tcp: track data delivery rate for a TCP connection

This patch generates data delivery rate (throughput) samples on a
per-ACK basis. These rate samples can be used by congestion control
modules, and specifically will be used by TCP BBR in later patches in
this series.

Key state:

tp->delivered: Tracks the total number of data packets (original or not)
	       delivered so far. This is an already-existing field.

tp->delivered_mstamp: the last time tp->delivered was updated.

Algorithm:

A rate sample is calculated as (d1 - d0)/(t1 - t0) on a per-ACK basis:

  d1: the current tp->delivered after processing the ACK
  t1: the current time after processing the ACK

  d0: the prior tp->delivered when the acked skb was transmitted
  t0: the prior tp->delivered_mstamp when the acked skb was transmitted

When an skb is transmitted, we snapshot d0 and t0 in its control
block in tcp_rate_skb_sent().

When an ACK arrives, it may SACK and ACK some skbs. For each SACKed
or ACKed skb, tcp_rate_skb_delivered() updates the rate_sample struct
to reflect the latest (d0, t0).

Finally, tcp_rate_gen() generates a rate sample by storing
(d1 - d0) in rs->delivered and (t1 - t0) in rs->interval_us.

One caveat: if an skb was sent with no packets in flight, then
tp->delivered_mstamp may be either invalid (if the connection is
starting) or outdated (if the connection was idle). In that case,
we'll re-stamp tp->delivered_mstamp.

At first glance it seems t0 should always be the time when an skb was
transmitted, but actually this could over-estimate the rate due to
phase mismatch between transmit and ACK events. To track the delivery
rate, we ensure that if packets are in flight then t0 and and t1 are
times at which packets were marked delivered.

If the initial and final RTTs are different then one may be corrupted
by some sort of noise. The noise we see most often is sending gaps
caused by delayed, compressed, or stretched acks. This either affects
both RTTs equally or artificially reduces the final RTT. We approach
this by recording the info we need to compute the initial RTT
(duration of the "send phase" of the window) when we recorded the
associated inflight. Then, for a filter to avoid bandwidth
overestimates, we generalize the per-sample bandwidth computation
from:

    bw = delivered / ack_phase_rtt

to the following:

    bw = delivered / max(send_phase_rtt, ack_phase_rtt)

In large-scale experiments, this filtering approach incorporating
send_phase_rtt is effective at avoiding bandwidth overestimates due to
ACK compression or stretched ACKs.

Signed-off-by: Van Jacobson <vanj@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Nandita Dukkipati <nanditad@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

include/linux/tcp.h

@@ -268,6 +268,8 @@ struct tcp_sock {
 	u32	prr_out;	/* Total number of pkts sent during Recovery. */
 	u32	delivered;	/* Total data packets delivered incl. rexmits */
 	u32	lost;		/* Total data packets lost incl. rexmits */
+	struct skb_mstamp first_tx_mstamp;  /* start of window send phase */
+	struct skb_mstamp delivered_mstamp; /* time we reached "delivered" */
 
  	u32	rcv_wnd;	/* Current receiver window		*/
 	u32	write_seq;	/* Tail(+1) of data held in tcp send buffer */

include/net/tcp.h

@@ -763,8 +763,14 @@ struct tcp_skb_cb {
 	__u32		ack_seq;	/* Sequence number ACK'd	*/
 	union {
 		struct {
-			/* There is space for up to 20 bytes */
+			/* There is space for up to 24 bytes */
 			__u32 in_flight;/* Bytes in flight when packet sent */
+			/* pkts S/ACKed so far upon tx of skb, incl retrans: */
+			__u32 delivered;
+			/* start of send pipeline phase */
+			struct skb_mstamp first_tx_mstamp;
+			/* when we reached the "delivered" count */
+			struct skb_mstamp delivered_mstamp;
 		} tx;   /* only used for outgoing skbs */
 		union {
 			struct inet_skb_parm	h4;
@@ -860,6 +866,26 @@ struct ack_sample {
 	u32 in_flight;
 };
 
