1659a8ad5SYuchung Cheng #include <linux/tcp.h> 2659a8ad5SYuchung Cheng #include <net/tcp.h> 3659a8ad5SYuchung Cheng 44f41b1c5SYuchung Cheng int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOST_RETRANS; 54f41b1c5SYuchung Cheng 6db8da6bbSYuchung Cheng static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb) 7db8da6bbSYuchung Cheng { 8db8da6bbSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 9db8da6bbSYuchung Cheng 10db8da6bbSYuchung Cheng tcp_skb_mark_lost_uncond_verify(tp, skb); 11db8da6bbSYuchung Cheng if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 12db8da6bbSYuchung Cheng /* Account for retransmits that are lost again */ 13db8da6bbSYuchung Cheng TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 14db8da6bbSYuchung Cheng tp->retrans_out -= tcp_skb_pcount(skb); 15db8da6bbSYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); 16db8da6bbSYuchung Cheng } 17db8da6bbSYuchung Cheng } 18db8da6bbSYuchung Cheng 194f41b1c5SYuchung Cheng /* Marks a packet lost, if some packet sent later has been (s)acked. 204f41b1c5SYuchung Cheng * The underlying idea is similar to the traditional dupthresh and FACK 214f41b1c5SYuchung Cheng * but they look at different metrics: 224f41b1c5SYuchung Cheng * 234f41b1c5SYuchung Cheng * dupthresh: 3 OOO packets delivered (packet count) 244f41b1c5SYuchung Cheng * FACK: sequence delta to highest sacked sequence (sequence space) 254f41b1c5SYuchung Cheng * RACK: sent time delta to the latest delivered packet (time domain) 264f41b1c5SYuchung Cheng * 274f41b1c5SYuchung Cheng * The advantage of RACK is it applies to both original and retransmitted 284f41b1c5SYuchung Cheng * packet and therefore is robust against tail losses. Another advantage 294f41b1c5SYuchung Cheng * is being more resilient to reordering by simply allowing some 304f41b1c5SYuchung Cheng * "settling delay", instead of tweaking the dupthresh. 314f41b1c5SYuchung Cheng * 324f41b1c5SYuchung Cheng * The current version is only used after recovery starts but can be 334f41b1c5SYuchung Cheng * easily extended to detect the first loss. 344f41b1c5SYuchung Cheng */ 35*e636f8b0SYuchung Cheng static void tcp_rack_detect_loss(struct sock *sk) 364f41b1c5SYuchung Cheng { 374f41b1c5SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 384f41b1c5SYuchung Cheng struct sk_buff *skb; 39*e636f8b0SYuchung Cheng u32 reo_wnd; 404f41b1c5SYuchung Cheng 414f41b1c5SYuchung Cheng /* To be more reordering resilient, allow min_rtt/4 settling delay 424f41b1c5SYuchung Cheng * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed 434f41b1c5SYuchung Cheng * RTT because reordering is often a path property and less related 444f41b1c5SYuchung Cheng * to queuing or delayed ACKs. 454f41b1c5SYuchung Cheng * 464f41b1c5SYuchung Cheng * TODO: measure and adapt to the observed reordering delay, and 474f41b1c5SYuchung Cheng * use a timer to retransmit like the delayed early retransmit. 