xref: /openbmc/linux/net/ipv4/tcp_recovery.c (revision c997aabb)
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(u64 t1, u64 t2, u32 seq1, u32 seq2)
21 {
22 	return t1 > t2 || (t1 == t2 && after(seq1, seq2));
23 }
24 
25 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
26  *
27  * Marks a packet lost, if some packet sent later has been (s)acked.
28  * The underlying idea is similar to the traditional dupthresh and FACK
29  * but they look at different metrics:
30  *
31  * dupthresh: 3 OOO packets delivered (packet count)
32  * FACK: sequence delta to highest sacked sequence (sequence space)
33  * RACK: sent time delta to the latest delivered packet (time domain)
34  *
35  * The advantage of RACK is it applies to both original and retransmitted
36  * packet and therefore is robust against tail losses. Another advantage
37  * is being more resilient to reordering by simply allowing some
38  * "settling delay", instead of tweaking the dupthresh.
39  *
40  * When tcp_rack_detect_loss() detects some packets are lost and we
41  * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
42  * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
43  * make us enter the CA_Recovery state.
44  */
45 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout)
46 {
47 	struct tcp_sock *tp = tcp_sk(sk);
48 	struct sk_buff *skb;
49 	u32 reo_wnd;
50 
51 	*reo_timeout = 0;
52 	/* To be more reordering resilient, allow min_rtt/4 settling delay
53 	 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed
54 	 * RTT because reordering is often a path property and less related
55 	 * to queuing or delayed ACKs.
56 	 */
57 	reo_wnd = 1000;
58 	if ((tp->rack.reord || !tp->lost_out) && tcp_min_rtt(tp) != ~0U)
59 		reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd);
60 
61 	tcp_for_write_queue(skb, sk) {
62 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
63 
64 		if (skb == tcp_send_head(sk))
65 			break;
66 
67 		/* Skip ones already (s)acked */
68 		if (!after(scb->end_seq, tp->snd_una) ||
69 		    scb->sacked & TCPCB_SACKED_ACKED)
70 			continue;
71 
72 		if (tcp_rack_sent_after(tp->rack.mstamp, skb->skb_mstamp,
73 					tp->rack.end_seq, scb->end_seq)) {
74 			/* Step 3 in draft-cheng-tcpm-rack-00.txt:
75 			 * A packet is lost if its elapsed time is beyond
76 			 * the recent RTT plus the reordering window.
77 			 */
78 			u32 elapsed = tcp_stamp_us_delta(tp->tcp_mstamp,
79 							 skb->skb_mstamp);
80 			s32 remaining = tp->rack.rtt_us + reo_wnd - elapsed;
81 
82 			if (remaining < 0) {
83 				tcp_rack_mark_skb_lost(sk, skb);
84 				continue;
85 			}
86 
87 			/* Skip ones marked lost but not yet retransmitted */
88 			if ((scb->sacked & TCPCB_LOST) &&
89 			    !(scb->sacked & TCPCB_SACKED_RETRANS))
90 				continue;
91 
92 			/* Record maximum wait time (+1 to avoid 0) */
93 			*reo_timeout = max_t(u32, *reo_timeout, 1 + remaining);
94 
95 		} else if (!(scb->sacked & TCPCB_RETRANS)) {
96 			/* Original data are sent sequentially so stop early
97 			 * b/c the rest are all sent after rack_sent
98 			 */
99 			break;
100 		}
101 	}
102 }
103 
104 void tcp_rack_mark_lost(struct sock *sk)
105 {
106 	struct tcp_sock *tp = tcp_sk(sk);
107 	u32 timeout;
108 
109 	if (!tp->rack.advanced)
110 		return;
111 
112 	/* Reset the advanced flag to avoid unnecessary queue scanning */
113 	tp->rack.advanced = 0;
114 	tcp_rack_detect_loss(sk, &timeout);
115 	if (timeout) {
116 		timeout = usecs_to_jiffies(timeout + TCP_REO_TIMEOUT_MIN);
117 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
118 					  timeout, inet_csk(sk)->icsk_rto);
119 	}
120 }
121 
122 /* Record the most recently (re)sent time among the (s)acked packets
123  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
124  * draft-cheng-tcpm-rack-00.txt
125  */
126 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
127 		      u64 xmit_time)
128 {
129 	u32 rtt_us;
130 
131 	if (tp->rack.mstamp &&
132 	    !tcp_rack_sent_after(xmit_time, tp->rack.mstamp,
133 				 end_seq, tp->rack.end_seq))
134 		return;
135 
136 	rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
137 	if (sacked & TCPCB_RETRANS) {
138 		/* If the sacked packet was retransmitted, it's ambiguous
139 		 * whether the retransmission or the original (or the prior
140 		 * retransmission) was sacked.
141 		 *
142 		 * If the original is lost, there is no ambiguity. Otherwise
143 		 * we assume the original can be delayed up to aRTT + min_rtt.
144 		 * the aRTT term is bounded by the fast recovery or timeout,
145 		 * so it's at least one RTT (i.e., retransmission is at least
146 		 * an RTT later).
147 		 */
148 		if (rtt_us < tcp_min_rtt(tp))
149 			return;
150 	}
151 	tp->rack.rtt_us = rtt_us;
152 	tp->rack.mstamp = xmit_time;
153 	tp->rack.end_seq = end_seq;
154 	tp->rack.advanced = 1;
155 }
156 
157 /* We have waited long enough to accommodate reordering. Mark the expired
158  * packets lost and retransmit them.
159  */
160 void tcp_rack_reo_timeout(struct sock *sk)
161 {
162 	struct tcp_sock *tp = tcp_sk(sk);
163 	u32 timeout, prior_inflight;
164 
165 	prior_inflight = tcp_packets_in_flight(tp);
166 	tcp_rack_detect_loss(sk, &timeout);
167 	if (prior_inflight != tcp_packets_in_flight(tp)) {
168 		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
169 			tcp_enter_recovery(sk, false);
170 			if (!inet_csk(sk)->icsk_ca_ops->cong_control)
171 				tcp_cwnd_reduction(sk, 1, 0);
172 		}
173 		tcp_xmit_retransmit_queue(sk);
174 	}
175 	if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
176 		tcp_rearm_rto(sk);
177 }
178