TCP on IPv6 when IPV6_USE_MIN_MTU is set

[kato%wide.ad.jp@localhost]

Folks,

This is not limited to NetBSD current but with all NetBSD releases
with IPv6, and other BSD based varients as well.

Applications which do not want to depend on Path MTU Discovery may set
IPV6_USE_MIN_MTU=1 to use IPv6_MMTU (1280) as defined in RFC3542. This
works well for UDP as long as fragmentation is allowed and no other
method to reduce message size is available.

But when the socket is set with this option, TCP doesn't see it and it
may send a larger segment especially when the peer notifies larger MSS
value. In this case, the segment is fragmented, and delivery
efficiency decreases. In some cases, if a middle box refuses fragments
(some of the boxes only pass the first fragment), the TCP session
breaks.

There was a short discussion on IETF 6man mailing list but no
conclusion was made:

https://www.ietf.org/mail-archive/web/ipv6/current/msg24977.html

I understood that it may not be the best to use IPV6_USE_MIN_MTU to
control this, however, there is no suitable knob defined at this
moment, and some TCP sessions currently fail in IPv6.

When IPV6_USE_MIN_MTU is defined to be 1 (IP6PO_MINMTU_ALL),
followings are necessary to avoid IPv6 fragmentation:
- it should advertise TCP MSS to be 1220 or less
- when TCP MSS is advertised from its peer, it should be clipped
  to be 1220 or less
- when it send a TCP segment, its segment size should be 1220 or less
Note: 1220 is IPV6_MMTU - IP6/TCP header size.

Enclosed is a quick ugly patch for NetBSD-7 to mitigate this
situation, which may not be complete.

-- Akira Kato, WIDE Project

*** tcp_input.c.ORG	Sun Aug  2 15:08:15 2015
--- tcp_input.c	Mon Nov  7 17:12:05 2016
***************
*** 4443,4448 ****
--- 4443,4461 ----
  	sc->sc_ourmaxseg = tcp_mss_to_advertise(m->m_flags & M_PKTHDR ?
  						m->m_pkthdr.rcvif : NULL,
  						sc->sc_src.sa.sa_family);
+ #ifdef	INET6
+ 	if (tp && tp->t_in6pcb && tp->t_in6pcb->in6p_outputopts) {
+ 		if (tp->t_in6pcb->in6p_outputopts->ip6po_minmtu ==
+ 			IP6PO_MINMTU_ALL) {
+ 			sc->sc_ourmaxseg = min(sc->sc_ourmaxseg,
+ 				IPV6_MMTU - sizeof(struct ip6_hdr)
+ 				- sizeof(struct tcphdr));
+ 			sc->sc_peermaxseg = min(sc->sc_peermaxseg,
+ 				IPV6_MMTU - sizeof(struct ip6_hdr)
+ 				- sizeof(struct tcphdr));
+ 		}
+ 	}
+ #endif
  	sc->sc_win = win;
  	sc->sc_timebase = tcp_now - 1;	/* see tcp_newtcpcb() */
  	sc->sc_timestamp = tb.ts_recent;
*** tcp_output.c.ORG	Sun Aug  2 15:08:15 2015
--- tcp_output.c	Tue Nov  1 08:29:38 2016
***************
*** 1124,1129 ****
--- 1124,1138 ----
  		tp->snd_nxt = tp->iss;
  		tp->t_ourmss = tcp_mss_to_advertise(synrt != NULL ?
  						    synrt->rt_ifp : NULL, af);
+ #ifdef INET6
+ 		if (tp->t_in6pcb && tp->t_in6pcb->in6p_outputopts) {
+ 			if (tp->t_in6pcb->in6p_outputopts->ip6po_minmtu ==
+ 				IP6PO_MINMTU_ALL)
+ 				tp->t_ourmss = min(tp->t_ourmss, 
+ 					IPV6_MMTU - sizeof(struct ip6_hdr)
+ 					- sizeof(struct tcphdr));
+ 		}
+ #endif
  		if ((tp->t_flags & TF_NOOPT) == 0 && OPT_FITS(4)) {
  			opt[0] = TCPOPT_MAXSEG;
  			opt[1] = 4;
*** tcp_subr.c.ORG	Thu Feb 26 12:44:13 2015
--- tcp_subr.c	Thu Nov  3 09:25:45 2016
***************
*** 2001,2006 ****
--- 2001,2014 ----
  	mss = tcp_mssdflt;
  	if (offer)
  		mss = offer;
+ #ifdef INET6
+ 	if (tp->t_in6pcb && tp->t_in6pcb->in6p_outputopts) {
+ 		if (tp->t_in6pcb->in6p_outputopts->ip6po_minmtu ==
+ 		    IP6PO_MINMTU_ALL)
+ 			mss = min(mss, IPV6_MMTU - sizeof(struct ip6_hdr)
+ 				- sizeof(struct tcphdr));
+ 	}
+ #endif
  	mss = max(mss, 256);		/* sanity */
  	tp->t_peermss = mss;
  	mss -= tcp_optlen(tp);