+/* A rate sample measures the number of (original/retransmitted) data
+ * packets delivered "delivered" over an interval of time "interval_us".
+ * The tcp_rate.c code fills in the rate sample, and congestion
+ * control modules that define a cong_control function to run at the end
+ * of ACK processing can optionally chose to consult this sample when
+ * setting cwnd and pacing rate.
+ * A sample is invalid if "delivered" or "interval_us" is negative.
+ */
+struct rate_sample {
+	struct	skb_mstamp prior_mstamp; /* starting timestamp for interval */
+	u32  prior_delivered;	/* tp->delivered at "prior_mstamp" */
+	s32  delivered;		/* number of packets delivered over interval */
+	long interval_us;	/* time for tp->delivered to incr "delivered" */
+	long rtt_us;		/* RTT of last (S)ACKed packet (or -1) */
+	int  losses;		/* number of packets marked lost upon ACK */
+	u32  acked_sacked;	/* number of packets newly (S)ACKed upon ACK */
+	u32  prior_in_flight;	/* in flight before this ACK */
+	bool is_retrans;	/* is sample from retransmission? */
+};
+
 struct tcp_congestion_ops {
 	struct list_head	list;
 	u32 key;
@@ -946,6 +972,13 @@ static inline void tcp_ca_event(struct sock *sk, const enum tcp_ca_event event)
 		icsk->icsk_ca_ops->cwnd_event(sk, event);
 }
 
+/* From tcp_rate.c */
+void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb);
+void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
+			    struct rate_sample *rs);
+void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
+		  struct skb_mstamp *now, struct rate_sample *rs);
+
 /* These functions determine how the current flow behaves in respect of SACK
  * handling. SACK is negotiated with the peer, and therefore it can vary
  * between different flows.

net/ipv4/Makefile

@@ -8,7 +8,7 @@ obj-y     := route.o inetpeer.o protocol.o \
 	     inet_timewait_sock.o inet_connection_sock.o \
 	     tcp.o tcp_input.o tcp_output.o tcp_timer.o tcp_ipv4.o \
 	     tcp_minisocks.o tcp_cong.o tcp_metrics.o tcp_fastopen.o \
-	     tcp_recovery.o \
+	     tcp_rate.o tcp_recovery.o \
 	     tcp_offload.o datagram.o raw.o udp.o udplite.o \
 	     udp_offload.o arp.o icmp.o devinet.o af_inet.o igmp.o \
 	     fib_frontend.o fib_semantics.o fib_trie.o \

net/ipv4/tcp_input.c

@@ -1112,6 +1112,7 @@ struct tcp_sacktag_state {
 	 */
 	struct skb_mstamp first_sackt;
 	struct skb_mstamp last_sackt;
+	struct rate_sample *rate;
 	int	flag;
 };
 
@@ -1279,6 +1280,7 @@ static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
 	tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
 			start_seq, end_seq, dup_sack, pcount,
 			&skb->skb_mstamp);
+	tcp_rate_skb_delivered(sk, skb, state->rate);
 
 	if (skb == tp->lost_skb_hint)
 		tp->lost_cnt_hint += pcount;
@@ -1329,6 +1331,9 @@ static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
 		tcp_advance_highest_sack(sk, skb);
 
 	tcp_skb_collapse_tstamp(prev, skb);
+	if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp.v64))
+		TCP_SKB_CB(prev)->tx.delivered_mstamp.v64 = 0;
+
 	tcp_unlink_write_queue(skb, sk);
 	sk_wmem_free_skb(sk, skb);
 
@@ -1558,6 +1563,7 @@ static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
 						dup_sack,
 						tcp_skb_pcount(skb),
 						&skb->skb_mstamp);
+			tcp_rate_skb_delivered(sk, skb, state->rate);
 
 			if (!before(TCP_SKB_CB(skb)->seq,
 				    tcp_highest_sack_seq(tp)))
@@ -1640,8 +1646,10 @@ tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
 
 	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
 					 num_sacks, prior_snd_una);
-	if (found_dup_sack)
+	if (found_dup_sack) {
 		state->flag |= FLAG_DSACKING_ACK;
+		tp->delivered++; /* A spurious retransmission is delivered */
+	}
 
