[Python-Dev] Is it kosher to use a buffer after release?

Serhiy Storchaka storchaka at gmail.com
Sun May 10 20:46:16 CEST 2015

On 10.05.15 21:28, Larry Hastings wrote:
> In Python's argument parsing code (convertsimple in Python/getargs.c), a
> couple of format units* accept "read-only bytes-like objects", aka
> read-only buffer objects.  They call a helper function called
> convertbuffer() which uses the buffer protocol to extract a pointer to
> the memory.
> Here's the relevant bit of code:
>     static Py_ssize_t
>     convertbuffer(PyObject *arg, void **p, char **errmsg)
>     {
>     Py_buffer view;
>     ...
>     if (getbuffer(arg, &view, errmsg) < 0)
>          return -1;
>     count = view.len;
>     *p = view.buf;
>     PyBuffer_Release(&view);
>     return count;
>     }
> getbuffer() uses the buffer protocol to fill in the "view" buffer. If
> it's successful, "view" is a valid buffer.  We store the pointer to the
> buffer's memory in output parameter p.
> In case you missed the big helpful capital letters, we are returning a
> pointer given to us by PyObject_GetBuffer(), which we have already
> released by calling PyBuffer_Release().  The buffer protocol
> documentation for bf_releasebuffer makes it sound like this pointer
> could easily be invalid after the release call finishes.
> Am I missing something, or is this code relying on an implementation
> detail it shouldn't--namely that you can continue using a pointer to
> some (most? all?) buffer memory even after releasing it?

You are missing following code:

     if (pb != NULL && pb->bf_releasebuffer != NULL) {
         *errmsg = "read-only bytes-like object";
         return -1;

convertbuffer() is applicable only for types for which 
PyBuffer_Release() is no-op. That is why there are different format 
units for read-only buffers and for general buffers. That is why new 
buffer protocol was introduced.

More information about the Python-Dev mailing list