1// SPDX-License-Identifier: GPL-2.0 2#include <linux/tcp.h> 3#include <net/tcp.h> 4 5static bool tcp_rack_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2) 6{ 7 return t1 > t2 || (t1 == t2 && after(seq1, seq2)); 8} 9 10static u32 tcp_rack_reo_wnd(const struct sock *sk) 11{ 12 struct tcp_sock *tp = tcp_sk(sk); 13 14 if (!tp->reord_seen) { 15 /* If reordering has not been observed, be aggressive during 16 * the recovery or starting the recovery by DUPACK threshold. 17 */ 18 if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery) 19 return 0; 20 21 if (tp->sacked_out >= tp->reordering && 22 !(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 23 TCP_RACK_NO_DUPTHRESH)) 24 return 0; 25 } 26 27 /* To be more reordering resilient, allow min_rtt/4 settling delay. 28 * Use min_rtt instead of the smoothed RTT because reordering is 29 * often a path property and less related to queuing or delayed ACKs. 30 * Upon receiving DSACKs, linearly increase the window up to the 31 * smoothed RTT. 32 */ 33 return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps, 34 tp->srtt_us >> 3); 35} 36 37s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd) 38{ 39 return tp->rack.rtt_us + reo_wnd - 40 tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb)); 41} 42 43/* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01): 44 * 45 * Marks a packet lost, if some packet sent later has been (s)acked. 46 * The underlying idea is similar to the traditional dupthresh and FACK 47 * but they look at different metrics: 48 * 49 * dupthresh: 3 OOO packets delivered (packet count) 50 * FACK: sequence delta to highest sacked sequence (sequence space) 51 * RACK: sent time delta to the latest delivered packet (time domain) 52 * 53 * The advantage of RACK is it applies to both original and retransmitted 54 * packet and therefore is robust against tail losses. Another advantage 55 * is being more resilient to reordering by simply allowing some 56 * "settling delay", instead of tweaking the dupthresh. 57 * 58 * When tcp_rack_detect_loss() detects some packets are lost and we 59 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout() 60 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will 61 * make us enter the CA_Recovery state. 62 */ 63static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout) 64{ 65 struct tcp_sock *tp = tcp_sk(sk); 66 struct sk_buff *skb, *n; 67 u32 reo_wnd; 68 69 *reo_timeout = 0; 70 reo_wnd = tcp_rack_reo_wnd(sk); 71 list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue, 72 tcp_tsorted_anchor) { 73 struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 74 s32 remaining; 75 76 /* Skip ones marked lost but not yet retransmitted */ 77 if ((scb->sacked & TCPCB_LOST) && 78 !(scb->sacked & TCPCB_SACKED_RETRANS)) 79 continue; 80 81 if (!tcp_rack_sent_after(tp->rack.mstamp, 82 tcp_skb_timestamp_us(skb), 83 tp->rack.end_seq, scb->end_seq)) 84 break; 85 86 /* A packet is lost if it has not been s/acked beyond 87 * the recent RTT plus the reordering window. 88 */ 89 remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd); 90 if (remaining <= 0) { 91 tcp_mark_skb_lost(sk, skb); 92 list_del_init(&skb->tcp_tsorted_anchor); 93 } else { 94 /* Record maximum wait time */ 95 *reo_timeout = max_t(u32, *reo_timeout, remaining); 96 } 97 } 98} 99 100bool tcp_rack_mark_lost(struct sock *sk) 101{ 102 struct tcp_sock *tp = tcp_sk(sk); 103 u32 timeout; 104 105 if (!tp->rack.advanced) 106 return false; 107 108 /* Reset the advanced flag to avoid unnecessary queue scanning */ 109 tp->rack.advanced = 0; 110 tcp_rack_detect_loss(sk, &timeout); 111 if (timeout) { 112 timeout = usecs_to_jiffies(timeout + TCP_TIMEOUT_MIN_US); 113 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT, 114 timeout, inet_csk(sk)->icsk_rto); 115 } 116 return !!timeout; 117} 118 119/* Record the most recently (re)sent time among the (s)acked packets 120 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from 121 * draft-cheng-tcpm-rack-00.txt 122 */ 123void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, 124 u64 xmit_time) 125{ 126 u32 rtt_us; 127 128 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time); 129 if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) { 130 /* If the sacked packet was retransmitted, it's ambiguous 131 * whether the retransmission or the original (or the prior 132 * retransmission) was sacked. 