 	/* Eliminate too old ACKs, but take into
 	 * account more or less fresh ones, they can
@@ -3071,10 +3079,11 @@ static void tcp_ack_tstamp(struct sock *sk, struct sk_buff *skb,
  */
 static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
 			       u32 prior_snd_una, int *acked,
-			       struct tcp_sacktag_state *sack)
+			       struct tcp_sacktag_state *sack,
+			       struct skb_mstamp *now)
 {
 	const struct inet_connection_sock *icsk = inet_csk(sk);
-	struct skb_mstamp first_ackt, last_ackt, now;
+	struct skb_mstamp first_ackt, last_ackt;
 	struct tcp_sock *tp = tcp_sk(sk);
 	u32 prior_sacked = tp->sacked_out;
 	u32 reord = tp->packets_out;
@@ -3106,7 +3115,6 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
 			acked_pcount = tcp_tso_acked(sk, skb);
 			if (!acked_pcount)
 				break;
-
 			fully_acked = false;
 		} else {
 			/* Speedup tcp_unlink_write_queue() and next loop */
@@ -3142,6 +3150,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
 
 		tp->packets_out -= acked_pcount;
 		pkts_acked += acked_pcount;
+		tcp_rate_skb_delivered(sk, skb, sack->rate);
 
 		/* Initial outgoing SYN's get put onto the write_queue
 		 * just like anything else we transmit.  It is not
@@ -3174,16 +3183,15 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
 	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
 		flag |= FLAG_SACK_RENEGING;
 
-	skb_mstamp_get(&now);
 	if (likely(first_ackt.v64) && !(flag & FLAG_RETRANS_DATA_ACKED)) {
-		seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
-		ca_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
+		seq_rtt_us = skb_mstamp_us_delta(now, &first_ackt);
+		ca_rtt_us = skb_mstamp_us_delta(now, &last_ackt);
 	}
 	if (sack->first_sackt.v64) {
-		sack_rtt_us = skb_mstamp_us_delta(&now, &sack->first_sackt);
-		ca_rtt_us = skb_mstamp_us_delta(&now, &sack->last_sackt);
+		sack_rtt_us = skb_mstamp_us_delta(now, &sack->first_sackt);
+		ca_rtt_us = skb_mstamp_us_delta(now, &sack->last_sackt);
 	}
-
+	sack->rate->rtt_us = ca_rtt_us; /* RTT of last (S)ACKed packet, or -1 */
 	rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us,
 					ca_rtt_us);
 
@@ -3211,7 +3219,7 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
 		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
 
 	} else if (skb && rtt_update && sack_rtt_us >= 0 &&
-		   sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
+		   sack_rtt_us > skb_mstamp_us_delta(now, &skb->skb_mstamp)) {
 		/* Do not re-arm RTO if the sack RTT is measured from data sent
 		 * after when the head was last (re)transmitted. Otherwise the
 		 * timeout may continue to extend in loss recovery.
@@ -3548,17 +3556,21 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 	struct inet_connection_sock *icsk = inet_csk(sk);
 	struct tcp_sock *tp = tcp_sk(sk);
 	struct tcp_sacktag_state sack_state;
+	struct rate_sample rs = { .prior_delivered = 0 };
 	u32 prior_snd_una = tp->snd_una;
 	u32 ack_seq = TCP_SKB_CB(skb)->seq;
 	u32 ack = TCP_SKB_CB(skb)->ack_seq;
 	bool is_dupack = false;
 	u32 prior_fackets;
 	int prior_packets = tp->packets_out;
-	u32 prior_delivered = tp->delivered;
+	u32 delivered = tp->delivered;
+	u32 lost = tp->lost;
 	int acked = 0; /* Number of packets newly acked */
 	int rexmit = REXMIT_NONE; /* Flag to (re)transmit to recover losses */
+	struct skb_mstamp now;
 
 	sack_state.first_sackt.v64 = 0;
+	sack_state.rate = &rs;
 
 	/* We very likely will need to access write queue head. */
 	prefetchw(sk->sk_write_queue.next);
@@ -3581,6 +3593,8 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 	if (after(ack, tp->snd_nxt))
 		goto invalid_ack;
 
+	skb_mstamp_get(&now);
+
 	if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
 	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
 		tcp_rearm_rto(sk);
@@ -3591,6 +3605,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 	}
 
 	prior_fackets = tp->fackets_out;
+	rs.prior_in_flight = tcp_packets_in_flight(tp);
 
