xref: /openbmc/linux/net/ipv4/tcp_rate.c (revision 0a94608f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <net/tcp.h>
3 
4 /* The bandwidth estimator estimates the rate at which the network
5  * can currently deliver outbound data packets for this flow. At a high
6  * level, it operates by taking a delivery rate sample for each ACK.
7  *
8  * A rate sample records the rate at which the network delivered packets
9  * for this flow, calculated over the time interval between the transmission
10  * of a data packet and the acknowledgment of that packet.
11  *
12  * Specifically, over the interval between each transmit and corresponding ACK,
13  * the estimator generates a delivery rate sample. Typically it uses the rate
14  * at which packets were acknowledged. However, the approach of using only the
15  * acknowledgment rate faces a challenge under the prevalent ACK decimation or
16  * compression: packets can temporarily appear to be delivered much quicker
17  * than the bottleneck rate. Since it is physically impossible to do that in a
18  * sustained fashion, when the estimator notices that the ACK rate is faster
19  * than the transmit rate, it uses the latter:
20  *
21  *    send_rate = #pkts_delivered/(last_snd_time - first_snd_time)
22  *    ack_rate  = #pkts_delivered/(last_ack_time - first_ack_time)
23  *    bw = min(send_rate, ack_rate)
24  *
25  * Notice the estimator essentially estimates the goodput, not always the
26  * network bottleneck link rate when the sending or receiving is limited by
27  * other factors like applications or receiver window limits.  The estimator
28  * deliberately avoids using the inter-packet spacing approach because that
29  * approach requires a large number of samples and sophisticated filtering.
30  *
31  * TCP flows can often be application-limited in request/response workloads.
32  * The estimator marks a bandwidth sample as application-limited if there
33  * was some moment during the sampled window of packets when there was no data
34  * ready to send in the write queue.
35  */
36 
37 /* Snapshot the current delivery information in the skb, to generate
38  * a rate sample later when the skb is (s)acked in tcp_rate_skb_delivered().
39  */
40 void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb)
41 {
42 	struct tcp_sock *tp = tcp_sk(sk);
43 
44 	 /* In general we need to start delivery rate samples from the
45 	  * time we received the most recent ACK, to ensure we include
46 	  * the full time the network needs to deliver all in-flight
47 	  * packets. If there are no packets in flight yet, then we
48 	  * know that any ACKs after now indicate that the network was
49 	  * able to deliver those packets completely in the sampling
50 	  * interval between now and the next ACK.
51 	  *
52 	  * Note that we use packets_out instead of tcp_packets_in_flight(tp)
53 	  * because the latter is a guess based on RTO and loss-marking
54 	  * heuristics. We don't want spurious RTOs or loss markings to cause
55 	  * a spuriously small time interval, causing a spuriously high
56 	  * bandwidth estimate.
57 	  */
58 	if (!tp->packets_out) {
59 		u64 tstamp_us = tcp_skb_timestamp_us(skb);
60 
61 		tp->first_tx_mstamp  = tstamp_us;
62 		tp->delivered_mstamp = tstamp_us;
63 	}
64 
65 	TCP_SKB_CB(skb)->tx.first_tx_mstamp	= tp->first_tx_mstamp;
66 	TCP_SKB_CB(skb)->tx.delivered_mstamp	= tp->delivered_mstamp;
67 	TCP_SKB_CB(skb)->tx.delivered		= tp->delivered;
68 	TCP_SKB_CB(skb)->tx.delivered_ce	= tp->delivered_ce;
69 	TCP_SKB_CB(skb)->tx.is_app_limited	= tp->app_limited ? 1 : 0;
70 }
71 
72 /* When an skb is sacked or acked, we fill in the rate sample with the (prior)
73  * delivery information when the skb was last transmitted.
74  *
75  * If an ACK (s)acks multiple skbs (e.g., stretched-acks), this function is
76  * called multiple times. We favor the information from the most recently
77  * sent skb, i.e., the skb with the most recently sent time and the highest
78  * sequence.
79  */
80 void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
81 			    struct rate_sample *rs)
82 {
83 	struct tcp_sock *tp = tcp_sk(sk);
84 	struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
85 	u64 tx_tstamp;
86 
87 	if (!scb->tx.delivered_mstamp)
88 		return;
89 
90 	tx_tstamp = tcp_skb_timestamp_us(skb);
91 	if (!rs->prior_delivered ||
92 	    tcp_skb_sent_after(tx_tstamp, tp->first_tx_mstamp,
93 			       scb->end_seq, rs->last_end_seq)) {
94 		rs->prior_delivered_ce  = scb->tx.delivered_ce;
95 		rs->prior_delivered  = scb->tx.delivered;
96 		rs->prior_mstamp     = scb->tx.delivered_mstamp;
97 		rs->is_app_limited   = scb->tx.is_app_limited;
98 		rs->is_retrans	     = scb->sacked & TCPCB_RETRANS;
99 		rs->last_end_seq     = scb->end_seq;
100 
101 		/* Record send time of most recently ACKed packet: */
102 		tp->first_tx_mstamp  = tx_tstamp;
103 		/* Find the duration of the "send phase" of this window: */
104 		rs->interval_us = tcp_stamp_us_delta(tp->first_tx_mstamp,
105 						     scb->tx.first_tx_mstamp);
106 
107 	}
108 	/* Mark off the skb delivered once it's sacked to avoid being
109 	 * used again when it's cumulatively acked. For acked packets
110 	 * we don't need to reset since it'll be freed soon.
