Dropping TCP options from retransmitted SYNs considered harmful

[Gleb Smirnoff]

On Fri, Oct 12, 2012 at 02:15:04PM -0400, George Neville-Neil wrote:
G> > Back in 2001 FreeBSD added a hack to strip TCP options from retransmitted SYNs 
G> > starting with the 3rd SYN in this block in tcp_timer.c:
G> > 
G> > 	/*
G> > 	 * Disable rfc1323 if we haven't got any response to
G> > 	 * our third SYN to work-around some broken terminal servers
G> > 	 * (most of which have hopefully been retired) that have bad VJ
G> > 	 * header compression code which trashes TCP segments containing
G> > 	 * unknown-to-them TCP options.
G> > 	 */
G> > 	if ((tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3))
G> > 		tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP);
G> > 
G> > There is even a PR for the original bug report: kern/1689
G> > 
G> > However, there is an unintended consequence of this change that can be 
G> > disastrous.  Specifically, suppose you have a FreeBSD client connecting to a 
G> > server, and that the SYNs are arriving at the server successfully, but the 
G> > first few return SYN/ACKs are dropped.  Eventually a SYN/ACK makes it through 
G> > and the connection is established.
G> > 
G> > The server (based on the first SYN it saw) believes it has negotiated window 
G> > scaling with the client.  The client, however, has broken what it promised in 
G> > that first SYN and believes it is not using any window scaling at all.  This 
G> > causes two forms of breakage:
G> > 
G> > 1) When the server advertises a scaled window (e.g. '8' for a 64k window
G> >    scaled at 13), the client thinks it is an unscaled window ('8') and
G> >    sends data to the server very slowly.
G> > 
G> > 2) When the client advertises an unscaled window (e.g. '65535' for a 64k
G> >    window), the server thinks it has a huge window (65535 << 13 == 511MB)
G> >    to send into.
G> > 
G> > I'm not sure that 2) is a problem per se, but I have definitely seen instances 
G> > of 1) (and examined the 'struct tcpcb' in kgdb on both the server and client 
G> > end of the connections to verify they disagreed on the scaling).
G> > 
G> > The original motivation of this change is to work around broken terminal 
G> > servers that were old when this change was added in 2001.  Over 10 years later 
G> > I think we should at least have an option to turn this work-around off, and 
G> > possibly disable it by default.
G> > 
G> > Thoughts?
G> > 
G> 
G> I'm all for taking that code out.

+1

-- 
Totus tuus, Glebius.