Arg Clinic would grow support for this. I think the tuple should be unpacked inside the implementation body, just as a native Py 3 function has to unpack tuple arguments. BTW I couldn’t get either Victor’s "\udbff" test case, nor Serhiy’s BadSequence test case to crash, but the original MyNum class from Issue 6083 does crash. |
|
|
| msg282898 - (view) |
Author: Larry Hastings (larry) * |
Date: 2016-12-11 03:17 |
| Sorry, Argument Clinic doesn't support automatic tuple unpacking for arguments. It was almost never used, I don't think it was ever a good idea, and it would have made an already-too-complicated program even more complicated-er. |
|
|
| msg282909 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) * |
Date: 2016-12-11 07:43 |
| Following patch applies to resource.prlimit() the same solution as to resource.setrlimit(). |
|
|
| msg283584 - (view) |
Author: Martin Panter (martin.panter) * |
Date: 2016-12-19 05:39 |
| Patch looks good to me. Although maybe you don’t need the IndexError check in the test. Won’t limit[key] already handle that for you (as long as key isn’t -1 etc). |
|
|
| msg283586 - (view) |
Author: Roundup Robot (python-dev)  |
Date: 2016-12-19 06:07 |
| New changeset dac72bc14c00 by Serhiy Storchaka in branch '3.5': Issue #20191: Fixed a crash in resource.prlimit() when pass a sequence that https://hg.python.org/cpython/rev/dac72bc14c00 New changeset 7bc2923a41b6 by Serhiy Storchaka in branch '3.6': Issue #20191: Fixed a crash in resource.prlimit() when pass a sequence that https://hg.python.org/cpython/rev/7bc2923a41b6 New changeset b4d2bff1c5f8 by Serhiy Storchaka in branch 'default': Issue #20191: Fixed a crash in resource.prlimit() when pass a sequence that https://hg.python.org/cpython/rev/b4d2bff1c5f8 |
|
|
| msg283588 - (view) |
Author: Serhiy Storchaka (serhiy.storchaka) * |
Date: 2016-12-19 06:09 |
| > Although maybe you don’t need the IndexError check in the test. Good point. Thank you for your review Martin! |
|
|