Issues with TCP Timestamps allocation

[Paul]

Hi team,

Recently we had an upgrade to 12 Stable. Immediately after, we have started 
seeing some strange connection establishment timeouts to some fixed number
of external (world) hosts. The issue was persistent and easy to reproduce.
Thanks to a patience and dedication of our system engineer we have tracked  
this issue down to a specific commit:

https://svnweb.freebsd.org/base?view=revision&revision=338053

This patch was also back-ported into 11 Stable:

https://svnweb.freebsd.org/base?view=revision&revision=348435

Among other things this patch changes the timestamp allocation strategy,
by introducing a deterministic randomness via a hash function that takes
into account a random key as well as source address, source port, dest
address and dest port. As the result, timestamp offsets of different
tuples (SA,SP,DA,DP) will be wildly different and will jump from small 
to large numbers and back, as long as something in the tuple changes.

After performing various tests of hosts that produce the above mentioned 
issue we came to conclusion that there are some interesting implementations 
that drop SYN packets with timestamps smaller  than the largest timestamp 
value from streams of all recent or current connections from a specific 
address. This looks as some kind of SYN flood protection.

To ensure that each external host is not going to see a wild jumps of 
timestamp values I propose a patch that removes ports from the equation
all together, when calculating the timestamp offset:

Index: sys/netinet/tcp_subr.c
===================================================================
--- sys/netinet/tcp_subr.c	(revision 348435)
+++ sys/netinet/tcp_subr.c	(working copy)
@@ -2224,7 +2224,22 @@
 uint32_t
 tcp_new_ts_offset(struct in_conninfo *inc)
 {
-	return (tcp_keyed_hash(inc, V_ts_offset_secret));
+        /* 
+         * Some implementations show a strange behaviour when a wildly random 
+         * timestamps allocated for different streams. It seems that only the
+         * SYN packets are affected. Observed implementations drop SYN packets
+         * with timestamps smaller than the largest timestamp value of all 
+         * recent or current connections from specific a address. To mitigate 
+         * this we are going to ensure that each host will always observe 
+         * timestamps as increasing no matter the stream: by dropping ports
+         * from the equation.
+         */ 
+        struct in_conninfo inc_copy = *inc;
+
+        inc_copy.inc_fport = 0;
+        inc_copy.inc_lport = 0;
+
+	return (tcp_keyed_hash(&inc_copy, V_ts_offset_secret));
 }
 
 /*

In any case, the solution of the uptime leak, implemented in rev338053 is 
not going to suffer, because a supposed attacker is currently able to use 
any fixed values of SP and DP, albeit not 0, anyway, to remove them out 
of the equation.

There is the list of example hosts that we were able to reproduce the 
issue with:

curl -v http://88.99.60.171:80
curl -v http://163.172.71.252:80
curl -v http://5.9.242.150:80
curl -v https://185.134.205.105:443
curl -v https://136.243.1.231:443
curl -v https://144.76.196.4:443
curl -v http://94.127.191.194:80

To reproduce, call curl repeatedly with a same URL some number of times. 
You are going  to see some of the requests stuck in 
`*    Trying XXX.XXX.XXX.XXX...`

For some reason, the easiest way to reproduce the issue is with nc:

$ echo "foooooo" | nc -v 88.99.60.171 80

Only a few such calls are required until one of them is stuck on connect():
issuing SYN packets with an exponential backoff.