MFC 300981

    tcp: Don't prematurely drop receiving-only connections

    If the connection was persistent and receiving-only, several (12)
    sporadic device insufficient buffers would cause the connection be
    dropped prematurely:
    Upon ENOBUFS in tcp_output() for an ACK, retransmission timer is
    started.  No one will stop this retransmission timer for receiving-
    only connection, so the retransmission timer promises to expire and
    t_rxtshift is promised to be increased.  And t_rxtshift will not be
    reset to 0, since no RTT measurement will be done for receiving-only
    connection.  If this receiving-only connection lived long enough
    (e.g. >350sec, given the RTO starts from 200ms), and it suffered 12
    sporadic device insufficient buffers, i.e. t_rxtshift >= 12, this
    receiving-only connection would be dropped prematurely by the
    retransmission timer.

    We now assert that for data segments, SYNs or FINs either rexmit or
    persist timer was wired upon ENOBUFS.  And don't set rexmit timer
    for other cases, i.e. ENOBUFS upon ACKs.

    Discussed with:     lstewart, hiren, jtl, Mike Karels
    MFC after:  3 weeks
    Sponsored by:       Microsoft OSTC
    Differential Revision:      https://reviews.freebsd.org/D5872

1 files changed, 11 insertions, 3 deletions

diff --git a/sys/netinet/tcp_output.c b/sys/netinet/tcp_output.c
index 57332be..b5781d3 100644
--- a/sys/netinet/tcp_output.c
+++ b/sys/netinet/tcp_output.c
@@ -127,6 +127,16 @@ SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
 	&VNET_NAME(tcp_autosndbuf_max), 0,
 	"Max size of automatic send buffer");
 
+/*
+ * Make sure that either retransmit or persist timer is set for SYN, FIN and
+ * non-ACK.
+ */
+#define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)			\
+	KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
+	    tcp_timer_active((tp), TT_REXMT) ||				\
+	    tcp_timer_active((tp), TT_PERSIST),				\
+	    ("neither rexmt nor persist timer is set"))
+
 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
 			    struct tcphdr *th, struct tcpopt *to,
 			    long len, int tso);
@@ -1531,9 +1541,7 @@ timer:
 			tp->t_softerror = error;
 			return (error);
 		case ENOBUFS:
-	                if (!tcp_timer_active(tp, TT_REXMT) &&
-			    !tcp_timer_active(tp, TT_PERSIST))
-	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
+			TCP_XMIT_TIMER_ASSERT(tp, len, flags);
 			tp->snd_cwnd = tp->t_maxseg;
 			return (0);
 		case EMSGSIZE: