tcp: properly handle stretch acks in slow start

Slow start now increases cwnd by 1 if an ACK acknowledges some packets,
regardless the number of packets. Consequently slow start performance
is highly dependent on the degree of the stretch ACKs caused by
receiver or network ACK compression mechanisms (e.g., delayed-ACK,
GRO, etc).  But slow start algorithm is to send twice the amount of
packets of packets left so it should process a stretch ACK of degree
N as if N ACKs of degree 1, then exits when cwnd exceeds ssthresh. A
follow up patch will use the remainder of the N (if greater than 1)
to adjust cwnd in the congestion avoidance phase.

In addition this patch retires the experimental limited slow start
(LSS) feature. LSS has multiple drawbacks but questionable benefit. The
fractional cwnd increase in LSS requires a loop in slow start even
though it's rarely used. Configuring such an increase step via a global
sysctl on different BDPS seems hard. Finally and most importantly the
slow start overshoot concern is now better covered by the Hybrid slow
start (hystart) enabled by default.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

14 files changed, 56 insertions, 67 deletions

diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c
index d5b1390eebbe..3d69ec8dac57 100644
--- a/net/ipv4/sysctl_net_ipv4.c
+++ b/net/ipv4/sysctl_net_ipv4.c
@@ -701,13 +701,6 @@ static struct ctl_table ipv4_table[] = {
 		.proc_handler   = proc_allowed_congestion_control,
 	},
 	{
-		.procname	= "tcp_max_ssthresh",
-		.data		= &sysctl_tcp_max_ssthresh,
-		.maxlen		= sizeof(int),
-		.mode		= 0644,
-		.proc_handler	= proc_dointvec,
-	},
-	{
 		.procname       = "tcp_thin_linear_timeouts",
 		.data           = &sysctl_tcp_thin_linear_timeouts,
 		.maxlen         = sizeof(int),
diff --git a/net/ipv4/tcp_bic.c b/net/ipv4/tcp_bic.c
index f45e1c242440..821846fb0a7e 100644
--- a/net/ipv4/tcp_bic.c
+++ b/net/ipv4/tcp_bic.c
@@ -140,7 +140,8 @@ static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 		ca->cnt = 1;
 }
 
-static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+			      u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct bictcp *ca = inet_csk_ca(sk);
@@ -149,7 +150,7 @@ static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		return;
 
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	else {
 		bictcp_update(ca, tp->snd_cwnd);
 		tcp_cong_avoid_ai(tp, ca->cnt);
diff --git a/net/ipv4/tcp_cong.c b/net/ipv4/tcp_cong.c
index 019c2389a341..ad37bf18ae4b 100644
--- a/net/ipv4/tcp_cong.c
+++ b/net/ipv4/tcp_cong.c
@@ -15,8 +15,6 @@
 #include <linux/gfp.h>
 #include <net/tcp.h>
 
-int sysctl_tcp_max_ssthresh = 0;
-
 static DEFINE_SPINLOCK(tcp_cong_list_lock);
 static LIST_HEAD(tcp_cong_list);
 
@@ -299,35 +297,24 @@ bool tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
 }
 EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
 
-/*
- * Slow start is used when congestion window is less than slow start
- * threshold. This version implements the basic RFC2581 version
- * and optionally supports:
- * 	RFC3742 Limited Slow Start  	  - growth limited to max_ssthresh
- *	RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
+/* Slow start is used when congestion window is no greater than the slow start
+ * threshold. We base on RFC2581 and also handle stretch ACKs properly.
+ * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
+ * something better;) a packet is only considered (s)acked in its entirety to
+ * defend the ACK attacks described in the RFC. Slow start processes a stretch
+ * ACK of degree N as if N acks of degree 1 are received back to back except
+ * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
+ * returns the leftover acks to adjust cwnd in congestion avoidance mode.
  */
-void tcp_slow_start(struct tcp_sock *tp)
+int tcp_slow_start(struct tcp_sock *tp, u32 acked)
 {
-	int cnt; /* increase in packets */
-	unsigned int delta = 0;
-	u32 snd_cwnd = tp->snd_cwnd;
-
-	if (unlikely(!snd_cwnd)) {
-		pr_err_once("snd_cwnd is nul, please report this bug.\n");
-		snd_cwnd = 1U;
-	}
+	u32 cwnd = tp->snd_cwnd + acked;
 
-	if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
-		cnt = sysctl_tcp_max_ssthresh >> 1;	/* limited slow start */
-	else
-		cnt = snd_cwnd;				/* exponential increase */
-
-	tp->snd_cwnd_cnt += cnt;
-	while (tp->snd_cwnd_cnt >= snd_cwnd) {
-		tp->snd_cwnd_cnt -= snd_cwnd;
-		delta++;
-	}
-	tp->snd_cwnd = min(snd_cwnd + delta, tp->snd_cwnd_clamp);
+	if (cwnd > tp->snd_ssthresh)
+		cwnd = tp->snd_ssthresh + 1;
+	acked -= cwnd - tp->snd_cwnd;
+	tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
+	return acked;
 }
 EXPORT_SYMBOL_GPL(tcp_slow_start);
 
