1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/tcp.h> 3 #include <net/tcp.h> 4 5 static u32 tcp_rack_reo_wnd(const struct sock *sk) 6 { 7 const struct tcp_sock *tp = tcp_sk(sk); 8 9 if (!tp->reord_seen) { 10 /* If reordering has not been observed, be aggressive during 11 * the recovery or starting the recovery by DUPACK threshold. 12 */ 13 if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery) 14 return 0; 15 16 if (tp->sacked_out >= tp->reordering && 17 !(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 18 TCP_RACK_NO_DUPTHRESH)) 19 return 0; 20 } 21 22 /* To be more reordering resilient, allow min_rtt/4 settling delay. 23 * Use min_rtt instead of the smoothed RTT because reordering is 24 * often a path property and less related to queuing or delayed ACKs. 25 * Upon receiving DSACKs, linearly increase the window up to the 26 * smoothed RTT. 27 */ 28 return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps, 29 tp->srtt_us >> 3); 30 } 31 32 s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd) 33 { 34 return tp->rack.rtt_us + reo_wnd - 35 tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb)); 36 } 37 38 /* RACK loss detection (IETF RFC8985): 39 * 40 * Marks a packet lost, if some packet sent later has been (s)acked. 41 * The underlying idea is similar to the traditional dupthresh and FACK 42 * but they look at different metrics: 43 * 44 * dupthresh: 3 OOO packets delivered (packet count) 45 * FACK: sequence delta to highest sacked sequence (sequence space) 46 * RACK: sent time delta to the latest delivered packet (time domain) 47 * 48 * The advantage of RACK is it applies to both original and retransmitted 49 * packet and therefore is robust against tail losses. Another advantage 50 * is being more resilient to reordering by simply allowing some 51 * "settling delay", instead of tweaking the dupthresh. 52 * 53 * When tcp_rack_detect_loss() detects some packets are lost and we 54 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout() 55 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will 56 * make us enter the CA_Recovery state. 57 */ 58 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout) 59 { 60 struct tcp_sock *tp = tcp_sk(sk); 61 struct sk_buff *skb, *n; 62 u32 reo_wnd; 63 64 *reo_timeout = 0; 65 reo_wnd = tcp_rack_reo_wnd(sk); 66 list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue, 67 tcp_tsorted_anchor) { 68 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 69 s32 remaining; 70 71 /* Skip ones marked lost but not yet retransmitted */ 72 if ((scb->sacked & TCPCB_LOST) && 73 !(scb->sacked & TCPCB_SACKED_RETRANS)) 74 continue; 75 76 if (!tcp_skb_sent_after(tp->rack.mstamp, 77 tcp_skb_timestamp_us(skb), 78 tp->rack.end_seq, scb->end_seq)) 79 break; 80 81 /* A packet is lost if it has not been s/acked beyond 82 * the recent RTT plus the reordering window. 83 */ 84 remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd); 85 if (remaining <= 0) { 86 tcp_mark_skb_lost(sk, skb); 87 list_del_init(&skb->tcp_tsorted_anchor); 88 } else { 89 /* Record maximum wait time */ 90 *reo_timeout = max_t(u32, *reo_timeout, remaining); 91 } 92 } 93 } 94 95 bool tcp_rack_mark_lost(struct sock *sk) 96 { 97 struct tcp_sock *tp = tcp_sk(sk); 98 u32 timeout; 99 100 if (!tp->rack.advanced) 101 return false; 102 103 /* Reset the advanced flag to avoid unnecessary queue scanning */ 104 tp->rack.advanced = 0; 105 tcp_rack_detect_loss(sk, &timeout); 106 if (timeout) { 107 timeout = usecs_to_jiffies(timeout + TCP_TIMEOUT_MIN_US); 108 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT, 109 timeout, inet_csk(sk)->icsk_rto); 110 } 111 return !!timeout; 112 } 113 114 /* We have waited long enough to accommodate reordering. Mark the expired 115 * packets lost and retransmit them. 116 */ 117 void tcp_rack_reo_timeout(struct sock *sk) 118 { 119 struct tcp_sock *tp = tcp_sk(sk); 120 u32 timeout, prior_inflight; 121 u32 lost = tp->lost; 122 123 prior_inflight = tcp_packets_in_flight(tp); 124 tcp_rack_detect_loss(sk, &timeout); 125 if (prior_inflight != tcp_packets_in_flight(tp)) { 126 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) { 127 tcp_enter_recovery(sk, false); 128 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 129 tcp_cwnd_reduction(sk, 1, tp->lost - lost, 0); 130 } 131 tcp_xmit_retransmit_queue(sk); 132 } 133 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS) 134 tcp_rearm_rto(sk); 135 } 136 137 /* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits 138 * the next unacked packet upon receiving 139 * a) three or more DUPACKs to start the fast recovery 140 * b) an ACK acknowledging new data during the fast recovery. 141 */ 142 void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced) 143 { 144 const u8 state = inet_csk(sk)->icsk_ca_state; 145 struct tcp_sock *tp = tcp_sk(sk); 146 147 if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) || 148 (state == TCP_CA_Recovery && snd_una_advanced)) { 149 struct sk_buff *skb = tcp_rtx_queue_head(sk); 150 u32 mss; 151 152 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 153 return; 154 155 mss = tcp_skb_mss(skb); 156 if (tcp_skb_pcount(skb) > 1 && skb->len > mss) 157 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 158 mss, mss, GFP_ATOMIC); 159 160 tcp_mark_skb_lost(sk, skb); 161 } 162 } 163