111 	 */
112 	if (scb->sacked & TCPCB_SACKED_ACKED)
113 		scb->tx.delivered_mstamp = 0;
114 }
115 
116 /* Update the connection delivery information and generate a rate sample. */
117 void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
118 		  bool is_sack_reneg, struct rate_sample *rs)
119 {
120 	struct tcp_sock *tp = tcp_sk(sk);
121 	u32 snd_us, ack_us;
122 
123 	/* Clear app limited if bubble is acked and gone. */
124 	if (tp->app_limited && after(tp->delivered, tp->app_limited))
125 		tp->app_limited = 0;
126 
127 	/* TODO: there are multiple places throughout tcp_ack() to get
128 	 * current time. Refactor the code using a new "tcp_acktag_state"
129 	 * to carry current time, flags, stats like "tcp_sacktag_state".
130 	 */
131 	if (delivered)
132 		tp->delivered_mstamp = tp->tcp_mstamp;
133 
134 	rs->acked_sacked = delivered;	/* freshly ACKed or SACKed */
135 	rs->losses = lost;		/* freshly marked lost */
136 	/* Return an invalid sample if no timing information is available or
137 	 * in recovery from loss with SACK reneging. Rate samples taken during
138 	 * a SACK reneging event may overestimate bw by including packets that
139 	 * were SACKed before the reneg.
140 	 */
141 	if (!rs->prior_mstamp || is_sack_reneg) {
142 		rs->delivered = -1;
143 		rs->interval_us = -1;
144 		return;
145 	}
146 	rs->delivered   = tp->delivered - rs->prior_delivered;
147 
148 	rs->delivered_ce = tp->delivered_ce - rs->prior_delivered_ce;
149 	/* delivered_ce occupies less than 32 bits in the skb control block */
150 	rs->delivered_ce &= TCPCB_DELIVERED_CE_MASK;
151 
152 	/* Model sending data and receiving ACKs as separate pipeline phases
153 	 * for a window. Usually the ACK phase is longer, but with ACK
154 	 * compression the send phase can be longer. To be safe we use the
155 	 * longer phase.
156 	 */
157 	snd_us = rs->interval_us;				/* send phase */
158 	ack_us = tcp_stamp_us_delta(tp->tcp_mstamp,
159 				    rs->prior_mstamp); /* ack phase */
160 	rs->interval_us = max(snd_us, ack_us);
161 
162 	/* Record both segment send and ack receive intervals */
163 	rs->snd_interval_us = snd_us;
164 	rs->rcv_interval_us = ack_us;
165 
166 	/* Normally we expect interval_us >= min-rtt.
167 	 * Note that rate may still be over-estimated when a spuriously
168 	 * retransmistted skb was first (s)acked because "interval_us"
169 	 * is under-estimated (up to an RTT). However continuously
170 	 * measuring the delivery rate during loss recovery is crucial
171 	 * for connections suffer heavy or prolonged losses.
172 	 */
173 	if (unlikely(rs->interval_us < tcp_min_rtt(tp))) {
174 		if (!rs->is_retrans)
175 			pr_debug("tcp rate: %ld %d %u %u %u\n",
176 				 rs->interval_us, rs->delivered,
177 				 inet_csk(sk)->icsk_ca_state,
178 				 tp->rx_opt.sack_ok, tcp_min_rtt(tp));
179 		rs->interval_us = -1;
180 		return;
181 	}
182 
183 	/* Record the last non-app-limited or the highest app-limited bw */
184 	if (!rs->is_app_limited ||
185 	    ((u64)rs->delivered * tp->rate_interval_us >=
186 	     (u64)tp->rate_delivered * rs->interval_us)) {
187 		tp->rate_delivered = rs->delivered;
188 		tp->rate_interval_us = rs->interval_us;
189 		tp->rate_app_limited = rs->is_app_limited;
190 	}
191 }
192 
193 /* If a gap is detected between sends, mark the socket application-limited. */
194 void tcp_rate_check_app_limited(struct sock *sk)
195 {
196 	struct tcp_sock *tp = tcp_sk(sk);
197 
198 	if (/* We have less than one packet to send. */
199 	    tp->write_seq - tp->snd_nxt < tp->mss_cache &&
200 	    /* Nothing in sending host's qdisc queues or NIC tx queue. */
201 	    sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) &&
202 	    /* We are not limited by CWND. */
203 	    tcp_packets_in_flight(tp) < tp->snd_cwnd &&
204 	    /* All lost packets have been retransmitted. */
205 	    tp->lost_out <= tp->retrans_out)
206 		tp->app_limited =
207 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
208 }
209 EXPORT_SYMBOL_GPL(tcp_rate_check_app_limited);
210