The memoryview object in Python 2.6 would help in the transition to
Python 3.0.  It is a lower-priority and could wait until 2.7 if it
doesn't get finished.

But the warning in 2.6 for buffer() says to use memoryview(), which does
not exist. So either the backport needs to happen in 2.6 or the warning
needs to go away for generic buffer() usage and instead be changed to
only occur when the 2to3 fixer won't work (which, I am guessing, is only
needed when more than one argument to buffer() is used).

Is there still time to do the backport for 2.6 at this late date?

Not only there may not be enough time, but:
1. the memoryview implementation itself is not finished (that is, in py3k)
2. polishing and documenting the buffer API is more important
3. there doesn't seem to be any use for memoryview objects right now

So, IMO, this can be re-targeted to 2.7.

I concur.

Not a release blocker then :-)

Here is a patch. Some remarks:
- I had to disable a few tests in test_memoryview.py, mainly because
array.array does not support the new buffer API in 2.x
- I kept the PyBytes_* aliases in memoryobject.c rather than using the
PyString_* equivalents. Is it ok?
- I didn't update the Windows build files, I suppose someone else can do
it (I'm unable to do it myself)
- I'm not sure about the tp_flags juggling in typeobject.c, my additions
may be redundant (although very lightweight)

This new patch includes documentation update and fixes as well.

Some comments:

- PyBytes_* API is fine.
- memory_str should die.
- Py_TPFLAGS_HAVE_NEWBUFFER needs to be documented.

Otherwise it looks good.

Here is an updated patch removing memory_str.
I know Py_TPFLAGS_HAVE_NEWBUFFER should be documented, but I'm not sure
where and how (I didn't add it in the first place).

Committed in r71070.