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, *n; 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 list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue, 62 tcp_tsorted_anchor) { 63 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 64 s32 remaining; 65 66 /* Skip ones marked lost but not yet retransmitted */ 67 if ((scb->sacked & TCPCB_LOST) && 68 !(scb->sacked & TCPCB_SACKED_RETRANS)) 69 continue; 70 71 if (!tcp_rack_sent_after(tp->rack.mstamp, skb->skb_mstamp, 72 tp->rack.end_seq, scb->end_seq)) 73 break; 74 75 /* A packet is lost if it has not been s/acked beyond 76 * the recent RTT plus the reordering window. 77 */ 78 remaining = tp->rack.rtt_us + reo_wnd - 79 tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp); 80 if (remaining < 0) { 81 tcp_rack_mark_skb_lost(sk, skb); 82 list_del_init(&skb->tcp_tsorted_anchor); 83 } else { 84 /* Record maximum wait time (+1 to avoid 0) */ 85 *reo_timeout = max_t(u32, *reo_timeout, 1 + remaining); 86 } 87 } 88 } 89 90 void tcp_rack_mark_lost(struct sock *sk) 91 { 92 struct tcp_sock *tp = tcp_sk(sk); 93 u32 timeout; 94 95 if (!tp->rack.advanced) 96 return; 97 98 /* Reset the advanced flag to avoid unnecessary queue scanning */ 99 tp->rack.advanced = 0; 100 tcp_rack_detect_loss(sk, &timeout); 101 if (timeout) { 102 timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN; 103 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT, 104 timeout, inet_csk(sk)->icsk_rto); 105 } 106 } 107 108 /* Record the most recently (re)sent time among the (s)acked packets 109 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from 110 * draft-cheng-tcpm-rack-00.txt 111 */ 112 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, 113 u64 xmit_time) 114 { 115 u32 rtt_us; 116 117 if (tp->rack.mstamp && 118 !tcp_rack_sent_after(xmit_time, tp->rack.mstamp, 119 end_seq, tp->rack.end_seq)) 120 return; 121 122 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time); 123 if (sacked & TCPCB_RETRANS) { 124 /* If the sacked packet was retransmitted, it's ambiguous 125 * whether the retransmission or the original (or the prior 126 * retransmission) was sacked. 127 * 128 * If the original is lost, there is no ambiguity. Otherwise 129 * we assume the original can be delayed up to aRTT + min_rtt. 130 * the aRTT term is bounded by the fast recovery or timeout, 131 * so it's at least one RTT (i.e., retransmission is at least 132 * an RTT later). 133 */ 134 if (rtt_us < tcp_min_rtt(tp)) 135 return; 136 } 137 tp->rack.rtt_us = rtt_us; 138 tp->rack.mstamp = xmit_time; 139 tp->rack.end_seq = end_seq; 140 tp->rack.advanced = 1; 141 } 142 143 /* We have waited long enough to accommodate reordering. Mark the expired 144 * packets lost and retransmit them. 145 */ 146 void tcp_rack_reo_timeout(struct sock *sk) 147 { 148 struct tcp_sock *tp = tcp_sk(sk); 149 u32 timeout, prior_inflight; 150 151 prior_inflight = tcp_packets_in_flight(tp); 152 tcp_rack_detect_loss(sk, &timeout); 153 if (prior_inflight != tcp_packets_in_flight(tp)) { 154 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) { 155 tcp_enter_recovery(sk, false); 156 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 157 tcp_cwnd_reduction(sk, 1, 0); 158 } 159 tcp_xmit_retransmit_queue(sk); 160 } 161 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS) 162 tcp_rearm_rto(sk); 163 } 164