I've just been enlightened about the use of per-route mtu settings as an
alternate (and probably better) workaround for this problem.  As far as I
can see, there is no advantage to using my patch over setting the mtu on
the appropriate route.  Sorry for the noise.

Rick

On Sun, 2 Dec 2001, Rick Byers wrote:

> Hi,
> In order to work around buggy networks suffering from the PMTU blackhole
> problem (see RFC 2923), I've written up a quick patch which adds a sysctl
> to limit the advertised TCP MSS (I this this is preferable to lowering
> the interface MTU).  Ideally, this could be configured per interface or
> per route, or even auto-detected on a host-by-host basis - but all of
> those options require much more work.
>
> This should be most usefull for machines behind a NetBSD box using the new
> PPPoE implementation (which doesn't have any built-in support for TCP MSS
> clamping like most other PPPoE software does).
>
> The best way to work around this problem is debateable.  Ideally
> we could just convince people to fix their networks, but thats not so
> easy (I've spend over a month trying to convince the Bank of Montreal -
> www.bmo.com - that they have a problem that needs to be fixed).  Anyway,
> this is one possible work around, and I post it here in the hopes that
> others will find it usefull.  If this is comitted, it will close
> kern/10943.
>
> Rick
>
> Patch against NetBSD-1.5Y
> Index: sys/netinet/tcp_var.h
> ===================================================================
> RCS file: /cvsroot/syssrc/sys/netinet/tcp_var.h,v
> retrieving revision 1.87
> diff -c -r1.87 tcp_var.h
> *** tcp_var.h	2001/09/11 21:03:21	1.87
> --- tcp_var.h	2001/12/02 19:06:00
> ***************
> *** 559,565 ****
>   #endif
>   #define	TCPCTL_RSTPPSLIMIT	24	/* RST pps limit */
>   #define	TCPCTL_DELACK_TICKS	25	/* # ticks to delay ACK */
> ! #define	TCPCTL_MAXID		26
>
>   #define	TCPCTL_NAMES { \
>   	{ 0, 0 }, \
> --- 559,566 ----
>   #endif
>   #define	TCPCTL_RSTPPSLIMIT	24	/* RST pps limit */
>   #define	TCPCTL_DELACK_TICKS	25	/* # ticks to delay ACK */
> ! #define TCPCTL_MSS_MAXADV	26	/* Maximum advertised MSS */
> ! #define	TCPCTL_MAXID		27
>
>   #define	TCPCTL_NAMES { \
>   	{ 0, 0 }, \
> ***************
> *** 588,593 ****
> --- 589,595 ----
>   	{ 0, 0 }, \
>   	{ "rstppslimit", CTLTYPE_INT }, \
>   	{ "delack_ticks", CTLTYPE_INT }, \
> + 	{ "mss_maxadv", CTLTYPE_INT }, \
>   }
>
>   #ifdef _KERNEL
> ***************
> *** 602,608 ****
>   extern	int tcp_do_win_scale;	/* RFC1323 window scaling enabled/disabled? */
>   extern	int tcp_do_timestamps;	/* RFC1323 timestamps enabled/disabled? */
>   extern	int tcp_do_newreno;	/* Use the New Reno algorithms */
> ! extern	int tcp_mssdflt;	/* default seg size */
>   extern	int tcp_init_win;	/* initial window */
>   extern	int tcp_mss_ifmtu;	/* take MSS from interface, not in_maxmtu */
>   extern	int tcp_compat_42;	/* work around ancient broken TCP peers */
> --- 604,611 ----
>   extern	int tcp_do_win_scale;	/* RFC1323 window scaling enabled/disabled? */
>   extern	int tcp_do_timestamps;	/* RFC1323 timestamps enabled/disabled? */
>   extern	int tcp_do_newreno;	/* Use the New Reno algorithms */
> ! extern	int tcp_mssdflt;	/* default maximum seg size */
> ! extern	int tcp_mss_maxadv;	/* maximum advertised mss */
