net/ipv4/tcp_recovery.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/tcp.h>
   3 #include <net/tcp.h>
   4
   5 static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
   6 {
   7         struct tcp_sock *tp = tcp_sk(sk);
   8
   9         tcp_skb_mark_lost_uncond_verify(tp, skb);
  10         if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
  11                 /* Account for retransmits that are lost again */
  12                 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
  13                 tp->retrans_out -= tcp_skb_pcount(skb);
  14                 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT,
  15                               tcp_skb_pcount(skb));
  16         }
  17 }
  18
  19 static bool tcp_rack_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2)
  20 {
  21         return t1 > t2 || (t1 == t2 && after(seq1, seq2));
  22 }
  23
  24 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
  25  *
  26  * Marks a packet lost, if some packet sent later has been (s)acked.
  27  * The underlying idea is similar to the traditional dupthresh and FACK
  28  * but they look at different metrics:
  29  *
  30  * dupthresh: 3 OOO packets delivered (packet count)
  31  * FACK: sequence delta to highest sacked sequence (sequence space)
  32  * RACK: sent time delta to the latest delivered packet (time domain)
  33  *
  34  * The advantage of RACK is it applies to both original and retransmitted
  35  * packet and therefore is robust against tail losses. Another advantage
  36  * is being more resilient to reordering by simply allowing some
  37  * "settling delay", instead of tweaking the dupthresh.
  38  *
  39  * When tcp_rack_detect_loss() detects some packets are lost and we
  40  * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
  41  * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
  42  * make us enter the CA_Recovery state.
  43  */
  44 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout)
  45 {
  46         struct tcp_sock *tp = tcp_sk(sk);
  47         u32 min_rtt = tcp_min_rtt(tp);
  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 || inet_csk(sk)->icsk_ca_state < TCP_CA_Recovery) &&
  59             min_rtt != ~0U) {
  60                 reo_wnd = max((min_rtt >> 2) * tp->rack.reo_wnd_steps, reo_wnd);
  61                 reo_wnd = min(reo_wnd, tp->srtt_us >> 3);
  62         }
  63
  64         list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue,
  65                                  tcp_tsorted_anchor) {
  66                 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
  67                 s32 remaining;
  68
  69                 /* Skip ones marked lost but not yet retransmitted */
  70                 if ((scb->sacked & TCPCB_LOST) &&
  71                     !(scb->sacked & TCPCB_SACKED_RETRANS))
  72                         continue;
  73
  74                 if (!tcp_rack_sent_after(tp->rack.mstamp, skb->skb_mstamp,
  75                                          tp->rack.end_seq, scb->end_seq))
  76                         break;
  77
  78                 /* A packet is lost if it has not been s/acked beyond
  79                  * the recent RTT plus the reordering window.
  80                  */
  81                 remaining = tp->rack.rtt_us + reo_wnd -
  82                             tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp);
  83                 if (remaining <= 0) {
  84                         tcp_rack_mark_skb_lost(sk, skb);
  85                         list_del_init(&skb->tcp_tsorted_anchor);
  86                 } else {
  87                         /* Record maximum wait time */
  88                         *reo_timeout = max_t(u32, *reo_timeout, remaining);
  89                 }
  90         }
  91 }
  92
  93 void tcp_rack_mark_lost(struct sock *sk)
  94 {
  95         struct tcp_sock *tp = tcp_sk(sk);
  96         u32 timeout;
  97
  98         if (!tp->rack.advanced)
  99                 return;
 100
 101         /* Reset the advanced flag to avoid unnecessary queue scanning */
 102         tp->rack.advanced = 0;
 103         tcp_rack_detect_loss(sk, &timeout);
 104         if (timeout) {
 105                 timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN;
 106                 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
 107                                           timeout, inet_csk(sk)->icsk_rto);
 108         }
 109 }
 110
 111 /* Record the most recently (re)sent time among the (s)acked packets
 112  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
 113  * draft-cheng-tcpm-rack-00.txt
 114  */
 115 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
 116                       u64 xmit_time)
 117 {
 118         u32 rtt_us;
 119
 120         rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
 121         if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) {
 122                 /* If the sacked packet was retransmitted, it's ambiguous
 123                  * whether the retransmission or the original (or the prior
 124                  * retransmission) was sacked.
 125                  *
 126                  * If the original is lost, there is no ambiguity. Otherwise
 127                  * we assume the original can be delayed up to aRTT + min_rtt.
 128                  * the aRTT term is bounded by the fast recovery or timeout,
 129                  * so it's at least one RTT (i.e., retransmission is at least
 130                  * an RTT later).
 131                  */
 132                 return;
 133         }
 134         tp->rack.advanced = 1;
 135         tp->rack.rtt_us = rtt_us;
 136         if (tcp_rack_sent_after(xmit_time, tp->rack.mstamp,
 137                                 end_seq, tp->rack.end_seq)) {
 138                 tp->rack.mstamp = xmit_time;
 139                 tp->rack.end_seq = end_seq;
 140         }
 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
 165 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries.
 166  *
 167  * If DSACK is received, increment reo_wnd by min_rtt/4 (upper bounded
 168  * by srtt), since there is possibility that spurious retransmission was
 169  * due to reordering delay longer than reo_wnd.
 170  *
 171  * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16)
 172  * no. of successful recoveries (accounts for full DSACK-based loss
 173  * recovery undo). After that, reset it to default (min_rtt/4).
 174  *
 175  * At max, reo_wnd is incremented only once per rtt. So that the new
 176  * DSACK on which we are reacting, is due to the spurious retx (approx)
 177  * after the reo_wnd has been updated last time.
 178  *
 179  * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than
 180  * absolute value to account for change in rtt.
 181  */
 182 void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs)
 183 {
 184         struct tcp_sock *tp = tcp_sk(sk);
 185
 186         if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_STATIC_REO_WND ||
 187             !rs->prior_delivered)
 188                 return;
 189
 190         /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */
 191         if (before(rs->prior_delivered, tp->rack.last_delivered))
 192                 tp->rack.dsack_seen = 0;
 193
 194         /* Adjust the reo_wnd if update is pending */
 195         if (tp->rack.dsack_seen) {
 196                 tp->rack.reo_wnd_steps = min_t(u32, 0xFF,
 197                                                tp->rack.reo_wnd_steps + 1);
 198                 tp->rack.dsack_seen = 0;
 199                 tp->rack.last_delivered = tp->delivered;
 200                 tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH;
 201         } else if (!tp->rack.reo_wnd_persist) {
 202                 tp->rack.reo_wnd_steps = 1;
 203         }
 204 }