@@ -351,7 +338,7 @@ EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
 /* This is Jacobson's slow start and congestion avoidance.
  * SIGCOMM '88, p. 328.
  */
-void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 
@@ -360,7 +347,7 @@ void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 
 	/* In "safe" area, increase. */
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	/* In dangerous area, increase slowly. */
 	else
 		tcp_cong_avoid_ai(tp, tp->snd_cwnd);
diff --git a/net/ipv4/tcp_cubic.c b/net/ipv4/tcp_cubic.c
index b6ae92a51f58..828e4c3ffbaf 100644
--- a/net/ipv4/tcp_cubic.c
+++ b/net/ipv4/tcp_cubic.c
@@ -304,7 +304,8 @@ static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 		ca->cnt = 1;
 }
 
-static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+			      u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct bictcp *ca = inet_csk_ca(sk);
@@ -315,7 +316,7 @@ static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 	if (tp->snd_cwnd <= tp->snd_ssthresh) {
 		if (hystart && after(ack, ca->end_seq))
 			bictcp_hystart_reset(sk);
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	} else {
 		bictcp_update(ca, tp->snd_cwnd);
 		tcp_cong_avoid_ai(tp, ca->cnt);
diff --git a/net/ipv4/tcp_highspeed.c b/net/ipv4/tcp_highspeed.c
index 30f27f6b3655..8ed9305dfdf4 100644
--- a/net/ipv4/tcp_highspeed.c
+++ b/net/ipv4/tcp_highspeed.c
@@ -109,7 +109,7 @@ static void hstcp_init(struct sock *sk)
 	tp->snd_cwnd_clamp = min_t(u32, tp->snd_cwnd_clamp, 0xffffffff/128);
 }
 
-static void hstcp_cong_avoid(struct sock *sk, u32 adk, u32 in_flight)
+static void hstcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct hstcp *ca = inet_csk_ca(sk);
@@ -118,7 +118,7 @@ static void hstcp_cong_avoid(struct sock *sk, u32 adk, u32 in_flight)
 		return;
 
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	else {
 		/* Update AIMD parameters.
 		 *
diff --git a/net/ipv4/tcp_htcp.c b/net/ipv4/tcp_htcp.c
index c1a8175361e8..4a194acfd923 100644
--- a/net/ipv4/tcp_htcp.c
+++ b/net/ipv4/tcp_htcp.c
@@ -227,7 +227,7 @@ static u32 htcp_recalc_ssthresh(struct sock *sk)
 	return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
 }
 
-static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct htcp *ca = inet_csk_ca(sk);
@@ -236,7 +236,7 @@ static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		return;
 
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	else {
 		/* In dangerous area, increase slowly.
 		 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
diff --git a/net/ipv4/tcp_hybla.c b/net/ipv4/tcp_hybla.c
index 57bdd17dff4d..478fe82611bf 100644
--- a/net/ipv4/tcp_hybla.c
+++ b/net/ipv4/tcp_hybla.c
@@ -85,7 +85,8 @@ static inline u32 hybla_fraction(u32 odds)
  *     o Give cwnd a new value based on the model proposed
  *     o remember increments <1
  */
-static void hybla_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void hybla_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+			     u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct hybla *ca = inet_csk_ca(sk);
@@ -102,7 +103,7 @@ static void hybla_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		return;
 
 	if (!ca->hybla_en) {
-		tcp_reno_cong_avoid(sk, ack, in_flight);
+		tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 		return;
 	}
 
diff --git a/net/ipv4/tcp_illinois.c b/net/ipv4/tcp_illinois.c
index 834857f3c871..8a520996f3d2 100644
--- a/net/ipv4/tcp_illinois.c
+++ b/net/ipv4/tcp_illinois.c
@@ -256,7 +256,8 @@ static void tcp_illinois_state(struct sock *sk, u8 new_state)
 /*
  * Increase window in response to successful acknowledgment.
  */
-static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+				    u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct illinois *ca = inet_csk_ca(sk);
@@ -270,7 +271,7 @@ static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 
 	/* In slow start */
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 
 	else {
 		u32 delta;
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 63095b218b4a..c53b7f35c51d 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -2934,10 +2934,10 @@ static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
 		tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1);
 }
 
-static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 {
 	const struct inet_connection_sock *icsk = inet_csk(sk);
-	icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight);
+	icsk->icsk_ca_ops->cong_avoid(sk, ack, acked, in_flight);
 	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
 }
 