 	/* ts_recent update must be made after we are sure that the packet
 	 * is in window.
@@ -3646,7 +3661,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 
 	/* See if we can take anything off of the retransmit queue. */
 	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, &acked,
-				    &sack_state);
+				    &sack_state, &now);
 
 	if (tcp_ack_is_dubious(sk, flag)) {
 		is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
@@ -3663,7 +3678,10 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
 
 	if (icsk->icsk_pending == ICSK_TIME_RETRANS)
 		tcp_schedule_loss_probe(sk);
-	tcp_cong_control(sk, ack, tp->delivered - prior_delivered, flag);
+	delivered = tp->delivered - delivered;	/* freshly ACKed or SACKed */
+	lost = tp->lost - lost;			/* freshly marked lost */
+	tcp_rate_gen(sk, delivered, lost, &now, &rs);
+	tcp_cong_control(sk, ack, delivered, flag);
 	tcp_xmit_recovery(sk, rexmit);
 	return 1;
 

net/ipv4/tcp_output.c

@@ -918,6 +918,7 @@ static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
 		skb_mstamp_get(&skb->skb_mstamp);
 		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
 			- tp->snd_una;
+		tcp_rate_skb_sent(sk, skb);
 
 		if (unlikely(skb_cloned(skb)))
 			skb = pskb_copy(skb, gfp_mask);
@@ -1213,6 +1214,9 @@ int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
 	tcp_set_skb_tso_segs(skb, mss_now);
 	tcp_set_skb_tso_segs(buff, mss_now);
 
+	/* Update delivered info for the new segment */
+	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
+
 	/* If this packet has been sent out already, we must
 	 * adjust the various packet counters.
 	 */

