1b2441318SGreg Kroah-Hartman // SPDX-License-Identifier: GPL-2.0 2659a8ad5SYuchung Cheng #include <linux/tcp.h> 3659a8ad5SYuchung Cheng #include <net/tcp.h> 4659a8ad5SYuchung Cheng 59a568de4SEric Dumazet static bool tcp_rack_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2) 61d0833dfSYuchung Cheng { 79a568de4SEric Dumazet return t1 > t2 || (t1 == t2 && after(seq1, seq2)); 81d0833dfSYuchung Cheng } 91d0833dfSYuchung Cheng 101f7455c3Skbuild test robot static u32 tcp_rack_reo_wnd(const struct sock *sk) 1120b654dfSYuchung Cheng { 1220b654dfSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 1320b654dfSYuchung Cheng 147ec65372SWei Wang if (!tp->reord_seen) { 1520b654dfSYuchung Cheng /* If reordering has not been observed, be aggressive during 1620b654dfSYuchung Cheng * the recovery or starting the recovery by DUPACK threshold. 1720b654dfSYuchung Cheng */ 1820b654dfSYuchung Cheng if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery) 1920b654dfSYuchung Cheng return 0; 2020b654dfSYuchung Cheng 2120b654dfSYuchung Cheng if (tp->sacked_out >= tp->reordering && 2220b654dfSYuchung Cheng !(sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_NO_DUPTHRESH)) 2320b654dfSYuchung Cheng return 0; 2420b654dfSYuchung Cheng } 2520b654dfSYuchung Cheng 2620b654dfSYuchung Cheng /* To be more reordering resilient, allow min_rtt/4 settling delay. 2720b654dfSYuchung Cheng * Use min_rtt instead of the smoothed RTT because reordering is 2820b654dfSYuchung Cheng * often a path property and less related to queuing or delayed ACKs. 2920b654dfSYuchung Cheng * Upon receiving DSACKs, linearly increase the window up to the 3020b654dfSYuchung Cheng * smoothed RTT. 3120b654dfSYuchung Cheng */ 3220b654dfSYuchung Cheng return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps, 3320b654dfSYuchung Cheng tp->srtt_us >> 3); 3420b654dfSYuchung Cheng } 3520b654dfSYuchung Cheng 36b8fef65aSYuchung Cheng s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd) 37b8fef65aSYuchung Cheng { 38b8fef65aSYuchung Cheng return tp->rack.rtt_us + reo_wnd - 392fd66ffbSEric Dumazet tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb)); 40b8fef65aSYuchung Cheng } 41b8fef65aSYuchung Cheng 42a0370b3fSYuchung Cheng /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01): 43a0370b3fSYuchung Cheng * 44a0370b3fSYuchung Cheng * Marks a packet lost, if some packet sent later has been (s)acked. 454f41b1c5SYuchung Cheng * The underlying idea is similar to the traditional dupthresh and FACK 464f41b1c5SYuchung Cheng * but they look at different metrics: 474f41b1c5SYuchung Cheng * 484f41b1c5SYuchung Cheng * dupthresh: 3 OOO packets delivered (packet count) 494f41b1c5SYuchung Cheng * FACK: sequence delta to highest sacked sequence (sequence space) 504f41b1c5SYuchung Cheng * RACK: sent time delta to the latest delivered packet (time domain) 514f41b1c5SYuchung Cheng * 524f41b1c5SYuchung Cheng * The advantage of RACK is it applies to both original and retransmitted 534f41b1c5SYuchung Cheng * packet and therefore is robust against tail losses. Another advantage 544f41b1c5SYuchung Cheng * is being more resilient to reordering by simply allowing some 554f41b1c5SYuchung Cheng * "settling delay", instead of tweaking the dupthresh. 