net/ipv4/tcp_recovery.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 #include <linux/tcp.h>
0003 #include <net/tcp.h>
0004
0005 static u32 tcp_rack_reo_wnd(const struct sock *sk)
0006 {
0007     struct tcp_sock *tp = tcp_sk(sk);
0008
0009     if (!tp->reord_seen) {
0010         /* If reordering has not been observed, be aggressive during
0011          * the recovery or starting the recovery by DUPACK threshold.
0012          */
0013         if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery)
0014             return 0;
0015
0016         if (tp->sacked_out >= tp->reordering &&
0017             !(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
0018               TCP_RACK_NO_DUPTHRESH))
0019             return 0;
0020     }
0021
0022     /* To be more reordering resilient, allow min_rtt/4 settling delay.
0023      * Use min_rtt instead of the smoothed RTT because reordering is
0024      * often a path property and less related to queuing or delayed ACKs.
0025      * Upon receiving DSACKs, linearly increase the window up to the
0026      * smoothed RTT.
0027      */
0028     return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps,
0029            tp->srtt_us >> 3);
0030 }
0031
0032 s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd)
0033 {
0034     return tp->rack.rtt_us + reo_wnd -
0035            tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb));
0036 }
0037
0038 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
0039  *
0040  * Marks a packet lost, if some packet sent later has been (s)acked.
0041  * The underlying idea is similar to the traditional dupthresh and FACK
0042  * but they look at different metrics:
0043  *
0044  * dupthresh: 3 OOO packets delivered (packet count)
0045  * FACK: sequence delta to highest sacked sequence (sequence space)
0046  * RACK: sent time delta to the latest delivered packet (time domain)
0047  *
0048  * The advantage of RACK is it applies to both original and retransmitted
0049  * packet and therefore is robust against tail losses. Another advantage
0050  * is being more resilient to reordering by simply allowing some
0051  * "settling delay", instead of tweaking the dupthresh.
0052  *
0053  * When tcp_rack_detect_loss() detects some packets are lost and we
0054  * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
0055  * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
0056  * make us enter the CA_Recovery state.
0057  */
0058 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout)
0059 {
0060     struct tcp_sock *tp = tcp_sk(sk);
0061     struct sk_buff *skb, *n;
0062     u32 reo_wnd;
0063
0064     *reo_timeout = 0;
0065     reo_wnd = tcp_rack_reo_wnd(sk);
0066     list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue,
0067                  tcp_tsorted_anchor) {
0068         struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
0069         s32 remaining;
0070
0071         /* Skip ones marked lost but not yet retransmitted */
0072         if ((scb->sacked & TCPCB_LOST) &&
0073             !(scb->sacked & TCPCB_SACKED_RETRANS))
0074             continue;
0075
0076         if (!tcp_skb_sent_after(tp->rack.mstamp,
0077                     tcp_skb_timestamp_us(skb),
0078                     tp->rack.end_seq, scb->end_seq))
0079             break;
0080
0081         /* A packet is lost if it has not been s/acked beyond
0082          * the recent RTT plus the reordering window.
0083          */
0084         remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd);
0085         if (remaining <= 0) {
0086             tcp_mark_skb_lost(sk, skb);
0087             list_del_init(&skb->tcp_tsorted_anchor);
0088         } else {
0089             /* Record maximum wait time */
0090             *reo_timeout = max_t(u32, *reo_timeout, remaining);
0091         }
0092     }
0093 }
0094
0095 bool tcp_rack_mark_lost(struct sock *sk)
0096 {
0097     struct tcp_sock *tp = tcp_sk(sk);
0098     u32 timeout;
0099
0100     if (!tp->rack.advanced)
0101         return false;
0102
0103     /* Reset the advanced flag to avoid unnecessary queue scanning */
0104     tp->rack.advanced = 0;
0105     tcp_rack_detect_loss(sk, &timeout);
0106     if (timeout) {
0107         timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN;
0108         inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
0109                       timeout, inet_csk(sk)->icsk_rto);
0110     }
0111     return !!timeout;
0112 }
0113
0114 /* Record the most recently (re)sent time among the (s)acked packets
0115  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
0116  * draft-cheng-tcpm-rack-00.txt
0117  */
0118 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
0119               u64 xmit_time)
0120 {
0121     u32 rtt_us;
0122
0123     rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
0124     if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) {
0125         /* If the sacked packet was retransmitted, it's ambiguous
0126          * whether the retransmission or the original (or the prior
0127          * retransmission) was sacked.
