Date: Sun, 5 Apr 2009 22:02:04 +0000 (UTC) From: "Bjoern A. Zeeb" <bzeeb-lists@lists.zabbadoz.net> To: sthaug@nethelp.no Cc: freebsd-net@freebsd.org Subject: Re: IPv6 window scaling factor always 1 on initial SYN Message-ID: <20090405215842.C15361@maildrop.int.zabbadoz.net> In-Reply-To: <20090405214757.E15361@maildrop.int.zabbadoz.net> References: <20090405.231044.74688369.sthaug@nethelp.no> <20090405214757.E15361@maildrop.int.zabbadoz.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, 5 Apr 2009, Bjoern A. Zeeb wrote: > On Sun, 5 Apr 2009, sthaug@nethelp.no wrote: > >> On 7-STABLE, with kern.ipc.maxsockbuf=2621440, both sides set a window >> scaling factor of 6 (i.e. SYN wscale 6, SYN-ACK wscale 6) using IPv4. >> >> With the same value of kern.ipc.maxsockbuf, using IPv6, the side which >> sends the initial SYN sets a window scaling factor of only 1, while >> the other side sets a scaling factor of 6 in the SYN-ACK. This will >> obviously limit throughput in many cases. >> >> In both cases net.inet.tcp.rfc1323=1. >> >> Anybody know why IPv6 behaves differently here? >> >> tcpdump example: >> >> 22:20:37.282415 IP 193.75.4.50.53981 > 193.75.110.66.5555: S >> 1580765626:1580765626(0) win 65535 <mss 1460,nop,wscale 6,sackOK,timestamp >> 661320721 0> >> 22:20:37.282442 IP 193.75.110.66.5555 > 193.75.4.50.53981: S >> 1408884711:1408884711(0) ack 1580765627 win 65535 <mss 1460,nop,wscale >> 6,sackOK,timestamp 1581013561 661320721> >> >> 22:21:49.749586 IP6 2001:8c0:9a00:1::2.53983 > 2001:8c0:8500:1::2.5555: S >> 565631163:565631163(0) win 65535 <mss 1440,nop,wscale 1,sackOK,timestamp >> 661393190 0> >> 22:21:49.749633 IP6 2001:8c0:8500:1::2.5555 > 2001:8c0:9a00:1::2.53983: S >> 627173961:627173961(0) ack 565631164 win 65535 <mss 1440,nop,wscale >> 6,sackOK,timestamp 8 > > I think the answer to tthat is in sys/netinet/tcp_usrreq.c in the > functuoins: > tcp_connect > > 1106 /* > 1107 * Compute window scaling to request: > 1108 * Scale to fit into sweet spot. See tcp_syncache.c. > 1109 * XXX: This should move to tcp_output(). > 1110 */ > 1111 while (tp->request_r_scale < TCP_MAX_WINSHIFT && > 1112 (TCP_MAXWIN << tp->request_r_scale) < sb_max) > > ^^^^^^^^^^^ > > 1113 tp->request_r_scale++; > > > and tcp6_connect > > 1174 /* Compute window scaling to request. */ > 1175 while (tp->request_r_scale < TCP_MAX_WINSHIFT && > 1176 (TCP_MAXWIN << tp->request_r_scale) < so->so_rcv.sb_hiwat) > > ^^^^^^^^^^^ > > 1177 tp->request_r_scale++; > > > I'll have to check why they are un-equal... Ok, both versions had: < so->so_rcv.sb_hiwat) http://svn.freebsd.org/viewvc/base?view=revision&revision=166403 changed it for IPv4 the first time, http://svn.freebsd.org/viewvc/base?view=revision&revision=172795 changed it a second time for IPv4. Noone changed the IPv6 version. The syncache already seems to do it for both v4/v6 (common code). Can you try changing it to < sb_max) for IPv6 as well and see if things work (better) for you? /bz -- Bjoern A. Zeeb The greatest risk is not taking one.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20090405215842.C15361>