xref: /linux/net/ipv4/tcp_recovery.c (revision e5c86679d5e864947a52fb31e45a425dea3e7fa9)
1 #include <linux/tcp.h>
2 #include <net/tcp.h>
3 
4 int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOSS_DETECTION;
5 
6 static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
7 {
8 	struct tcp_sock *tp = tcp_sk(sk);
9 
10 	tcp_skb_mark_lost_uncond_verify(tp, skb);
11 	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
12 		/* Account for retransmits that are lost again */
13 		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
14 		tp->retrans_out -= tcp_skb_pcount(skb);
15 		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
16 			      tcp_skb_pcount(skb));
17 	}
18 }
19 
20 static bool tcp_rack_sent_after(const struct skb_mstamp *t1,
21 				const struct skb_mstamp *t2,
22 				u32 seq1, u32 seq2)
23 {
24 	return skb_mstamp_after(t1, t2) ||
25 	       (t1->v64 == t2->v64 && after(seq1, seq2));
26 }
27 
28 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
29  *
30  * Marks a packet lost, if some packet sent later has been (s)acked.
31  * The underlying idea is similar to the traditional dupthresh and FACK
32  * but they look at different metrics:
33  *
34  * dupthresh: 3 OOO packets delivered (packet count)
35  * FACK: sequence delta to highest sacked sequence (sequence space)
36  * RACK: sent time delta to the latest delivered packet (time domain)
37  *
38  * The advantage of RACK is it applies to both original and retransmitted
39  * packet and therefore is robust against tail losses. Another advantage
40  * is being more resilient to reordering by simply allowing some
41  * "settling delay", instead of tweaking the dupthresh.
42  *
43  * When tcp_rack_detect_loss() detects some packets are lost and we
44  * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
45  * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
46  * make us enter the CA_Recovery state.
47  */
48 static void tcp_rack_detect_loss(struct sock *sk, const struct skb_mstamp *now,
49 				 u32 *reo_timeout)
50 {
51 	struct tcp_sock *tp = tcp_sk(sk);
52 	struct sk_buff *skb;
53 	u32 reo_wnd;
54 
55 	*reo_timeout = 0;
56 	/* To be more reordering resilient, allow min_rtt/4 settling delay
57 	 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed
58 	 * RTT because reordering is often a path property and less related
59 	 * to queuing or delayed ACKs.
60 	 */
61 	reo_wnd = 1000;
62 	if ((tp->rack.reord || !tp->lost_out) && tcp_min_rtt(tp) != ~0U)
63 		reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd);
64 
65 	tcp_for_write_queue(skb, sk) {
66 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
67 
68 		if (skb == tcp_send_head(sk))
69 			break;
70 
71 		/* Skip ones already (s)acked */
72 		if (!after(scb->end_seq, tp->snd_una) ||
73 		    scb->sacked & TCPCB_SACKED_ACKED)
74 			continue;
75 
76 		if (tcp_rack_sent_after(&tp->rack.mstamp, &skb->skb_mstamp,
77 					tp->rack.end_seq, scb->end_seq)) {
78 			/* Step 3 in draft-cheng-tcpm-rack-00.txt:
79 			 * A packet is lost if its elapsed time is beyond
80 			 * the recent RTT plus the reordering window.
81 			 */
82 			u32 elapsed = skb_mstamp_us_delta(now,
83 							  &skb->skb_mstamp);
84 			s32 remaining = tp->rack.rtt_us + reo_wnd - elapsed;
85 
86 			if (remaining < 0) {
87 				tcp_rack_mark_skb_lost(sk, skb);
88 				continue;
89 			}
90 
91 			/* Skip ones marked lost but not yet retransmitted */
92 			if ((scb->sacked & TCPCB_LOST) &&
93 			    !(scb->sacked & TCPCB_SACKED_RETRANS))
94 				continue;
95 
96 			/* Record maximum wait time (+1 to avoid 0) */
97 			*reo_timeout = max_t(u32, *reo_timeout, 1 + remaining);
98 
99 		} else if (!(scb->sacked & TCPCB_RETRANS)) {
100 			/* Original data are sent sequentially so stop early
101 			 * b/c the rest are all sent after rack_sent
102 			 */
103 			break;
104 		}
105 	}
106 }
107 
108 void tcp_rack_mark_lost(struct sock *sk, const struct skb_mstamp *now)
109 {
110 	struct tcp_sock *tp = tcp_sk(sk);
111 	u32 timeout;
112 
113 	if (!tp->rack.advanced)
114 		return;
115 
116 	/* Reset the advanced flag to avoid unnecessary queue scanning */
117 	tp->rack.advanced = 0;
118 	tcp_rack_detect_loss(sk, now, &timeout);
119 	if (timeout) {
120 		timeout = usecs_to_jiffies(timeout + TCP_REO_TIMEOUT_MIN);
121 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
122 					  timeout, inet_csk(sk)->icsk_rto);
123 	}
124 }
125 
126 /* Record the most recently (re)sent time among the (s)acked packets
127  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
128  * draft-cheng-tcpm-rack-00.txt
129  */
130 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
131 		      const struct skb_mstamp *xmit_time,
132 		      const struct skb_mstamp *ack_time)
133 {
134 	u32 rtt_us;
135 
136 	if (tp->rack.mstamp.v64 &&
137 	    !tcp_rack_sent_after(xmit_time, &tp->rack.mstamp,
138 				 end_seq, tp->rack.end_seq))
139 		return;
140 
141 	rtt_us = skb_mstamp_us_delta(ack_time, xmit_time);
142 	if (sacked & TCPCB_RETRANS) {
143 		/* If the sacked packet was retransmitted, it's ambiguous
144 		 * whether the retransmission or the original (or the prior
145 		 * retransmission) was sacked.
146 		 *
147 		 * If the original is lost, there is no ambiguity. Otherwise
148 		 * we assume the original can be delayed up to aRTT + min_rtt.
149 		 * the aRTT term is bounded by the fast recovery or timeout,
150 		 * so it's at least one RTT (i.e., retransmission is at least
151 		 * an RTT later).
152 		 */
153 		if (rtt_us < tcp_min_rtt(tp))
154 			return;
155 	}
156 	tp->rack.rtt_us = rtt_us;
157 	tp->rack.mstamp = *xmit_time;
158 	tp->rack.end_seq = end_seq;
159 	tp->rack.advanced = 1;
160 }
161 
162 /* We have waited long enough to accommodate reordering. Mark the expired
163  * packets lost and retransmit them.
164  */
165 void tcp_rack_reo_timeout(struct sock *sk)
166 {
167 	struct tcp_sock *tp = tcp_sk(sk);
168 	struct skb_mstamp now;
169 	u32 timeout, prior_inflight;
170 
171 	skb_mstamp_get(&now);
172 	prior_inflight = tcp_packets_in_flight(tp);
173 	tcp_rack_detect_loss(sk, &now, &timeout);
174 	if (prior_inflight != tcp_packets_in_flight(tp)) {
175 		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
176 			tcp_enter_recovery(sk, false);
177 			if (!inet_csk(sk)->icsk_ca_ops->cong_control)
178 				tcp_cwnd_reduction(sk, 1, 0);
179 		}
180 		tcp_xmit_retransmit_queue(sk);
181 	}
182 	if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
183 		tcp_rearm_rto(sk);
184 }
185