xref: /openbmc/linux/net/ipv4/tcp_rate.c (revision 3b23dc52)
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 		tp->first_tx_mstamp  = skb->skb_mstamp;
59 		tp->delivered_mstamp = skb->skb_mstamp;
60 	}
61 
62 	TCP_SKB_CB(skb)->tx.first_tx_mstamp	= tp->first_tx_mstamp;
63 	TCP_SKB_CB(skb)->tx.delivered_mstamp	= tp->delivered_mstamp;
64 	TCP_SKB_CB(skb)->tx.delivered		= tp->delivered;
65 	TCP_SKB_CB(skb)->tx.is_app_limited	= tp->app_limited ? 1 : 0;
66 }
67 
68 /* When an skb is sacked or acked, we fill in the rate sample with the (prior)
69  * delivery information when the skb was last transmitted.
70  *
71  * If an ACK (s)acks multiple skbs (e.g., stretched-acks), this function is
72  * called multiple times. We favor the information from the most recently
73  * sent skb, i.e., the skb with the highest prior_delivered count.
74  */
75 void tcp_rate_skb_delivered(struct sock *sk, struct sk_buff *skb,
76 			    struct rate_sample *rs)
77 {
78 	struct tcp_sock *tp = tcp_sk(sk);
79 	struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
80 
81 	if (!scb->tx.delivered_mstamp)
82 		return;
83 
84 	if (!rs->prior_delivered ||
85 	    after(scb->tx.delivered, rs->prior_delivered)) {
86 		rs->prior_delivered  = scb->tx.delivered;
87 		rs->prior_mstamp     = scb->tx.delivered_mstamp;
88 		rs->is_app_limited   = scb->tx.is_app_limited;
89 		rs->is_retrans	     = scb->sacked & TCPCB_RETRANS;
90 
91 		/* Find the duration of the "send phase" of this window: */
92 		rs->interval_us      = tcp_stamp_us_delta(
93 						skb->skb_mstamp,
94 						scb->tx.first_tx_mstamp);
95 
96 		/* Record send time of most recently ACKed packet: */
97 		tp->first_tx_mstamp  = skb->skb_mstamp;
98 	}
99 	/* Mark off the skb delivered once it's sacked to avoid being
100 	 * used again when it's cumulatively acked. For acked packets
101 	 * we don't need to reset since it'll be freed soon.
102 	 */
103 	if (scb->sacked & TCPCB_SACKED_ACKED)
104 		scb->tx.delivered_mstamp = 0;
105 }
106 
107 /* Update the connection delivery information and generate a rate sample. */
108 void tcp_rate_gen(struct sock *sk, u32 delivered, u32 lost,
109 		  bool is_sack_reneg, struct rate_sample *rs)
110 {
111 	struct tcp_sock *tp = tcp_sk(sk);
112 	u32 snd_us, ack_us;
113 
114 	/* Clear app limited if bubble is acked and gone. */
115 	if (tp->app_limited && after(tp->delivered, tp->app_limited))
116 		tp->app_limited = 0;
117 
118 	/* TODO: there are multiple places throughout tcp_ack() to get
119 	 * current time. Refactor the code using a new "tcp_acktag_state"
120 	 * to carry current time, flags, stats like "tcp_sacktag_state".
121 	 */
122 	if (delivered)
123 		tp->delivered_mstamp = tp->tcp_mstamp;
124 
125 	rs->acked_sacked = delivered;	/* freshly ACKed or SACKed */
126 	rs->losses = lost;		/* freshly marked lost */
127 	/* Return an invalid sample if no timing information is available or
128 	 * in recovery from loss with SACK reneging. Rate samples taken during
129 	 * a SACK reneging event may overestimate bw by including packets that
130 	 * were SACKed before the reneg.
131 	 */
132 	if (!rs->prior_mstamp || is_sack_reneg) {
133 		rs->delivered = -1;
134 		rs->interval_us = -1;
135 		return;
136 	}
137 	rs->delivered   = tp->delivered - rs->prior_delivered;
138 
139 	/* Model sending data and receiving ACKs as separate pipeline phases
140 	 * for a window. Usually the ACK phase is longer, but with ACK
141 	 * compression the send phase can be longer. To be safe we use the
142 	 * longer phase.
143 	 */
144 	snd_us = rs->interval_us;				/* send phase */
145 	ack_us = tcp_stamp_us_delta(tp->tcp_mstamp,
146 				    rs->prior_mstamp); /* ack phase */
147 	rs->interval_us = max(snd_us, ack_us);
148 
149 	/* Normally we expect interval_us >= min-rtt.
150 	 * Note that rate may still be over-estimated when a spuriously
151 	 * retransmistted skb was first (s)acked because "interval_us"
152 	 * is under-estimated (up to an RTT). However continuously
153 	 * measuring the delivery rate during loss recovery is crucial
154 	 * for connections suffer heavy or prolonged losses.
155 	 */
156 	if (unlikely(rs->interval_us < tcp_min_rtt(tp))) {
157 		if (!rs->is_retrans)
158 			pr_debug("tcp rate: %ld %d %u %u %u\n",
159 				 rs->interval_us, rs->delivered,
160 				 inet_csk(sk)->icsk_ca_state,
161 				 tp->rx_opt.sack_ok, tcp_min_rtt(tp));
162 		rs->interval_us = -1;
163 		return;
164 	}
165 
166 	/* Record the last non-app-limited or the highest app-limited bw */
167 	if (!rs->is_app_limited ||
168 	    ((u64)rs->delivered * tp->rate_interval_us >=
169 	     (u64)tp->rate_delivered * rs->interval_us)) {
170 		tp->rate_delivered = rs->delivered;
171 		tp->rate_interval_us = rs->interval_us;
172 		tp->rate_app_limited = rs->is_app_limited;
173 	}
174 }
175 
176 /* If a gap is detected between sends, mark the socket application-limited. */
177 void tcp_rate_check_app_limited(struct sock *sk)
178 {
179 	struct tcp_sock *tp = tcp_sk(sk);
180 
181 	if (/* We have less than one packet to send. */
182 	    tp->write_seq - tp->snd_nxt < tp->mss_cache &&
183 	    /* Nothing in sending host's qdisc queues or NIC tx queue. */
184 	    sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) &&
185 	    /* We are not limited by CWND. */
186 	    tcp_packets_in_flight(tp) < tp->snd_cwnd &&
187 	    /* All lost packets have been retransmitted. */
188 	    tp->lost_out <= tp->retrans_out)
189 		tp->app_limited =
190 			(tp->delivered + tcp_packets_in_flight(tp)) ? : 1;
191 }
192 EXPORT_SYMBOL_GPL(tcp_rate_check_app_limited);
193