0128          *
0129          * If the original is lost, there is no ambiguity. Otherwise
0130          * we assume the original can be delayed up to aRTT + min_rtt.
0131          * the aRTT term is bounded by the fast recovery or timeout,
0132          * so it's at least one RTT (i.e., retransmission is at least
0133          * an RTT later).
0134          */
0135         return;
0136     }
0137     tp->rack.advanced = 1;
0138     tp->rack.rtt_us = rtt_us;
0139     if (tcp_skb_sent_after(xmit_time, tp->rack.mstamp,
0140                    end_seq, tp->rack.end_seq)) {
0141         tp->rack.mstamp = xmit_time;
0142         tp->rack.end_seq = end_seq;
0143     }
0144 }
0145
0146 /* We have waited long enough to accommodate reordering. Mark the expired
0147  * packets lost and retransmit them.
0148  */
0149 void tcp_rack_reo_timeout(struct sock *sk)
0150 {
0151     struct tcp_sock *tp = tcp_sk(sk);
0152     u32 timeout, prior_inflight;
0153     u32 lost = tp->lost;
0154
0155     prior_inflight = tcp_packets_in_flight(tp);
0156     tcp_rack_detect_loss(sk, &timeout);
0157     if (prior_inflight != tcp_packets_in_flight(tp)) {
0158         if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
0159             tcp_enter_recovery(sk, false);
0160             if (!inet_csk(sk)->icsk_ca_ops->cong_control)
0161                 tcp_cwnd_reduction(sk, 1, tp->lost - lost, 0);
0162         }
0163         tcp_xmit_retransmit_queue(sk);
0164     }
0165     if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
0166         tcp_rearm_rto(sk);
0167 }
0168
0169 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries.
0170  *
0171  * If a DSACK is received that seems like it may have been due to reordering
0172  * triggering fast recovery, increment reo_wnd by min_rtt/4 (upper bounded
0173  * by srtt), since there is possibility that spurious retransmission was
0174  * due to reordering delay longer than reo_wnd.
0175  *
0176  * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16)
0177  * no. of successful recoveries (accounts for full DSACK-based loss
0178  * recovery undo). After that, reset it to default (min_rtt/4).
0179  *
0180  * At max, reo_wnd is incremented only once per rtt. So that the new
0181  * DSACK on which we are reacting, is due to the spurious retx (approx)
0182  * after the reo_wnd has been updated last time.
0183  *
0184  * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than
0185  * absolute value to account for change in rtt.
0186  */
0187 void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs)
0188 {
0189     struct tcp_sock *tp = tcp_sk(sk);
0190
0191     if ((READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
0192          TCP_RACK_STATIC_REO_WND) ||
0193         !rs->prior_delivered)
0194         return;
0195
0196     /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */
0197     if (before(rs->prior_delivered, tp->rack.last_delivered))
0198         tp->rack.dsack_seen = 0;
0199
0200     /* Adjust the reo_wnd if update is pending */
0201     if (tp->rack.dsack_seen) {
0202         tp->rack.reo_wnd_steps = min_t(u32, 0xFF,
0203                            tp->rack.reo_wnd_steps + 1);
0204         tp->rack.dsack_seen = 0;
0205         tp->rack.last_delivered = tp->delivered;
0206         tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH;
0207     } else if (!tp->rack.reo_wnd_persist) {
0208         tp->rack.reo_wnd_steps = 1;
0209     }
0210 }
0211
0212 /* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits
0213  * the next unacked packet upon receiving
0214  * a) three or more DUPACKs to start the fast recovery
0215  * b) an ACK acknowledging new data during the fast recovery.
0216  */
0217 void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced)
0218 {
0219     const u8 state = inet_csk(sk)->icsk_ca_state;
0220     struct tcp_sock *tp = tcp_sk(sk);
0221
0222     if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) ||
0223         (state == TCP_CA_Recovery && snd_una_advanced)) {
0224         struct sk_buff *skb = tcp_rtx_queue_head(sk);
0225         u32 mss;
0226
0227         if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
0228             return;
0229
0230         mss = tcp_skb_mss(skb);
0231         if (tcp_skb_pcount(skb) > 1 && skb->len > mss)
0232             tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
0233                      mss, mss, GFP_ATOMIC);
0234
0235         tcp_mark_skb_lost(sk, skb);
0236     }
0237 }