564f41b1c5SYuchung Cheng * 57a0370b3fSYuchung Cheng * When tcp_rack_detect_loss() detects some packets are lost and we 58a0370b3fSYuchung Cheng * are not already in the CA_Recovery state, either tcp_rack_reo_timeout() 59a0370b3fSYuchung Cheng * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will 60a0370b3fSYuchung Cheng * make us enter the CA_Recovery state. 614f41b1c5SYuchung Cheng */ 627c1c7308SEric Dumazet static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout) 634f41b1c5SYuchung Cheng { 644f41b1c5SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 65043b87d7SYuchung Cheng struct sk_buff *skb, *n; 66e636f8b0SYuchung Cheng u32 reo_wnd; 674f41b1c5SYuchung Cheng 6857dde7f7SYuchung Cheng *reo_timeout = 0; 6920b654dfSYuchung Cheng reo_wnd = tcp_rack_reo_wnd(sk); 70043b87d7SYuchung Cheng list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue, 71043b87d7SYuchung Cheng tcp_tsorted_anchor) { 724f41b1c5SYuchung Cheng struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 73bef06223SYuchung Cheng s32 remaining; 7457dde7f7SYuchung Cheng 7557dde7f7SYuchung Cheng /* Skip ones marked lost but not yet retransmitted */ 7657dde7f7SYuchung Cheng if ((scb->sacked & TCPCB_LOST) && 7757dde7f7SYuchung Cheng !(scb->sacked & TCPCB_SACKED_RETRANS)) 7857dde7f7SYuchung Cheng continue; 7957dde7f7SYuchung Cheng 802fd66ffbSEric Dumazet if (!tcp_rack_sent_after(tp->rack.mstamp, 812fd66ffbSEric Dumazet tcp_skb_timestamp_us(skb), 82bef06223SYuchung Cheng tp->rack.end_seq, scb->end_seq)) 83bef06223SYuchung Cheng break; 84bef06223SYuchung Cheng 85bef06223SYuchung Cheng /* A packet is lost if it has not been s/acked beyond 86bef06223SYuchung Cheng * the recent RTT plus the reordering window. 87bef06223SYuchung Cheng */ 88b8fef65aSYuchung Cheng remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd); 89428aec5eSYuchung Cheng if (remaining <= 0) { 90d716bfdbSYuchung Cheng tcp_mark_skb_lost(sk, skb); 91bef06223SYuchung Cheng list_del_init(&skb->tcp_tsorted_anchor); 92bef06223SYuchung Cheng } else { 93428aec5eSYuchung Cheng /* Record maximum wait time */ 94428aec5eSYuchung Cheng *reo_timeout = max_t(u32, *reo_timeout, remaining); 954f41b1c5SYuchung Cheng } 964f41b1c5SYuchung Cheng } 97e636f8b0SYuchung Cheng } 98e636f8b0SYuchung Cheng 9962d9f1a6SPengcheng Yang bool tcp_rack_mark_lost(struct sock *sk) 100e636f8b0SYuchung Cheng { 101e636f8b0SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 10257dde7f7SYuchung Cheng u32 timeout; 103e636f8b0SYuchung Cheng 104a0370b3fSYuchung Cheng if (!tp->rack.advanced) 10562d9f1a6SPengcheng Yang return false; 10657dde7f7SYuchung Cheng 107e636f8b0SYuchung Cheng /* Reset the advanced flag to avoid unnecessary queue scanning */ 108e636f8b0SYuchung Cheng tp->rack.advanced = 0; 1097c1c7308SEric Dumazet tcp_rack_detect_loss(sk, &timeout); 11057dde7f7SYuchung Cheng if (timeout) { 111bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN; 11257dde7f7SYuchung Cheng inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT, 11357dde7f7SYuchung Cheng timeout, inet_csk(sk)->icsk_rto); 11457dde7f7SYuchung Cheng } 11562d9f1a6SPengcheng Yang return !!timeout; 1164f41b1c5SYuchung Cheng } 1174f41b1c5SYuchung Cheng 118deed7be7SYuchung Cheng /* Record the most recently (re)sent time among the (s)acked packets 119deed7be7SYuchung Cheng * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from 120deed7be7SYuchung Cheng * draft-cheng-tcpm-rack-00.txt 121deed7be7SYuchung Cheng */ 1221d0833dfSYuchung Cheng void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, 1239a568de4SEric Dumazet u64 xmit_time) 124659a8ad5SYuchung Cheng { 125deed7be7SYuchung Cheng u32 rtt_us; 126deed7be7SYuchung Cheng 1279a568de4SEric Dumazet rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time); 1286065fd0dSYuchung Cheng if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) { 129659a8ad5SYuchung Cheng /* If the sacked packet was retransmitted, it's ambiguous 130659a8ad5SYuchung Cheng * whether the retransmission or the original (or the prior 131659a8ad5SYuchung Cheng * retransmission) was sacked. 