1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/tcp.h> 3 #include <net/tcp.h> 4 5 int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOSS_DETECTION; 6 7 static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 8 { 9 struct tcp_sock *tp = tcp_sk(sk); 10 11 tcp_skb_mark_lost_uncond_verify(tp, skb); 12 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 13 /* Account for retransmits that are lost again */ 14 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 15 tp->retrans_out -= tcp_skb_pcount(skb); 16 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT, 17 tcp_skb_pcount(skb)); 18 } 19 } 20 21 static bool tcp_rack_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2) 22 { 23 return t1 > t2 || (t1 == t2 && after(seq1, seq2)); 24 } 25 26 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01): 27 * 28 * Marks a packet lost, if some packet sent later has been (s)acked. 29 * The underlying idea is similar to the traditional dupthresh and FACK 30 * but they look at different metrics: 31 * 32 * dupthresh: 3 OOO packets delivered (packet count) 33 * FACK: sequence delta to highest sacked sequence (sequence space) 34 * RACK: sent time delta to the latest delivered packet (time domain) 35 * 36 * The advantage of RACK is it applies to both original and retransmitted 37 * packet and therefore is robust against tail losses. Another advantage 38 * is being more resilient to reordering by simply allowing some 39 * "settling delay", instead of tweaking the dupthresh. 40 * 41 * When tcp_rack_detect_loss() detects some packets are lost and we 42 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout() 43 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will 44 * make us enter the CA_Recovery state. 45 */ 46 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout) 47 { 48 struct tcp_sock *tp = tcp_sk(sk); 49 struct sk_buff *skb; 50 u32 reo_wnd; 51 52 *reo_timeout = 0; 53 /* To be more reordering resilient, allow min_rtt/4 settling delay 54 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed 55 * RTT because reordering is often a path property and less related 56 * to queuing or delayed ACKs. 57 */ 58 reo_wnd = 1000; 59 if ((tp->rack.reord || !tp->lost_out) && tcp_min_rtt(tp) != ~0U) 60 reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd); 61 62 tcp_for_write_queue(skb, sk) { 63 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 64 65 if (skb == tcp_send_head(sk)) 66 break; 67 68 /* Skip ones already (s)acked */ 69 if (!after(scb->end_seq, tp->snd_una) || 70 scb->sacked & TCPCB_SACKED_ACKED) 71 continue; 72 73 if (tcp_rack_sent_after(tp->rack.mstamp, skb->skb_mstamp, 74 tp->rack.end_seq, scb->end_seq)) { 75 /* Step 3 in draft-cheng-tcpm-rack-00.txt: 76 * A packet is lost if its elapsed time is beyond 77 * the recent RTT plus the reordering window. 78 */ 79 u32 elapsed = tcp_stamp_us_delta(tp->tcp_mstamp, 80 skb->skb_mstamp); 81 s32 remaining = tp->rack.rtt_us + reo_wnd - elapsed; 82 83 if (remaining < 0) { 84 tcp_rack_mark_skb_lost(sk, skb); 85 continue; 86 } 87 88 /* Skip ones marked lost but not yet retransmitted */ 89 if ((scb->sacked & TCPCB_LOST) && 90 !(scb->sacked & TCPCB_SACKED_RETRANS)) 91 continue; 92 93 /* Record maximum wait time (+1 to avoid 0) */ 94 *reo_timeout = max_t(u32, *reo_timeout, 1 + remaining); 95 96 } else if (!(scb->sacked & TCPCB_RETRANS)) { 97 /* Original data are sent sequentially so stop early 98 * b/c the rest are all sent after rack_sent 99 */ 100 break; 101 } 102 } 103 } 104 105 void tcp_rack_mark_lost(struct sock *sk) 106 { 107 struct tcp_sock *tp = tcp_sk(sk); 108 u32 timeout; 109 110 if (!tp->rack.advanced) 111 return; 112 113 /* Reset the advanced flag to avoid unnecessary queue scanning */ 114 tp->rack.advanced = 0; 115 tcp_rack_detect_loss(sk, &timeout); 116 if (timeout) { 117 timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN; 118 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT, 119 timeout, inet_csk(sk)->icsk_rto); 120 } 121 } 122 123 /* Record the most recently (re)sent time among the (s)acked packets 124 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from 125 * draft-cheng-tcpm-rack-00.txt 126 */ 127 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, 128 u64 xmit_time) 129 { 130 u32 rtt_us; 131 132 if (tp->rack.mstamp && 133 !tcp_rack_sent_after(xmit_time, tp->rack.mstamp, 134 end_seq, tp->rack.end_seq)) 135 return; 136 137 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time); 138 if (sacked & TCPCB_RETRANS) { 139 /* If the sacked packet was retransmitted, it's ambiguous 140 * whether the retransmission or the original (or the prior 141 * retransmission) was sacked. 142 * 143 * If the original is lost, there is no ambiguity. Otherwise 144 * we assume the original can be delayed up to aRTT + min_rtt. 145 * the aRTT term is bounded by the fast recovery or timeout, 146 * so it's at least one RTT (i.e., retransmission is at least 147 * an RTT later). 148 */ 149 if (rtt_us < tcp_min_rtt(tp)) 150 return; 151 } 152 tp->rack.rtt_us = rtt_us; 153 tp->rack.mstamp = xmit_time; 154 tp->rack.end_seq = end_seq; 155 tp->rack.advanced = 1; 156 } 157 158 /* We have waited long enough to accommodate reordering. Mark the expired 159 * packets lost and retransmit them. 160 */ 161 void tcp_rack_reo_timeout(struct sock *sk) 162 { 163 struct tcp_sock *tp = tcp_sk(sk); 164 u32 timeout, prior_inflight; 165 166 prior_inflight = tcp_packets_in_flight(tp); 167 tcp_rack_detect_loss(sk, &timeout); 168 if (prior_inflight != tcp_packets_in_flight(tp)) { 169 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) { 170 tcp_enter_recovery(sk, false); 171 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 172 tcp_cwnd_reduction(sk, 1, 0); 173 } 174 tcp_xmit_retransmit_queue(sk); 175 } 176 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS) 177 tcp_rearm_rto(sk); 178 } 179