The bug is related to http://bugs.python.org/issue1415 and occurs only
with the latest patch from #1415.

Writing to an invalid fd doesn't raise an exception:

>>> f = open(100, 'w')
>>> f.fileno()
100
>>> f.write("test")
4

However reading or opening an invalid fd for reading and writing raises
an exception.

>>> f = open(100, 'r')
>>> f.read()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 1253, in read
    res += decoder.decode(self.buffer.read(), True)
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 756, in read
    current = self.raw.read(to_read)
IOError: [Errno 9] Bad file descriptor
>>> f = open(100, 'w+')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 195, in __new__
    return open(*args, **kwargs)
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 169, in open
    buffer = BufferedRandom(raw, buffering)
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 948, in __init__
    raw._checkSeekable()
  File "/home/heimes/dev/python/py3k/Lib/io.py", line 301, in _checkSeekable
    if msg is None else msg)
IOError: File or stream is not seekable.

I expected that fileio_write() raises an exception when fd is invalid:

        n = write(self->fd, ptr, n);
        if (n < 0) {
                if (errno == EAGAIN)
                        Py_RETURN_NONE;
                PyErr_SetFromErrno(PyExc_IOError);
                return NULL;
        }

Python 2.5 and probably 2.6 suffer from the same problem although it's
harder to trigger it.

Python 2.5.1 (r251:54863, Oct  5 2007, 13:36:32)
[GCC 4.1.3 20070929 (prerelease) (Ubuntu 4.1.2-16ubuntu2)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import os
>>> f = open("/tmp/test", 'w')
>>> f.fileno()
3
>>> os.close(f.fileno())
>>> f.write("test")
>>> f.write("test")
>>>
close failed: [Errno 9] Bad file descriptor
$ ls -la /tmp/test
-rw-r----- 1 heimes heimes 0 2007-11-11 20:46 /tmp/test

I don't think this is worth fixing; you'll get an I/O error as soon as
I/O is attempted, which is upon the first read() for input, or upon the
firsh (implicit or explicit) flush() on output.  You can't tell whether
a fd is valid or not without attempting I/O on it, so trying to test on
each write() call would defeat the purpose of buffering. Testing on
open() is insufficient as long as anybody could call os.close() any time.