132659a8ad5SYuchung Cheng * 133659a8ad5SYuchung Cheng * If the original is lost, there is no ambiguity. Otherwise 134659a8ad5SYuchung Cheng * we assume the original can be delayed up to aRTT + min_rtt. 135659a8ad5SYuchung Cheng * the aRTT term is bounded by the fast recovery or timeout, 136659a8ad5SYuchung Cheng * so it's at least one RTT (i.e., retransmission is at least 137659a8ad5SYuchung Cheng * an RTT later). 138659a8ad5SYuchung Cheng */ 139659a8ad5SYuchung Cheng return; 140659a8ad5SYuchung Cheng } 1416065fd0dSYuchung Cheng tp->rack.advanced = 1; 142deed7be7SYuchung Cheng tp->rack.rtt_us = rtt_us; 1436065fd0dSYuchung Cheng if (tcp_rack_sent_after(xmit_time, tp->rack.mstamp, 1446065fd0dSYuchung Cheng end_seq, tp->rack.end_seq)) { 1459a568de4SEric Dumazet tp->rack.mstamp = xmit_time; 1461d0833dfSYuchung Cheng tp->rack.end_seq = end_seq; 1476065fd0dSYuchung Cheng } 148659a8ad5SYuchung Cheng } 14957dde7f7SYuchung Cheng 15057dde7f7SYuchung Cheng /* We have waited long enough to accommodate reordering. Mark the expired 15157dde7f7SYuchung Cheng * packets lost and retransmit them. 15257dde7f7SYuchung Cheng */ 15357dde7f7SYuchung Cheng void tcp_rack_reo_timeout(struct sock *sk) 15457dde7f7SYuchung Cheng { 15557dde7f7SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 15657dde7f7SYuchung Cheng u32 timeout, prior_inflight; 1577e901ee7SYuchung Cheng u32 lost = tp->lost; 15857dde7f7SYuchung Cheng 15957dde7f7SYuchung Cheng prior_inflight = tcp_packets_in_flight(tp); 1607c1c7308SEric Dumazet tcp_rack_detect_loss(sk, &timeout); 16157dde7f7SYuchung Cheng if (prior_inflight != tcp_packets_in_flight(tp)) { 16257dde7f7SYuchung Cheng if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) { 16357dde7f7SYuchung Cheng tcp_enter_recovery(sk, false); 16457dde7f7SYuchung Cheng if (!inet_csk(sk)->icsk_ca_ops->cong_control) 1657e901ee7SYuchung Cheng tcp_cwnd_reduction(sk, 1, tp->lost - lost, 0); 16657dde7f7SYuchung Cheng } 16757dde7f7SYuchung Cheng tcp_xmit_retransmit_queue(sk); 16857dde7f7SYuchung Cheng } 16957dde7f7SYuchung Cheng if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS) 17057dde7f7SYuchung Cheng tcp_rearm_rto(sk); 17157dde7f7SYuchung Cheng } 1721f255691SPriyaranjan Jha 1731f255691SPriyaranjan Jha /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries. 1741f255691SPriyaranjan Jha * 175*a657db03SNeal Cardwell * If a DSACK is received that seems like it may have been due to reordering 176*a657db03SNeal Cardwell * triggering fast recovery, increment reo_wnd by min_rtt/4 (upper bounded 1771f255691SPriyaranjan Jha * by srtt), since there is possibility that spurious retransmission was 1781f255691SPriyaranjan Jha * due to reordering delay longer than reo_wnd. 1791f255691SPriyaranjan Jha * 1801f255691SPriyaranjan Jha * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16) 1811f255691SPriyaranjan Jha * no. of successful recoveries (accounts for full DSACK-based loss 1821f255691SPriyaranjan Jha * recovery undo). After that, reset it to default (min_rtt/4). 