net/ipv4/tcp_rate.c

@@ -0,0 +1,149 @@
+#include <net/tcp.h>
+
+/* The bandwidth estimator estimates the rate at which the network
+ * can currently deliver outbound data packets for this flow. At a high
+ * level, it operates by taking a delivery rate sample for each ACK.
+ *
+ * A rate sample records the rate at which the network delivered packets
+ * for this flow, calculated over the time interval between the transmission
+ * of a data packet and the acknowledgment of that packet.
+ *
+ * Specifically, over the interval between each transmit and corresponding ACK,
+ * the estimator generates a delivery rate sample. Typically it uses the rate
+ * at which packets were acknowledged. However, the approach of using only the
+ * acknowledgment rate faces a challenge under the prevalent ACK decimation or
+ * compression: packets can temporarily appear to be delivered much quicker
+ * than the bottleneck rate. Since it is physically impossible to do that in a
+ * sustained fashion, when the estimator notices that the ACK rate is faster
+ * than the transmit rate, it uses the latter:
+ *
+ *    send_rate = #pkts_delivered/(last_snd_time - first_snd_time)
+ *    ack_rate  = #pkts_delivered/(last_ack_time - first_ack_time)
+ *    bw = min(send_rate, ack_rate)
+ *
+ * Notice the estimator essentially estimates the goodput, not always the
+ * network bottleneck link rate when the sending or receiving is limited by
+ * other factors like applications or receiver window limits.  The estimator
+ * deliberately avoids using the inter-packet spacing approach because that
+ * approach requires a large number of samples and sophisticated filtering.
+ */
+
+
+/* Snapshot the current delivery information in the skb, to generate
+ * a rate sample later when the skb is (s)acked in tcp_rate_skb_delivered().
+ */
+void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+
+	 /* In general we need to start delivery rate samples from the
+	  * time we received the most recent ACK, to ensure we include
+	  * the full time the network needs to deliver all in-flight
+	  * packets. If there are no packets in flight yet, then we
+	  * know that any ACKs after now indicate that the network was
+	  * able to deliver those packets completely in the sampling
+	  * interval between now and the next ACK.
+	  *
+	  * Note that we use packets_out instead of tcp_packets_in_flight(tp)
+	  * because the latter is a guess based on RTO and loss-marking
+	  * heuristics. We don't want spurious RTOs or loss markings to cause
+	  * a spuriously small time interval, causing a spuriously high
+	  * bandwidth estimate.
+	  */
+	if (!tp->packets_out) {
+		tp->first_tx_mstamp  = skb->skb_mstamp;
+		tp->delivered_mstamp = skb->skb_mstamp;
+	}
+
+	TCP_SKB_CB(skb)->tx.first_tx_mstamp	= tp->first_tx_mstamp;
+	TCP_SKB_CB(skb)->tx.delivered_mstamp	= tp->delivered_mstamp;
+	TCP_SKB_CB(skb)->tx.delivered		= tp->delivered;
+}
+
+/* When an skb is sacked or acked, we fill in the rate sample with the (prior)
+ * delivery information when the skb was last transmitted.
+ *
+ * If an ACK (s)acks multiple skbs (e.g., stretched-acks), this function is
+ * called multiple times. We favor the information from the most recently
+ * sent skb, i.e., the skb with the highest prior_delivered count.
+ */
+void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
+			    struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
+
+	if (!scb->tx.delivered_mstamp.v64)
+		return;
+
+	if (!rs->prior_delivered ||
+	    after(scb->tx.delivered, rs->prior_delivered)) {
+		rs->prior_delivered  = scb->tx.delivered;
+		rs->prior_mstamp     = scb->tx.delivered_mstamp;
+		rs->is_retrans	     = scb->sacked & TCPCB_RETRANS;
+
+		/* Find the duration of the "send phase" of this window: */
+		rs->interval_us      = skb_mstamp_us_delta(
+						&skb->skb_mstamp,
+						&scb->tx.first_tx_mstamp);
+
+		/* Record send time of most recently ACKed packet: */
+		tp->first_tx_mstamp  = skb->skb_mstamp;
+	}
+	/* Mark off the skb delivered once it's sacked to avoid being
+	 * used again when it's cumulatively acked. For acked packets
+	 * we don't need to reset since it'll be freed soon.
+	 */
+	if (scb->sacked & TCPCB_SACKED_ACKED)
+		scb->tx.delivered_mstamp.v64 = 0;
+}
+
+/* Update the connection delivery information and generate a rate sample. */
+void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
+		  struct skb_mstamp *now, struct rate_sample *rs)
+{
+	struct tcp_sock *tp = tcp_sk(sk);
+	u32 snd_us, ack_us;
+
+	/* TODO: there are multiple places throughout tcp_ack() to get
+	 * current time. Refactor the code using a new "tcp_acktag_state"
+	 * to carry current time, flags, stats like "tcp_sacktag_state".
+	 */
+	if (delivered)
+		tp->delivered_mstamp = *now;
+
+	rs->acked_sacked = delivered;	/* freshly ACKed or SACKed */
+	rs->losses = lost;		/* freshly marked lost */
+	/* Return an invalid sample if no timing information is available. */
+	if (!rs->prior_mstamp.v64) {
+		rs->delivered = -1;
+		rs->interval_us = -1;
+		return;
+	}
+	rs->delivered   = tp->delivered - rs->prior_delivered;
+
+	/* Model sending data and receiving ACKs as separate pipeline phases
+	 * for a window. Usually the ACK phase is longer, but with ACK
+	 * compression the send phase can be longer. To be safe we use the
+	 * longer phase.
+	 */
+	snd_us = rs->interval_us;				/* send phase */
+	ack_us = skb_mstamp_us_delta(now, &rs->prior_mstamp);	/* ack phase */
+	rs->interval_us = max(snd_us, ack_us);
+
+	/* Normally we expect interval_us >= min-rtt.
+	 * Note that rate may still be over-estimated when a spuriously
+	 * retransmistted skb was first (s)acked because "interval_us"
+	 * is under-estimated (up to an RTT). However continuously
+	 * measuring the delivery rate during loss recovery is crucial
+	 * for connections suffer heavy or prolonged losses.
+	 */
+	if (unlikely(rs->interval_us < tcp_min_rtt(tp))) {
+		rs->interval_us = -1;
+		if (!rs->is_retrans)
+			pr_debug("tcp rate: %ld %d %u %u %u\n",
+				 rs->interval_us, rs->delivered,
+				 inet_csk(sk)->icsk_ca_state,
+				 tp->rx_opt.sack_ok, tcp_min_rtt(tp));
+	}
+}