484f41b1c5SYuchung Cheng */ 494f41b1c5SYuchung Cheng reo_wnd = 1000; 504f41b1c5SYuchung Cheng if (tp->rack.reord && tcp_min_rtt(tp) != ~0U) 514f41b1c5SYuchung Cheng reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd); 524f41b1c5SYuchung Cheng 534f41b1c5SYuchung Cheng tcp_for_write_queue(skb, sk) { 544f41b1c5SYuchung Cheng struct tcp_skb_cb *scb = TCP_SKB_CB(skb); 554f41b1c5SYuchung Cheng 564f41b1c5SYuchung Cheng if (skb == tcp_send_head(sk)) 574f41b1c5SYuchung Cheng break; 584f41b1c5SYuchung Cheng 594f41b1c5SYuchung Cheng /* Skip ones already (s)acked */ 604f41b1c5SYuchung Cheng if (!after(scb->end_seq, tp->snd_una) || 614f41b1c5SYuchung Cheng scb->sacked & TCPCB_SACKED_ACKED) 624f41b1c5SYuchung Cheng continue; 634f41b1c5SYuchung Cheng 644f41b1c5SYuchung Cheng if (skb_mstamp_after(&tp->rack.mstamp, &skb->skb_mstamp)) { 654f41b1c5SYuchung Cheng 664f41b1c5SYuchung Cheng if (skb_mstamp_us_delta(&tp->rack.mstamp, 674f41b1c5SYuchung Cheng &skb->skb_mstamp) <= reo_wnd) 684f41b1c5SYuchung Cheng continue; 694f41b1c5SYuchung Cheng 704f41b1c5SYuchung Cheng /* skb is lost if packet sent later is sacked */ 71db8da6bbSYuchung Cheng tcp_rack_mark_skb_lost(sk, skb); 724f41b1c5SYuchung Cheng } else if (!(scb->sacked & TCPCB_RETRANS)) { 734f41b1c5SYuchung Cheng /* Original data are sent sequentially so stop early 744f41b1c5SYuchung Cheng * b/c the rest are all sent after rack_sent 754f41b1c5SYuchung Cheng */ 764f41b1c5SYuchung Cheng break; 774f41b1c5SYuchung Cheng } 784f41b1c5SYuchung Cheng } 79*e636f8b0SYuchung Cheng } 80*e636f8b0SYuchung Cheng 81*e636f8b0SYuchung Cheng void tcp_rack_mark_lost(struct sock *sk) 82*e636f8b0SYuchung Cheng { 83*e636f8b0SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 84*e636f8b0SYuchung Cheng 85*e636f8b0SYuchung Cheng if (inet_csk(sk)->icsk_ca_state < TCP_CA_Recovery || !tp->rack.advanced) 86*e636f8b0SYuchung Cheng return; 87*e636f8b0SYuchung Cheng /* Reset the advanced flag to avoid unnecessary queue scanning */ 88*e636f8b0SYuchung Cheng tp->rack.advanced = 0; 89*e636f8b0SYuchung Cheng tcp_rack_detect_loss(sk); 904f41b1c5SYuchung Cheng } 914f41b1c5SYuchung Cheng 92659a8ad5SYuchung Cheng /* Record the most recently (re)sent time among the (s)acked packets */ 93659a8ad5SYuchung Cheng void tcp_rack_advance(struct tcp_sock *tp, 94659a8ad5SYuchung Cheng const struct skb_mstamp *xmit_time, u8 sacked) 95659a8ad5SYuchung Cheng { 96659a8ad5SYuchung Cheng if (tp->rack.mstamp.v64 && 97659a8ad5SYuchung Cheng !skb_mstamp_after(xmit_time, &tp->rack.mstamp)) 98659a8ad5SYuchung Cheng return; 99659a8ad5SYuchung Cheng 100659a8ad5SYuchung Cheng if (sacked & TCPCB_RETRANS) { 101659a8ad5SYuchung Cheng struct skb_mstamp now; 102659a8ad5SYuchung Cheng 103659a8ad5SYuchung Cheng /* If the sacked packet was retransmitted, it's ambiguous 104659a8ad5SYuchung Cheng * whether the retransmission or the original (or the prior 105659a8ad5SYuchung Cheng * retransmission) was sacked. 106659a8ad5SYuchung Cheng * 107659a8ad5SYuchung Cheng * If the original is lost, there is no ambiguity. Otherwise 108659a8ad5SYuchung Cheng * we assume the original can be delayed up to aRTT + min_rtt. 109659a8ad5SYuchung Cheng * the aRTT term is bounded by the fast recovery or timeout, 110659a8ad5SYuchung Cheng * so it's at least one RTT (i.e., retransmission is at least 111659a8ad5SYuchung Cheng * an RTT later). 112659a8ad5SYuchung Cheng */ 113659a8ad5SYuchung Cheng skb_mstamp_get(&now); 114659a8ad5SYuchung Cheng if (skb_mstamp_us_delta(&now, xmit_time) < tcp_min_rtt(tp)) 115659a8ad5SYuchung Cheng return; 116659a8ad5SYuchung Cheng } 117659a8ad5SYuchung Cheng 118659a8ad5SYuchung Cheng tp->rack.mstamp = *xmit_time; 119659a8ad5SYuchung Cheng tp->rack.advanced = 1; 120659a8ad5SYuchung Cheng } 121