1831f255691SPriyaranjan Jha * 1841f255691SPriyaranjan Jha * At max, reo_wnd is incremented only once per rtt. So that the new 1851f255691SPriyaranjan Jha * DSACK on which we are reacting, is due to the spurious retx (approx) 1861f255691SPriyaranjan Jha * after the reo_wnd has been updated last time. 1871f255691SPriyaranjan Jha * 1881f255691SPriyaranjan Jha * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than 1891f255691SPriyaranjan Jha * absolute value to account for change in rtt. 1901f255691SPriyaranjan Jha */ 1911f255691SPriyaranjan Jha void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs) 1921f255691SPriyaranjan Jha { 1931f255691SPriyaranjan Jha struct tcp_sock *tp = tcp_sk(sk); 1941f255691SPriyaranjan Jha 1951f255691SPriyaranjan Jha if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_STATIC_REO_WND || 1961f255691SPriyaranjan Jha !rs->prior_delivered) 1971f255691SPriyaranjan Jha return; 1981f255691SPriyaranjan Jha 1991f255691SPriyaranjan Jha /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */ 2001f255691SPriyaranjan Jha if (before(rs->prior_delivered, tp->rack.last_delivered)) 2011f255691SPriyaranjan Jha tp->rack.dsack_seen = 0; 2021f255691SPriyaranjan Jha 2031f255691SPriyaranjan Jha /* Adjust the reo_wnd if update is pending */ 2041f255691SPriyaranjan Jha if (tp->rack.dsack_seen) { 2051f255691SPriyaranjan Jha tp->rack.reo_wnd_steps = min_t(u32, 0xFF, 2061f255691SPriyaranjan Jha tp->rack.reo_wnd_steps + 1); 2071f255691SPriyaranjan Jha tp->rack.dsack_seen = 0; 2081f255691SPriyaranjan Jha tp->rack.last_delivered = tp->delivered; 2091f255691SPriyaranjan Jha tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH; 2101f255691SPriyaranjan Jha } else if (!tp->rack.reo_wnd_persist) { 2111f255691SPriyaranjan Jha tp->rack.reo_wnd_steps = 1; 2121f255691SPriyaranjan Jha } 2131f255691SPriyaranjan Jha } 2146ac06ecdSYuchung Cheng 2156ac06ecdSYuchung Cheng /* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits 2166ac06ecdSYuchung Cheng * the next unacked packet upon receiving 2176ac06ecdSYuchung Cheng * a) three or more DUPACKs to start the fast recovery 2186ac06ecdSYuchung Cheng * b) an ACK acknowledging new data during the fast recovery. 2196ac06ecdSYuchung Cheng */ 2206ac06ecdSYuchung Cheng void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced) 2216ac06ecdSYuchung Cheng { 2226ac06ecdSYuchung Cheng const u8 state = inet_csk(sk)->icsk_ca_state; 2236ac06ecdSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 2246ac06ecdSYuchung Cheng 2256ac06ecdSYuchung Cheng if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) || 2266ac06ecdSYuchung Cheng (state == TCP_CA_Recovery && snd_una_advanced)) { 2276ac06ecdSYuchung Cheng struct sk_buff *skb = tcp_rtx_queue_head(sk); 2286ac06ecdSYuchung Cheng u32 mss; 2296ac06ecdSYuchung Cheng 2306ac06ecdSYuchung Cheng if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 2316ac06ecdSYuchung Cheng return; 2326ac06ecdSYuchung Cheng 2336ac06ecdSYuchung Cheng mss = tcp_skb_mss(skb); 2346ac06ecdSYuchung Cheng if (tcp_skb_pcount(skb) > 1 && skb->len > mss) 2356ac06ecdSYuchung Cheng tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 2366ac06ecdSYuchung Cheng mss, mss, GFP_ATOMIC); 2376ac06ecdSYuchung Cheng 238179ac35fSYuchung Cheng tcp_mark_skb_lost(sk, skb); 2396ac06ecdSYuchung Cheng } 2406ac06ecdSYuchung Cheng } 241