Pass the number of segments coalesced by LRO up the stack by repurposing the

tso_segsz pkthdr field during RX processing, and use the information in TCP for
more correct accounting and as a congestion control input. This is only a start,
and an audit of other uses for the data is left as future work.

Reviewed by:	gallatin, rrs
Sponsored by:	Netflix, Inc.
Differential Revision:	https://reviews.freebsd.org/D7564

sys/netinet/cc/cc.h

@@ -86,6 +86,7 @@ struct cc_var {
 		struct tcpcb		*tcp;
 		struct sctp_nets	*sctp;
 	} ccvc;
+	uint16_t	nsegs; /* # segments coalesced into current chain. */
 };
 
 /* cc_var flags. */

sys/netinet/cc/cc_newreno.c

@@ -137,7 +137,8 @@ newreno_ack_received(struct cc_var *ccv, uint16_t type)
 			 */
 			if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max))
 				incr = min(ccv->bytes_this_ack,
-				    V_tcp_abc_l_var * CCV(ccv, t_maxseg));
+				    ccv->nsegs * V_tcp_abc_l_var *
+				    CCV(ccv, t_maxseg));
 			else
 				incr = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg));
 		}

sys/netinet/tcp_input.c

@@ -300,10 +300,12 @@ hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to)
  * CC wrapper hook functions
  */
 void
-cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t type)
+cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs,
+    uint16_t type)
 {
 	INP_WLOCK_ASSERT(tp->t_inpcb);
 
+	tp->ccv->nsegs = nsegs;
 	tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th);
 	if (tp->snd_cwnd <= tp->snd_wnd)
 		tp->ccv->flags |= CCF_CWND_LIMITED;
@@ -313,7 +315,7 @@ cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t type)
 	if (type == CC_ACK) {
 		if (tp->snd_cwnd > tp->snd_ssthresh) {
 			tp->t_bytes_acked += min(tp->ccv->bytes_this_ack,
-			     V_tcp_abc_l_var * tcp_maxseg(tp));
+			     nsegs * V_tcp_abc_l_var * tcp_maxseg(tp));
 			if (tp->t_bytes_acked >= tp->snd_cwnd) {
 				tp->t_bytes_acked -= tp->snd_cwnd;
 				tp->ccv->flags |= CCF_ABC_SENTAWND;
@@ -1502,6 +1504,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 	int thflags, acked, ourfinisacked, needoutput = 0, sack_changed;
 	int rstreason, todrop, win;
 	u_long tiwin;
+	uint16_t nsegs;
 	char *s;
 	struct in_conninfo *inc;
 	struct mbuf *mfree;
@@ -1521,6 +1524,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 	inc = &tp->t_inpcb->inp_inc;
 	tp->sackhint.last_sack_ack = 0;
 	sack_changed = 0;
+	nsegs = max(1, m->m_pkthdr.lro_nsegs);
 
 	/*
 	 * If this is either a state-changing packet or current state isn't
@@ -1759,7 +1763,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 				/* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */
 				hhook_run_tcp_est_in(tp, th, &to);
 
-				TCPSTAT_INC(tcps_rcvackpack);
+				TCPSTAT_ADD(tcps_rcvackpack, nsegs);
 				TCPSTAT_ADD(tcps_rcvackbyte, acked);
 				sbdrop(&so->so_snd, acked);
 				if (SEQ_GT(tp->snd_una, tp->snd_recover) &&
@@ -1772,7 +1776,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 				 * typically means increasing the congestion
 				 * window.
 				 */
-				cc_ack_received(tp, th, CC_ACK);
+				cc_ack_received(tp, th, nsegs, CC_ACK);
 
