This implements bug #457114 and implements timed socket
operations. If a timeout is set and the timeout period
elaspes before the socket operation has finished, a
socket.error exception is thrown.

This patch integrates the functionality at two levels:
the timeout capability is integrated at the C level in
socketmodule.c. Socket.py was also modified to update 
fileobject creation on a win platform to handle the
case of the underlying socket throwing an exception.
The tex documentation was also updated and a new
regression unit was provided as test_timeout.py.

Logged In: YES 
user_id=6380

For a detailed review, see

http://mail.python.org/pipermail/python-dev/2002-May/024340.html

Logged In: YES 
user_id=116038

I've addressed all the issues brought up by Guido. The 2nd
version of the patch is attached here. In this version, I've
modified test_socket.py to include tests for the _fileobject
class in socket.py that was modified by this patch.
_fileobject needed to be modified so that data would not be
lost when the underlying socket threw an expection (data was
no longer accumulated in local variables). The tests for the
_fileobject class succeed on older versions of python
(tested 2.1.3) and pass on the newer version of python.

Logged In: YES 
user_id=6380

Thanks for the new version! I've checked this in.  I made
considerable changes; the following is feedback but you
don't need to respond because I've addressed all these in
the checked-in code!

- Thanks for the cleanup of some non-standard formatting.
However, it's better not to do this so the diffs don't show
changes that are unrelated to the timeout patch.

- You are still importing the select module instead of
calling select() directly. I really think you should do the
latter -- the select module has an enormous overhead (it
allocates several large lists on the heap).

- Instead of explicitly testing the argument to settimeout
for being a float, int or long, you should simply call
PyFloat_AsDouble and handle the error; if someone passes
another object that implements __float__ that should be
acceptable.

- gettimeout() returns sock_timeout without checking if it
is NULL. It can be NULL when a socket object is never
initialized. E.g. I can do this:

>>> from socket import *
>>> s = socket.__new__(socket)
>>> s.gettimeout()

which gives me a segfault. There are probably other places
where this is assumed.

- I addressed the latter two issues by making sock_timeout a
double, whose value is < 0.0 when no timeout is set.

Logged In: YES 
user_id=6380

Keeping this open as a reminder of things still to finish.

Most is in the python-dev discussion; Michael Gilfix and
Bernard Yue have offered to produce more patches.

One feature we definitely want is a way to specify a timeout
to be applied to all new sockets.

Logged In: YES 
user_id=6380

The default timeout is now implemented in CVS.

There's a bug report from Andrew Macintyre (unfortunately on
python-dev) about test_socket.py failures on FreeBSD. I'll
try to keep an eye on that, so this patch *still* stays
open. Also, Bernie has promised some changes that I haven't
received yet and the details of which I don't recall (sorry
:-( ).

Logged In: YES 
user_id=116038

Now that I'm back :)

I checked the archive and this seems to have been handled by
you. Please let me know if it isn't resolved and I can give
it a closer look.

Also, perhaps I should contact Bernie and ask him if there's
anything he hasn't gotten around to in the test_timeout that
I can off-load from him.

Logged In: YES 
user_id=250749

In private mail to/from Guido, it appears that the FreeBSD 
issues were in test_socket.py, and have been addressed.

I still have outstanding issues on OS/2 EMX, which I sent to 
Guido privately but will add here as soon as I can.

Logged In: YES 
user_id=116038

If Guido is busy (And I'm sure he is), I'd be willing to
take a hack at the problem if you could email me privately
and provide a testing environment (No OS/2 EMX in my apt ;) ).

Logged In: YES 
user_id=6380

Michael and Andrew, if you can deal with this without my
involvement I would greatly appreciate it. ;-)

Logged In: YES 
user_id=250749

After discussing the OS/2 issues privately with Michael, the
outstanding issues are resolved with the socketmodule.c and
test_socket.py patches I've uploaded here.

socketmodule.c.nb-connect.diff:
in the non-blocking connect, OS/2 is returning EINPROGRESS
from the initial connection attempt, and after the
internal_select(), the subsequent connection attempt returns
EISCONN.
this appears to be perfectly legitimate, although FreeBSD
and Linux haven't been seen to return the EINPROGRESS.
the patch adds specific handling for the EISCONN after
EINPROGRESS case, matching the semantics already in place
for the Windows version of the code.

test_socket.py.sendall.diff:
the existing sendall() test is flawed as the recv() call
makes no guarantees about waiting for all the data requested.
OS/2 required a 100ms sleep in the recv loop to get all the
data.
rewriting the reciev test to allow for recv() not waiting
for data still in transit is more correct.

Note that these interpretations of "correctness" have been
based on FreeBSD manpages, which is the only sockets
documentation I currently have.

If these are acceptable to Guido, and Michael gets to test
them on Linux, I can relieve Guido of committing them and
closing this patch.

Logged In: YES 
user_id=6380

Andrew, I've checked in your patches. I had to review them
anyway and I decided to rewrite the testSendAll() check to
be more informative. Thanks!

If there are other unsolved issues, please open a new bug
report and assign it to me.