>   extern	int tcp_init_win;	/* initial window */
>   extern	int tcp_mss_ifmtu;	/* take MSS from interface, not in_maxmtu */
>   extern	int tcp_compat_42;	/* work around ancient broken TCP peers */
> ***************
> *** 646,651 ****
> --- 649,655 ----
>   	{ 0 },					\
>   	{ 1, 0, &tcp_rst_ppslim },		\
>   	{ 1, 0, &tcp_delack_ticks },		\
> + 	{ 1, 0, &tcp_mss_maxadv },		\
>   }
>
>   int	 tcp_attach __P((struct socket *));
> Index: sys/netinet/tcp_subr.c
> ===================================================================
> RCS file: /cvsroot/syssrc/sys/netinet/tcp_subr.c,v
> retrieving revision 1.122
> diff -c -r1.122 tcp_subr.c
> *** tcp_subr.c	2001/11/13 00:32:41	1.122
> --- tcp_subr.c	2001/12/02 19:06:04
> ***************
> *** 165,170 ****
> --- 165,171 ----
>
>   /* patchable/settable parameters for tcp */
>   int 	tcp_mssdflt = TCP_MSS;
> + int	tcp_mss_maxadv = 0;
>   int 	tcp_rttdflt = TCPTV_SRTTDFLT / PR_SLOWHZ;
>   int	tcp_do_rfc1323 = 1;	/* window scaling / timestamps (obsolete) */
>   #if NRND > 0
> ***************
> *** 1573,1579 ****
> --- 1574,1588 ----
>   	if (mss > hdrsiz)
>   		mss -= hdrsiz;
>
> + 	/* If a maximum MSS is set, don't advertise above it.
> + 	 */
> + 	if( tcp_mss_maxadv > 0 )
> + 	    mss = min(tcp_mss_maxadv, mss);
> +
> + 	/* Advertise atleast the default MSS (is there a reason for this?)
> + 	 */
>   	mss = max(tcp_mssdflt, mss);
> +
>   	return (mss);
>   }
>
> Index: sbin/sysctl/sysctl.8
> ===================================================================
> RCS file: /cvsroot/basesrc/sbin/sysctl/sysctl.8,v
> retrieving revision 1.70
> diff -c -r1.70 sysctl.8
> *** sysctl.8	2001/10/30 07:28:22	1.70
> --- sysctl.8	2001/12/02 19:06:05
> ***************
> *** 252,257 ****
> --- 252,258 ----
>   .It net.inet.tcp.keepintvl	integer	yes
>   .It net.inet.tcp.log_refused	integer	yes
>   .It net.inet.tcp.mss_ifmtu	integer	yes
> + .It net.inet.tcp.mss_maxadv	integer	yes
>   .It net.inet.tcp.mssdflt	integer	yes
>   .It net.inet.tcp.recvspace	integer	yes
>   .It net.inet.tcp.rfc1323	integer	yes
> Index: lib/libc/gen/sysctl.3
> ===================================================================
> RCS file: /cvsroot/basesrc/lib/libc/gen/sysctl.3,v
> retrieving revision 1.80
> diff -c -r1.80 sysctl.3
> *** sysctl.3	2001/10/30 06:43:21	1.80
> --- sysctl.3	2001/12/02 19:06:08
> ***************
> *** 705,710 ****
> --- 705,711 ----
>   .It tcp	newreno	integer	yes
>   .It tcp	log_refused	integer	yes
>   .It tcp	rstppslimit	integer	yes
> + .It tcp	mss_maxadv	integer	yes
>   .It udp	checksum	integer	yes
>   .It udp	sendspace	integer	yes
>   .It udp	recvspace	integer	yes
> ***************
> *** 796,801 ****
> --- 797,808 ----
>   and to use when the peer does not advertize a maximum segment size to
>   us during connection setup.  Do not change this value unless you really
>   know what you are doing.
> + .It Li tcp.mss_maxadv
> + Returns the maximum TCP MSS that will be advertized to peers, or 0 if there
> + is no limit.  Usually there is no need to change this setting, but it can
> + be usefull for working around buggy networks which suffer from the  PMTU
> + blackhole problem.  This setting does not affect the segment size for
> + outgoing packets.
>   .It Li tcp.syn_cache_limit
>   Returns the maximum number of entries allowed in the TCP compressed state
>   engine.
>
>
>
>