From: Alexander Bluhm <bluhm@openbsd.org>
Subject: Re: raw IPv6 in parallel
To: Vitaliy Makkoveev <otto@bsdbox.dev>
Cc: tech@openbsd.org
Date: Tue, 16 Apr 2024 00:19:42 +0200

On Tue, Apr 16, 2024 at 01:03:37AM +0300, Vitaliy Makkoveev wrote:
> > On 16 Apr 2024, at 00:25, Alexander Bluhm <bluhm@openbsd.org> wrote:
> > 
> > Hi,
> > 
> > This diff brings rip6_input() in line with shared net lock rip6_input().
> > 
> > The fields inp_icmp6filt and inp_cksum6 are protected by exclusive
> > net lock.  I have a follow up diff to tune them and document this.
> > 
> > IPv4 function rip_disconnect() calls soisdisconnected() and I think
> > IPv6 should do this, too.
> > 
> > Other BSDs are confusing.  NetBSD rip_disconnect() calls
> > soisdisconnected(), but rip6_disconnect() only clears SS_ISCONNECTED.
> > FreeBSD rip_disconnect() clears SS_ISCONNECTED, but rip6_disconnect()
> > calls soisdisconnected(), so it is the other way around.
> > 
> > Consistent would be to always call soisdisconnected().  The reuslt
> > is that SS_CANTRCVMORE and SS_CANTSENDMORE are set and subsequent
> > read and write result in EOF or EPIPE.
> > 
> > 4.4BSD allows reconnect for UDP, but not for raw IP.  I think we
> > should use the same behavior for IPv6.
> > 
> > ok?
> > 
> 
> ok mvs
> 
> soisdisconnected() awakes sleeping threads polling threads. This is
> better, because otherwise userland program continues to sleep and
> SS_ flag modification has no sense.
> 
> unix(4) sockets have the same problem in unp_detach(). We clean
> SS_ISCONNECTED and set so_error on disconnected peer, but don???t awake
> it???s sleeping threads.

soconnect() has this comment:

        /*
         * If protocol is connection-based, can only connect once.
         * Otherwise, if connected, try to disconnect first.
         * This allows user to disconnect by connecting to, e.g.,
         * a null address.
         */
        if (so->so_state & (SS_ISCONNECTED|SS_ISCONNECTING) &&
            ((so->so_proto->pr_flags & PR_CONNREQUIRED) ||
            (error = sodisconnect(so))))
                error = EISCONN;
        else
                error = pru_connect(so, nam);

It looks like the intension is to connect -> disconnect -> connect
again.  If userland is doing this, of course it must not sleep.
The disconnect call is synchronous from userland.

I am not aware of any program doing this.  For UDP it might work,
but I have not tested it.  Anyway, we do not need it for raw IP.

bluhm