 				tp->snd_una = th->th_ack;
 				/*
@@ -1838,7 +1842,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 			 * rcv_nxt.
 			 */
 			tp->rcv_up = tp->rcv_nxt;
-			TCPSTAT_INC(tcps_rcvpack);
+			TCPSTAT_ADD(tcps_rcvpack, nsegs);
 			TCPSTAT_ADD(tcps_rcvbyte, tlen);
 #ifdef TCPDEBUG
 			if (so->so_options & SO_DEBUG)
@@ -2570,7 +2574,8 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 					tp->t_dupacks = 0;
 				else if (++tp->t_dupacks > tcprexmtthresh ||
 				     IN_FASTRECOVERY(tp->t_flags)) {
-					cc_ack_received(tp, th, CC_DUPACK);
+					cc_ack_received(tp, th, nsegs,
+					    CC_DUPACK);
 					if ((tp->t_flags & TF_SACK_PERMIT) &&
 					    IN_FASTRECOVERY(tp->t_flags)) {
 						int awnd;
@@ -2620,7 +2625,8 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 					}
 					/* Congestion signal before ack. */
 					cc_cong_signal(tp, th, CC_NDUPACK);
-					cc_ack_received(tp, th, CC_DUPACK);
+					cc_ack_received(tp, th, nsegs,
+					    CC_DUPACK);
 					tcp_timer_activate(tp, TT_REXMT, 0);
 					tp->t_rtttime = 0;
 					if (tp->t_flags & TF_SACK_PERMIT) {
@@ -2654,7 +2660,8 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 					 * segment. Restore the original
 					 * snd_cwnd after packet transmission.
 					 */
-					cc_ack_received(tp, th, CC_DUPACK);
+					cc_ack_received(tp, th, nsegs,
+					    CC_DUPACK);
 					u_long oldcwnd = tp->snd_cwnd;
 					tcp_seq oldsndmax = tp->snd_max;
 					u_int sent;
@@ -2756,7 +2763,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 		KASSERT(acked >= 0, ("%s: acked unexepectedly negative "
 		    "(tp->snd_una=%u, th->th_ack=%u, tp=%p, m=%p)", __func__,
 		    tp->snd_una, th->th_ack, tp, m));
-		TCPSTAT_INC(tcps_rcvackpack);
+		TCPSTAT_ADD(tcps_rcvackpack, nsegs);
 		TCPSTAT_ADD(tcps_rcvackbyte, acked);
 
 		/*
@@ -2821,7 +2828,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
 		 * control related information. This typically means increasing
 		 * the congestion window.
 		 */
-		cc_ack_received(tp, th, CC_ACK);
+		cc_ack_received(tp, th, nsegs, CC_ACK);
 
 		SOCKBUF_LOCK(&so->so_snd);
 		if (acked > sbavail(&so->so_snd)) {

sys/netinet/tcp_lro.c

@@ -392,6 +392,7 @@ tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le)
 #endif
 	}
 
+	le->m_head->m_pkthdr.lro_nsegs = le->append_cnt + 1;
 	(*lc->ifp->if_input)(lc->ifp, le->m_head);
 	lc->lro_queued += le->append_cnt + 1;
 	lc->lro_flushed++;

sys/netinet/tcp_var.h

@@ -784,7 +784,7 @@ void	tcp_pulloutofband(struct socket *,
 void	tcp_xmit_timer(struct tcpcb *, int);
 void	tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *);
 void	cc_ack_received(struct tcpcb *tp, struct tcphdr *th,
-			    uint16_t type);
+			    uint16_t nsegs, uint16_t type);
 void 	cc_conn_init(struct tcpcb *tp);
 void 	cc_post_recovery(struct tcpcb *tp, struct tcphdr *th);
 void	cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type);

sys/sys/mbuf.h

@@ -174,6 +174,7 @@ struct pkthdr {
 #define	PH_vt		PH_per
 #define	vt_nrecs	sixteen[0]
 #define	tso_segsz	PH_per.sixteen[1]
+#define	lro_nsegs	tso_segsz
 #define	csum_phsum	PH_per.sixteen[2]
 #define	csum_data	PH_per.thirtytwo[1]