The Pointer Corruption Bug

Or an alternative name for this post...

Why API Docs Should Have Examples

As part of my continuation of libAttachSQL for HP's Advanced Technology Group I have recently been focusing on a Python based wrapper called pyAttachSQL. This is currently at an alpha level of release with no package builds yet.

Today I want to talk about one (silly on my part) very frustrating bug I found whilst working on pyAttachSQL and why this means API docs should have examples for every call.

The Crash

Whilst writing the group connection functions I was using a Py_BuildValue call to generate parameters to use in a callback PyObject_CallObject. So the code looked a little like this:

cbargs= Py_BuildValue("iOOO", &events, &pycon, &pycon->query, &self->cb_args);
PyObject_CallObject(self->cb_func, cbargs);

Those who are Python API veterans will be able to see straight away where I went wrong but I am quite new to the API. The code was segfaulting on PyObject_CallObject.

My Big Dumb Mistake

Whilst debugging this I found the problem was in cbargs, for some reason the pointers to pycon and alike were slightly different to when they were set. After some time going over it again and again in GDB and watching the pointers get incremented it suddenly hit me. The incrementation was happening during reference increment functions inside Py_BuildValue. Which makes sense because the one first items in a PyObject structure is the reference count. I was supposed to pass pointers, not pointers to pointers. The Py_BuildValue function has no type checking so was taking whatever you passed to it as a pointer to a structure.

So the question many of you would be asking is: why did you pass pointers to pointers? That is easy to answer... Earlier in the code I have been using the PyArg_ParseTuple and similar functions which are on the same documentation page, using a similar API and I assumed the API was consistent with no examples to show me otherwise. The fix was to simply remove the & symbols from the above code.

Lessons Learnt

I guess there is two lessons I have learnt from this:

  1. Don't assume that an API is consistent
  2. Add examples for every API call in the documentation

I'll shortly be opening a ticket for libAttachSQL and pyAttachSQL to implement the second item.

Comments !