133 * 134 * If the original is lost, there is no ambiguity. Otherwise 135 * we assume the original can be delayed up to aRTT + min_rtt. 136 * the aRTT term is bounded by the fast recovery or timeout, 137 * so it's at least one RTT (i.e., retransmission is at least 138 * an RTT later). 139 */ 140 return; 141 } 142 tp->rack.advanced = 1; 143 tp->rack.rtt_us = rtt_us; 144 if (tcp_rack_sent_after(xmit_time, tp->rack.mstamp, 145 end_seq, tp->rack.end_seq)) { 146 tp->rack.mstamp = xmit_time; 147 tp->rack.end_seq = end_seq; 148 } 149} 150 151/* We have waited long enough to accommodate reordering. Mark the expired 152 * packets lost and retransmit them. 153 */ 154void tcp_rack_reo_timeout(struct sock *sk) 155{ 156 struct tcp_sock *tp = tcp_sk(sk); 157 u32 timeout, prior_inflight; 158 159 prior_inflight = tcp_packets_in_flight(tp); 160 tcp_rack_detect_loss(sk, &timeout); 161 if (prior_inflight != tcp_packets_in_flight(tp)) { 162 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) { 163 tcp_enter_recovery(sk, false); 164 if (!inet_csk(sk)->icsk_ca_ops->cong_control) 165 tcp_cwnd_reduction(sk, 1, 0); 166 } 167 tcp_xmit_retransmit_queue(sk); 168 } 169 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS) 170 tcp_rearm_rto(sk); 171} 172 173/* Updates the RACK's reo_wnd based on DSACK and no. of recoveries. 174 * 175 * If DSACK is received, increment reo_wnd by min_rtt/4 (upper bounded 176 * by srtt), since there is possibility that spurious retransmission was 177 * due to reordering delay longer than reo_wnd. 178 * 179 * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16) 180 * no. of successful recoveries (accounts for full DSACK-based loss 181 * recovery undo). After that, reset it to default (min_rtt/4). 182 * 183 * At max, reo_wnd is incremented only once per rtt. So that the new 184 * DSACK on which we are reacting, is due to the spurious retx (approx) 185 * after the reo_wnd has been updated last time. 186 * 187 * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than 188 * absolute value to account for change in rtt. 189 */ 190void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs) 191{ 192 struct tcp_sock *tp = tcp_sk(sk); 193 194 if ((READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) & 195 TCP_RACK_STATIC_REO_WND) || 196 !rs->prior_delivered) 197 return; 198 199 /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */ 200 if (before(rs->prior_delivered, tp->rack.last_delivered)) 201 tp->rack.dsack_seen = 0; 202 203 /* Adjust the reo_wnd if update is pending */ 204 if (tp->rack.dsack_seen) { 205 tp->rack.reo_wnd_steps = min_t(u32, 0xFF, 206 tp->rack.reo_wnd_steps + 1); 207 tp->rack.dsack_seen = 0; 208 tp->rack.last_delivered = tp->delivered; 209 tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH; 210 } else if (!tp->rack.reo_wnd_persist) { 211 tp->rack.reo_wnd_steps = 1; 212 } 213} 214 215/* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits 216 * the next unacked packet upon receiving 217 * a) three or more DUPACKs to start the fast recovery 218 * b) an ACK acknowledging new data during the fast recovery. 219 */ 220void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced) 221{ 222 const u8 state = inet_csk(sk)->icsk_ca_state; 223 struct tcp_sock *tp = tcp_sk(sk); 224 225 if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) || 226 (state == TCP_CA_Recovery && snd_una_advanced)) { 227 struct sk_buff *skb = tcp_rtx_queue_head(sk); 228 u32 mss; 229 230 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 231 return; 232 233 mss = tcp_skb_mss(skb); 234 if (tcp_skb_pcount(skb) > 1 && skb->len > mss) 235 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 236 mss, mss, GFP_ATOMIC); 237 238 tcp_mark_skb_lost(sk, skb); 239 } 240} 241