xref: /openbmc/linux/net/ipv4/tcp_rate.c (revision 6d99a79c)
1 #include <net/tcp.h>
2 
3 /* The bandwidth estimator estimates the rate at which the network
4  * can currently deliver outbound data packets for this flow. At a high
5  * level, it operates by taking a delivery rate sample for each ACK.
6  *
7  * A rate sample records the rate at which the network delivered packets
8  * for this flow, calculated over the time interval between the transmission
9  * of a data packet and the acknowledgment of that packet.
10  *
11  * Specifically, over the interval between each transmit and corresponding ACK,
12  * the estimator generates a delivery rate sample. Typically it uses the rate
13  * at which packets were acknowledged. However, the approach of using only the
14  * acknowledgment rate faces a challenge under the prevalent ACK decimation or
15  * compression: packets can temporarily appear to be delivered much quicker
16  * than the bottleneck rate. Since it is physically impossible to do that in a
17  * sustained fashion, when the estimator notices that the ACK rate is faster
18  * than the transmit rate, it uses the latter:
19  *
20  *    send_rate = #pkts_delivered/(last_snd_time - first_snd_time)
21  *    ack_rate  = #pkts_delivered/(last_ack_time - first_ack_time)
22  *    bw = min(send_rate, ack_rate)
23  *
24  * Notice the estimator essentially estimates the goodput, not always the
25  * network bottleneck link rate when the sending or receiving is limited by
26  * other factors like applications or receiver window limits.  The estimator
27  * deliberately avoids using the inter-packet spacing approach because that
28  * approach requires a large number of samples and sophisticated filtering.
29  *
30  * TCP flows can often be application-limited in request/response workloads.
31  * The estimator marks a bandwidth sample as application-limited if there
32  * was some moment during the sampled window of packets when there was no data
33  * ready to send in the write queue.
34  */
35 
36 /* Snapshot the current delivery information in the skb, to generate
37  * a rate sample later when the skb is (s)acked in tcp_rate_skb_delivered().
38  */
39 void tcp_rate_skb_sent(struct sock *sk, struct sk_buff *skb)
40 {
41 	struct tcp_sock *tp = tcp_sk(sk);
42 
43 	 /* In general we need to start delivery rate samples from the
44 	  * time we received the most recent ACK, to ensure we include
45 	  * the full time the network needs to deliver all in-flight
46 	  * packets. If there are no packets in flight yet, then we
47 	  * know that any ACKs after now indicate that the network was
48 	  * able to deliver those packets completely in the sampling
49 	  * interval between now and the next ACK.
50 	  *
51 	  * Note that we use packets_out instead of tcp_packets_in_flight(tp)
52 	  * because the latter is a guess based on RTO and loss-marking
53 	  * heuristics. We don't want spurious RTOs or loss markings to cause
54 	  * a spuriously small time interval, causing a spuriously high
55 	  * bandwidth estimate.
56 	  */
57 	if (!tp->packets_out) {
58 		u64 tstamp_us = tcp_skb_timestamp_us(skb);
59 
60 		tp->first_tx_mstamp  = tstamp_us;
61 		tp->delivered_mstamp = tstamp_us;
62 	}
63 
64 	TCP_SKB_CB(skb)->tx.first_tx_mstamp	= tp->first_tx_mstamp;
65 	TCP_SKB_CB(skb)->tx.delivered_mstamp	= tp->delivered_mstamp;
66 	TCP_SKB_CB(skb)->tx.delivered		= tp->delivered;
67 	TCP_SKB_CB(skb)->tx.is_app_limited	= tp->app_limited ? 1 : 0;
68 }
69 
70 /* When an skb is sacked or acked, we fill in the rate sample with the (prior)
71  * delivery information when the skb was last transmitted.
72  *
73  * If an ACK (s)acks multiple skbs (e.g., stretched-acks), this function is
74  * called multiple times. We favor the information from the most recently
75  * sent skb, i.e., the skb with the highest prior_delivered count.
76  */
77 void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
78 			    struct rate_sample *rs)
79 {
80 	struct tcp_sock *tp = tcp_sk(sk);
81 	struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
82 
83 	if (!scb->tx.delivered_mstamp)
84 		return;
85 
86 	if (!rs->prior_delivered ||
87 	    after(scb->tx.delivered, rs->prior_delivered)) {
88 		rs->prior_delivered  = scb->tx.delivered;
89 		rs->prior_mstamp     = scb->tx.delivered_mstamp;
90 		rs->is_app_limited   = scb->tx.is_app_limited;
91 		rs->is_retrans	     = scb->sacked & TCPCB_RETRANS;
92 
93 		/* Record send time of most recently ACKed packet: */
94 		tp->first_tx_mstamp  = tcp_skb_timestamp_us(skb);
95 		/* Find the duration of the "send phase" of this window: */
96 		rs->interval_us = tcp_stamp_us_delta(tp->first_tx_mstamp,
97 						     scb->tx.first_tx_mstamp);
98 
99 	}
100 	/* Mark off the skb delivered once it's sacked to avoid being
101 	 * used again when it's cumulatively acked. For acked packets
102 	 * we don't need to reset since it'll be freed soon.
