Re: Review of gnio, round 1

On Mon, 2009-04-27 at 11:55 -0400, Dan Winship wrote:
> Ryan Lortie wrote:
> > Not sure I agree.  See shutdown() syscall.  The fact that this call
> > exists means that the designers of [unix or tcp or whatever] went out of
> > their way because they disagreed with you.
> I can't think of any time I've used SHUT_RD or SHUT_WR rather than
> SHUT_RDWR though...

You can use SHUT_WR to get reliable socket shutdown, making sure the
remote side got all the data (without using SO_LINGER). See e.g.:

Although this kind of thing looks like something the highlevel code
should support not by exposing shutdown, but rather by having a function
that does all of this.

> >> Furthermore, its actually a problem in the ipv4 vs ipv6 magic case. The
> >> current tcplistener code first tries to do an ipv6 socket and only if
> >> that fails it tries an ipv4 socket. This makes sense on linux, were an
> >> ipv6 socket also can accept ipv4 connections. However, this is not true
> >> on many other unixes, where you need two sockets to handle both ipv4 and
> >> ipv6. So in this case the listener object actually gets in the way, as
> >> we'd need to create two listener objects to handle this (or make the
> >> listener have two sockets).
> > 
> > The 4-over-6 functionality (and even the setsockopt to disable it) is
> > specified in some RFC somewhere.
> Right, but it's disabled (at the kernel level) by default on most OSes
> (everything but Linux?) because apparently the behavior is
> underspecified. (Eg, see
> So to avoid separate Linux-vs-everyone-else codepaths, it's probably
> best to use IPV6_V6ONLY and manage v4 and v6 sockets separately everywhere.

As per this doesn't
work for linux, as it only allows one socket per port. Isn't portable
programming fun!

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]