In python2.3, the following code works.  In python2.4
it fails with an AttributeError:

>>> import urllib2
>>> request = urllib2.Request("http://pheared.net")
>>> opener = urllib2.build_opener()
>>> r = opener.open(request)
>>> r.fileno()
Traceback (most recent call last):
  File "<stdin>", line 1, in ?
  File "/usr/lib/python2.4/socket.py", line 246, in fileno
    return self._sock.fileno()
AttributeError: HTTPResponse instance has no attribute
'fileno'

Without a fileno it's hard to do things like select.

Logged In: YES 
user_id=285914

I think the problem is at line 1010 in urllib2.py.

--- urllib2.py-orig     2005-10-16 11:04:30.000000000 -0400
+++ urllib2.py  2005-10-16 11:05:02.000000000 -0400
@@ -1007,7 +1007,7 @@
         # out of socket._fileobject() and into a base class.

         r.recv = r.read
-        fp = socket._fileobject(r)
+        fp = socket._fileobject(r.fp)

         resp = addinfourl(fp, r.msg, req.get_full_url())
         resp.code = r.status

Logged In: YES 
user_id=285914

Actually that's not entirely correct.  The fix is probably
more like the other hack in there:

--- urllib2.py-orig     2005-10-16 11:19:34.000000000 -0400
+++ urllib2.py  2005-10-16 11:22:30.000000000 -0400
@@ -1007,6 +1007,7 @@
         # out of socket._fileobject() and into a base class.

         r.recv = r.read
+        r.fileno = r.fp.fileno
         fp = socket._fileobject(r)

         resp = addinfourl(fp, r.msg, req.get_full_url())

Besides being outdated, the OP shows it's trivial to access the fileno:
"r.fileno = r.fp.fileno". I suggest closing.

Well, easy access is not the point. If the object is supposed to have
the file-like interface, and can provide a fileno(), it should.

Georg is right and this has been fixed apparently in 3.0, leaving 2.7
and older broken.

There are two possible solutions to this. One is to change
socket._fileobject.fileno() to simply try self._sock.fp.fileno() if
self._sock.fileno() does not exist. The other option is to add a
__getattr__ to httplib.HTTPResponse to redirect to self.fp. Anyone have
an opinion?

Another option is to change urllib.addinfourl to look for fileno() on an
HTTPResponse object.

OK, I'll work on a test + patch.

IMHO, using a fileno property looks better than __getattr__. Setting an
attribute in init works too, unless fp changes during the object life
(then 3.x is broken IIUC).

It works OK as a property of either urllib.addinfourl or of
httplib.HTTPResponse (socket would work, but is this broken there or a
higher layer). 

Tests and fixes for both solutions attached. As soon as one is chosen,
some docs on fileno would be nice, even for 3.x.


This (current) test in test_urllibnet.py passes, but I don't see how to
improve it, assuming it should fail/detect this bug (maybe self.urlopen
isn't testing urllib.urlopen correctly?):

    def test_fileno(self):
        if (sys.platform in ('win32',) or
                not hasattr(os, 'fdopen')):
            # On Windows, socket handles are not file descriptors; this
            # test can't pass on Windows.
            return
        # Make sure fd returned by fileno is valid.
        open_url = self.urlopen("http://www.python.org/")
        fd = open_url.fileno()
        FILE = os.fdopen(fd)

Hmm, always run all the tests :)

fileno is set in test_urllib2. New patches.

The urllib patch breaks test_pyclbr because test_pyclbr can't handle
properties. I think the test is wrong here, hence the hack attached...
otherwise, let's go with httplib.

These patches are small and marked as easy, can they be accepted?

Fix committed in revision 84932.  A property based fileno was not required, because it is hardly set, it is always often a read-only attribute. Also, py3k behavior is same as the current fix.

Documented it in r84933, r84934 and r84935