@@ -3454,7 +3454,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 
 	/* Advance cwnd if state allows */
 	if (tcp_may_raise_cwnd(sk, flag))
-		tcp_cong_avoid(sk, ack, prior_in_flight);
+		tcp_cong_avoid(sk, ack, acked, prior_in_flight);
 
 	if (tcp_ack_is_dubious(sk, flag)) {
 		is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
diff --git a/net/ipv4/tcp_lp.c b/net/ipv4/tcp_lp.c
index 72f7218b03f5..991d62a2f9bb 100644
--- a/net/ipv4/tcp_lp.c
+++ b/net/ipv4/tcp_lp.c
@@ -115,12 +115,13 @@ static void tcp_lp_init(struct sock *sk)
  * Will only call newReno CA when away from inference.
  * From TCP-LP's paper, this will be handled in additive increasement.
  */
-static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_lp_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+			      u32 in_flight)
 {
 	struct lp *lp = inet_csk_ca(sk);
 
 	if (!(lp->flag & LP_WITHIN_INF))
-		tcp_reno_cong_avoid(sk, ack, in_flight);
+		tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 }
 
 /**
diff --git a/net/ipv4/tcp_scalable.c b/net/ipv4/tcp_scalable.c
index 8ce55b8aaec8..19ea6c2951f3 100644
--- a/net/ipv4/tcp_scalable.c
+++ b/net/ipv4/tcp_scalable.c
@@ -15,7 +15,8 @@
 #define TCP_SCALABLE_AI_CNT	50U
 #define TCP_SCALABLE_MD_SCALE	3
 
-static void tcp_scalable_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_scalable_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+				    u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 
@@ -23,7 +24,7 @@ static void tcp_scalable_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		return;
 
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 	else
 		tcp_cong_avoid_ai(tp, min(tp->snd_cwnd, TCP_SCALABLE_AI_CNT));
 }
diff --git a/net/ipv4/tcp_vegas.c b/net/ipv4/tcp_vegas.c
index 80fa2bfd7ede..06cae62bf208 100644
--- a/net/ipv4/tcp_vegas.c
+++ b/net/ipv4/tcp_vegas.c
@@ -163,13 +163,14 @@ static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp)
 	return  min(tp->snd_ssthresh, tp->snd_cwnd-1);
 }
 
-static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+				 u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct vegas *vegas = inet_csk_ca(sk);
 
 	if (!vegas->doing_vegas_now) {
-		tcp_reno_cong_avoid(sk, ack, in_flight);
+		tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 		return;
 	}
 
@@ -194,7 +195,7 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 			/* We don't have enough RTT samples to do the Vegas
 			 * calculation, so we'll behave like Reno.
 			 */
-			tcp_reno_cong_avoid(sk, ack, in_flight);
+			tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 		} else {
 			u32 rtt, diff;
 			u64 target_cwnd;
@@ -243,7 +244,7 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 
 			} else if (tp->snd_cwnd <= tp->snd_ssthresh) {
 				/* Slow start.  */
-				tcp_slow_start(tp);
+				tcp_slow_start(tp, acked);
 			} else {
 				/* Congestion avoidance. */
 
@@ -283,7 +284,7 @@ static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 	}
 	/* Use normal slow start */
 	else if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 
 }
 
diff --git a/net/ipv4/tcp_veno.c b/net/ipv4/tcp_veno.c
index ac43cd747bce..326475a94865 100644
--- a/net/ipv4/tcp_veno.c
+++ b/net/ipv4/tcp_veno.c
@@ -114,13 +114,14 @@ static void tcp_veno_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 		tcp_veno_init(sk);
 }
 
-static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+				u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct veno *veno = inet_csk_ca(sk);
 
 	if (!veno->doing_veno_now) {
-		tcp_reno_cong_avoid(sk, ack, in_flight);
+		tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 		return;
 	}
 
@@ -133,7 +134,7 @@ static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		/* We don't have enough rtt samples to do the Veno
 		 * calculation, so we'll behave like Reno.
 		 */
-		tcp_reno_cong_avoid(sk, ack, in_flight);
+		tcp_reno_cong_avoid(sk, ack, acked, in_flight);
 	} else {
 		u64 target_cwnd;
 		u32 rtt;
@@ -152,7 +153,7 @@ static void tcp_veno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 
 		if (tp->snd_cwnd <= tp->snd_ssthresh) {
 			/* Slow start.  */
-			tcp_slow_start(tp);
+			tcp_slow_start(tp, acked);
 		} else {
 			/* Congestion avoidance. */
 			if (veno->diff < beta) {
diff --git a/net/ipv4/tcp_yeah.c b/net/ipv4/tcp_yeah.c
index 05c3b6f0e8e1..a347a078ee07 100644
--- a/net/ipv4/tcp_yeah.c
+++ b/net/ipv4/tcp_yeah.c
@@ -69,7 +69,8 @@ static void tcp_yeah_pkts_acked(struct sock *sk, u32 pkts_acked, s32 rtt_us)
 	tcp_vegas_pkts_acked(sk, pkts_acked, rtt_us);
 }
 
-static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
+static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 acked,
+				u32 in_flight)
 {
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct yeah *yeah = inet_csk_ca(sk);
@@ -78,7 +79,7 @@ static void tcp_yeah_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 		return;
 
 	if (tp->snd_cwnd <= tp->snd_ssthresh)
-		tcp_slow_start(tp);
+		tcp_slow_start(tp, acked);
 
 	else if (!yeah->doing_reno_now) {
 		/* Scalable */