Problem:

If the loop() function of asyncore is invoked with poll
rather than select, the function readwrite() is used
when I/O is available on a socket.  However, this
function does not check for hangup - provided by POLLHUP.

If a socket is attempting to write, then POLLOUT never
gets set, so the socket hangs.  

Because poll() is returning immediately, but the return
value is never used, asyncore busy-loops, consuming all
available CPU.


Possible solutions:

The easy solution is to check for POLLHUP in the
readwrite() function:

        if flags & (select.POLLOUT | select.POLLHUP):
            obj.handle_write_event()

This makes the poll work exactly like the select - the
application raises a socket.error set to EPIPE.

An alternate solution - possibly more graceful - is to
invoke the handle_close() method of the object:

        if flags & select.POLLHUP:
            obj.handle_close()
        else:
            if flags & select.POLLIN:
                obj.handle_read_event()
            if flags & select.pollout:
                obj.handle_write_event()

This is incompatible with the select model, but it
means that the read and write logic is now the same for
socket hangups - handle_close() is invoked.

Logged In: YES 
user_id=410460

Perhaps, it would be better to raise exception:

def readwrite(obj, flags):
    try:
        if flags & (select.POLLIN | select.POLLPRI):
            obj.handle_read_event()
        if flags & select.POLLOUT:
            obj.handle_write_event()
        if flags & (select.POLLERR | select.POLLHUP |
select.POLLNVAL):
            obj.handle_expt_event()
    except ExitNow:
        raise
    except:
        obj.handle_error()

...

    def handle_expt_event(self):
        err = self.socket.getsockopt(socket.SOL_SOCKET,
socket.SO_ERROR)
        assert(err != 0)
        raise socket.error, (err, errorcode[err])

Since asyncore closes socket in handle_error, this solves
the problem too.

The solution suggested by klimkin seems to have made it into revision 35513 as a fix to bug #887279.

I'm not sure that this is necessarily the right solution to this bug or #887279, as a socket disconnect isn't necessarily an error condition, otherwise .handle_close_event() shouldn't exist for select-based loops, and it should always be an error.

Suggest switching to the last if clause of readwrite() to...

if flags & (select.POLLERR | select.POLLNVAL):
    obj.handle_expt_event()
if flags & select.POLLHUP:
    obj.handle_close_event()

Fixed for 2.6 in changelist 64768.
Fixed for 3.0 in changelist 64770.