I've noticed a severe bug in my python 2.6.2 io module, and looking at _buffered_raw_seek in http://svn.python.org/view/python/trunk/Modules/_io/bufferedio.c?view=markup, it seems the bug is still there in the C reimplementation of buffered io classes.

The problem is, if we seek a buffered stream with whence=os.SEEK_CUR, that call is directly transmitted to the underlying raw stream, which has not the same file pointer position as the buffered stream. So in the end, the file pointer doesn't get moved at the right absolute offset, but at an offset which depends of the state of the read head buffer or the write buffer.
When using os.SEEK_CUR, it would be necessary to call tell() or to manually compute offsets, so that the seek() works as expected.

Regards, 
Pascal

Could you construct a test case?
(it's even better if you fix the offending code as well of course)

The problem seems to be fixed in trunk (the future 2.7).
BufferedRandom.seek() contains a block labelled "# Undo read ahead." which does not exists in 2.6.

I reproduce the problem with a file opened with mode "rb+"

from io import open
open("tmpfile", "wb").write("abcdefgh" * 1024 + "X")
f = open("tmpfile", "rb+")
print f.read(1) # "a"
f.seek(0, 1)
print f.read(1) # should be "b", 2.6 returns "X"

My bad, I had looked at _buffered_raw_seek, not buffered_seek >_<
Indeed, the code is OK in both trunk _io an _pyio modules.
And the SEEK_CUR flag (value : 1) seems more than sufficiently tested in test_io actually, for example in the function write_ops() :
http://svn.python.org/view/python/trunk/Lib/test/test_io.py?view=markup

So concerning python2.6, isn't that just possible to backport _pyio (and its test suite) to it (renamed as io.py) ? They seem to offer the same functionality, except that _pyio is much more robust than io.py (the code around seek(), in particular, is much more complete).
Just tell me, else I'll gather a patch for the seek() problem.

> So concerning python2.6, isn't that just possible to backport _pyio
> (and its test suite) to it (renamed as io.py) ?

I would not be against it, but someone has to provide a patch.
The current _pyio.py might rely on other new behaviour of 2.7, so
perhaps it won't be that easy.

Hum, with a selective merge (tortoiseSVN makes it easy), backporting _pyio should be doable in a decent time. Triaging pertinent tests should be more brain damaging :p
I'm looking at this.

Well, here is a patch for the seek() methods of io module, in python2.6 maintenance branch.
Finally, I've only backported some assertions and the offset stuffs - I'm not comfortable enough about recent io refactorings to do more (it changed pretty quickly while I looked away, actually :p).
Tests have been patched too (bufferedrandom wasn't tested as other buffer classes), and to have the whole suite pass, I had to modify testWriteNonBlocking() as well (mock objects returned wrong values - this test is marked as unreliable in sources anyway).
If more is needed to patch the py2.6 branch, just tell me B-)

BufferedRandom's seek() is still borken in the latest svn revision of python2.6 (I haven't checked python2.7 yet), so here is a new, smaller patch, tested with the 6 IO-related test suites this time, on win32 and linux.

The patch itself...

Hello

I advocate the inclusion of this patch to the 2.6 maintenance branch, because currently the io module in this branch (which is still the most recent 2.X version released) is simply broken. 
People using it will certainly encounter MAJOR file corruptions due to the buggy seek().

Or should we rather backport the much more evoluated io of trunk to py2.6 ?

Committed in r81203. Thank you, Pascal!