[Cython] tp_clear() of buffer client objects (was: Re: test crash in Py3.2 (NumPy/memoryview/refcounting related?))

[Stefan Behnel]

Stefan Behnel, 23.04.2012 10:23:
> mark florisson, 22.04.2012 16:41:
>> On 22 April 2012 15:31, mark florisson wrote:
>>> I think the problem here
>>> is that a single memoryview object is traversed multiple times through
>>> different traverse functions, and that the refcount doesn't match the
>>> number of traverses. Indeed, the refcount is only one, as the actual
>>> count is the acquisition count. So we shouldn't traverse the
>>> memoryview objects in memoryview slices, i.e. not
>>> _memoryviewslice.from_slice.memview. I'll come up with a commit
>>> shortly, would you be willing to test it?
>>
>> BTW, tp_clear calls Py_CLEAR on Py_buffer.obj, shouldn't it call
>> releasebuffer instead?
> 
> Where is that? The memoryview class calls __Pyx_ReleaseBuffer() here.

Ah, found it. Yes, tp_clear() would be called before __dealloc__() in
reference cycles, so that's a problem. I'm not sure tp_clear should do
something as (potentially) involved as freeing the buffer, but if the
Py_buffer is owned by the object, then I guess it just has to do that.
Otherwise, it would leak the buffer.

The problem is that this also impacts user code, though, so a change might
break code, e.g. when it needs to do some cleanup on its own before freeing
the buffer. It would make the code more correct, sure, but it would still
break it. Guess we have to take that route, though...

Stefan