103 	 */
104 	if (scb->sacked & TCPCB_SACKED_ACKED)
105 		scb->tx.delivered_mstamp = 0;
106 }
107 
108 /* Update the connection delivery information and generate a rate sample. */
109 void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
110 		  bool is_sack_reneg, struct rate_sample *rs)
111 {
112 	struct tcp_sock *tp = tcp_sk(sk);
113 	u32 snd_us, ack_us;
114 
115 	/* Clear app limited if bubble is acked and gone. */
116 	if (tp->app_limited && after(tp->delivered, tp->app_limited))
117 		tp->app_limited = 0;
118 
119 	/* TODO: there are multiple places throughout tcp_ack() to get
120 	 * current time. Refactor the code using a new "tcp_acktag_state"
121 	 * to carry current time, flags, stats like "tcp_sacktag_state".
122 	 */
123 	if (delivered)
124 		tp->delivered_mstamp = tp->tcp_mstamp;
125 
126 	rs->acked_sacked = delivered;	/* freshly ACKed or SACKed */
127 	rs->losses = lost;		/* freshly marked lost */
128 	/* Return an invalid sample if no timing information is available or
129 	 * in recovery from loss with SACK reneging. Rate samples taken during
130 	 * a SACK reneging event may overestimate bw by including packets that
131 	 * were SACKed before the reneg.
132 	 */
133 	if (!rs->prior_mstamp || is_sack_reneg) {
134 		rs->delivered = -1;
135 		rs->interval_us = -1;
136 		return;
137 	}
138 	rs->delivered   = tp->delivered - rs->prior_delivered;
139 
140 	/* Model sending data and receiving ACKs as separate pipeline phases
141 	 * for a window. Usually the ACK phase is longer, but with ACK
142 	 * compression the send phase can be longer. To be safe we use the
143 	 * longer phase.
144 	 */
145 	snd_us = rs->interval_us;				/* send phase */
146 	ack_us = tcp_stamp_us_delta(tp->tcp_mstamp,
147 				    rs->prior_mstamp); /* ack phase */
148 	rs->interval_us = max(snd_us, ack_us);
149 
150 	/* Record both segment send and ack receive intervals */
151 	rs->snd_interval_us = snd_us;
152 	rs->rcv_interval_us = ack_us;
153 
154 	/* Normally we expect interval_us >= min-rtt.
155 	 * Note that rate may still be over-estimated when a spuriously
156 	 * retransmistted skb was first (s)acked because "interval_us"
157 	 * is under-estimated (up to an RTT). However continuously
158 	 * measuring the delivery rate during loss recovery is crucial
159 	 * for connections suffer heavy or prolonged losses.
160 	 */
161 	if (unlikely(rs->interval_us < tcp_min_rtt(tp))) {
162 		if (!rs->is_retrans)
163 			pr_debug("tcp rate: %ld %d %u %u %u\n",
164 				 rs->interval_us, rs->delivered,
165 				 inet_csk(sk)->icsk_ca_state,
166 				 tp->rx_opt.sack_ok, tcp_min_rtt(tp));
167 		rs->interval_us = -1;
168 		return;
169 	}
170 
171 	/* Record the last non-app-limited or the highest app-limited bw */
172 	if (!rs->is_app_limited ||
173 	    ((u64)rs->delivered * tp->rate_interval_us >=
174 	     (u64)tp->rate_delivered * rs->interval_us)) {
175 		tp->rate_delivered = rs->delivered;
176 		tp->rate_interval_us = rs->interval_us;
177 		tp->rate_app_limited = rs->is_app_limited;
178 	}
179 }
180 
181 /* If a gap is detected between sends, mark the socket application-limited. */
182 void tcp_rate_check_app_limited(struct sock *sk)
183 {
184 	struct tcp_sock *tp = tcp_sk(sk);
185 
186 	if (/* We have less than one packet to send. */
187 	    tp->write_seq - tp->snd_nxt < tp->mss_cache &&
188 	    /* Nothing in sending host's qdisc queues or NIC tx queue. */
189 	    sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) &&
190 	    /* We are not limited by CWND. */
191 	    tcp_packets_in_flight(tp) < tp->snd_cwnd &&
192 	    /* All lost packets have been retransmitted. */
193 	    tp->lost_out <= tp->retrans_out)
194 		tp->app_limited =
195 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
196 }
197 EXPORT_SYMBOL_GPL(tcp_rate_check_app_limited);
198