[Python-Dev] Assertion in _PyManagedBuffer_FromObject()

[Stefan Behnel]

Nick Coghlan, 02.03.2012 14:22:
> On Fri, Mar 2, 2012 at 10:55 PM, Stefan Krah wrote:
>> Nick Coghlan wrote:
>>> However, given the lack of control, an assert() isn't the appropriate
>>> tool here - PyObject_GetBuffer itself should be *checking* the
>>> constraint and then reporting an error if the check fails. Otherwise a
>>> misbehaving extension module could trivially crash the Python
>>> interpreter by returning a bad Py_buffer.
>>
>> I'm not so sure. Extension modules that use the C-API in wrong or
>> undocumented ways can always crash the interpreter. This assert()
>> should be triggered in the first unit test of the module. Now, if
>> the module does not have unit tests or they don't test against a
>> new Python version is that really our problem?
> 
> Crashing out with a C assert when we can easily give them a nice
> Python traceback instead is unnecessarily unfriendly. As Stefan Behnel
> pointed out, by tightening up the API semantics, we're already running
> the risk of breaking applications that relied on looking at what the
> old code *did*, since it clearly deviated from both spec (the PEP) and
> the documentation (which didn't explain how ReleaseBuffer works at
> all).
> 
>> Modules do need to be recompiled anyway due to the removal of
>> Py_buffer.smalltable, otherwise they will almost certainly crash.
> 
>> Perhaps an addition to whatsnew/3.3 would be sufficient.
> 
> That, updating the 2.7 and 3.2 docs with a reference to the fleshed
> out 3.3 semantics and converting the assert() to a Python exception
> should cover it.

One problem here: if the code raises an exception, it should properly clean
up after itself. Meaning that it must call PyBuffer_Release() on the
already acquired buffer - thus proving that the code actually works, except
that it decides to raise an exception.

I keep failing to see the interest in making this an error in the first
place. Why would the object that bf_getbuffer() is being called on have to
be identical with the one that exports the buffer?

Stefan