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