[Monotone-devel] Performance improvement splitout

[Zack Weinberg]

[this was meant to go to the list, sorry Tim]

On 9/18/06, Timothy Brownawell <address@hidden> wrote:
> On Mon, 2006-09-18 at 11:08 -0700, Eric Anderson wrote:
> > It's a probe to see whether any "Error" data has arrived, which near
> > as I can tell included data sent over TCP with urgency.  When I went
> > searching for something that used it, I recall finding that the usher
> > code seemed to use this as a "something bad happened, go die now."
> > It's been a long time since I looked at this though.
>
> Um... that's probably a bad place to look, I just had it do what the
> example in select_tut(2) (and also IIRC monotone) did.
>
> Now that I know slightly more I'm fairly sure we can just not use that
> (give select() an empty FD_SET for that argument), since we never send
> OOB/high-priority data. But if we do look at it, killing the connection
> is probably a good response since we're receiving something we know the
> other side didn't send...

It seems to me that it would be adequate defensiveness if, when (and
only when) we are going to select() for read or write on a socket, we
also select for errors on that socket, and handle any sockets tagged
in the error set appropriately.  Note that I *think* things like TCP
RST packets get reported that way; it's not just for MSG_OOB / TCP
urgent.

zw