1 // SPDX-License-Identifier: GPL-2.0
5 void tcp_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
7 struct tcp_sock *tp = tcp_sk(sk);
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,
19 static bool tcp_rack_sent_after(u64 t1, u64 t2, u32 seq1, u32 seq2)
21 return t1 > t2 || (t1 == t2 && after(seq1, seq2));
24 static u32 tcp_rack_reo_wnd(const struct sock *sk)
26 struct tcp_sock *tp = tcp_sk(sk);
28 if (!tp->reord_seen) {
29 /* If reordering has not been observed, be aggressive during
30 * the recovery or starting the recovery by DUPACK threshold.
32 if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery)
35 if (tp->sacked_out >= tp->reordering &&
36 !(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
37 TCP_RACK_NO_DUPTHRESH))
41 /* To be more reordering resilient, allow min_rtt/4 settling delay.
42 * Use min_rtt instead of the smoothed RTT because reordering is
43 * often a path property and less related to queuing or delayed ACKs.
44 * Upon receiving DSACKs, linearly increase the window up to the
47 return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps,
51 s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd)
53 return tp->rack.rtt_us + reo_wnd -
54 tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb));
57 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
59 * Marks a packet lost, if some packet sent later has been (s)acked.
60 * The underlying idea is similar to the traditional dupthresh and FACK
61 * but they look at different metrics:
63 * dupthresh: 3 OOO packets delivered (packet count)
64 * FACK: sequence delta to highest sacked sequence (sequence space)
65 * RACK: sent time delta to the latest delivered packet (time domain)
67 * The advantage of RACK is it applies to both original and retransmitted
68 * packet and therefore is robust against tail losses. Another advantage
69 * is being more resilient to reordering by simply allowing some
70 * "settling delay", instead of tweaking the dupthresh.
72 * When tcp_rack_detect_loss() detects some packets are lost and we
73 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
74 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
75 * make us enter the CA_Recovery state.
77 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout)
79 struct tcp_sock *tp = tcp_sk(sk);
80 struct sk_buff *skb, *n;
84 reo_wnd = tcp_rack_reo_wnd(sk);
85 list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue,
87 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
90 /* Skip ones marked lost but not yet retransmitted */
91 if ((scb->sacked & TCPCB_LOST) &&
92 !(scb->sacked & TCPCB_SACKED_RETRANS))
95 if (!tcp_rack_sent_after(tp->rack.mstamp,
96 tcp_skb_timestamp_us(skb),
97 tp->rack.end_seq, scb->end_seq))
100 /* A packet is lost if it has not been s/acked beyond
101 * the recent RTT plus the reordering window.
103 remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd);
104 if (remaining <= 0) {
105 tcp_mark_skb_lost(sk, skb);
106 list_del_init(&skb->tcp_tsorted_anchor);
108 /* Record maximum wait time */
109 *reo_timeout = max_t(u32, *reo_timeout, remaining);
114 bool tcp_rack_mark_lost(struct sock *sk)
116 struct tcp_sock *tp = tcp_sk(sk);
119 if (!tp->rack.advanced)
122 /* Reset the advanced flag to avoid unnecessary queue scanning */
123 tp->rack.advanced = 0;
124 tcp_rack_detect_loss(sk, &timeout);
126 timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN;
127 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
128 timeout, inet_csk(sk)->icsk_rto);
133 /* Record the most recently (re)sent time among the (s)acked packets
134 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
135 * draft-cheng-tcpm-rack-00.txt
137 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
142 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
143 if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) {
144 /* If the sacked packet was retransmitted, it's ambiguous
145 * whether the retransmission or the original (or the prior
146 * retransmission) was sacked.
148 * If the original is lost, there is no ambiguity. Otherwise
149 * we assume the original can be delayed up to aRTT + min_rtt.
150 * the aRTT term is bounded by the fast recovery or timeout,
151 * so it's at least one RTT (i.e., retransmission is at least
156 tp->rack.advanced = 1;
157 tp->rack.rtt_us = rtt_us;
158 if (tcp_rack_sent_after(xmit_time, tp->rack.mstamp,
159 end_seq, tp->rack.end_seq)) {
160 tp->rack.mstamp = xmit_time;
161 tp->rack.end_seq = end_seq;
165 /* We have waited long enough to accommodate reordering. Mark the expired
166 * packets lost and retransmit them.
168 void tcp_rack_reo_timeout(struct sock *sk)
170 struct tcp_sock *tp = tcp_sk(sk);
171 u32 timeout, prior_inflight;
173 prior_inflight = tcp_packets_in_flight(tp);
174 tcp_rack_detect_loss(sk, &timeout);
175 if (prior_inflight != tcp_packets_in_flight(tp)) {
176 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
177 tcp_enter_recovery(sk, false);
178 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
179 tcp_cwnd_reduction(sk, 1, 0);
181 tcp_xmit_retransmit_queue(sk);
183 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
187 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries.
189 * If DSACK is received, increment reo_wnd by min_rtt/4 (upper bounded
190 * by srtt), since there is possibility that spurious retransmission was
191 * due to reordering delay longer than reo_wnd.
193 * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16)
194 * no. of successful recoveries (accounts for full DSACK-based loss
195 * recovery undo). After that, reset it to default (min_rtt/4).
197 * At max, reo_wnd is incremented only once per rtt. So that the new
198 * DSACK on which we are reacting, is due to the spurious retx (approx)
199 * after the reo_wnd has been updated last time.
201 * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than
202 * absolute value to account for change in rtt.
204 void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs)
206 struct tcp_sock *tp = tcp_sk(sk);
208 if ((READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_recovery) &
209 TCP_RACK_STATIC_REO_WND) ||
210 !rs->prior_delivered)
213 /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */
214 if (before(rs->prior_delivered, tp->rack.last_delivered))
215 tp->rack.dsack_seen = 0;
217 /* Adjust the reo_wnd if update is pending */
218 if (tp->rack.dsack_seen) {
219 tp->rack.reo_wnd_steps = min_t(u32, 0xFF,
220 tp->rack.reo_wnd_steps + 1);
221 tp->rack.dsack_seen = 0;
222 tp->rack.last_delivered = tp->delivered;
223 tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH;
224 } else if (!tp->rack.reo_wnd_persist) {
225 tp->rack.reo_wnd_steps = 1;
229 /* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits
230 * the next unacked packet upon receiving
231 * a) three or more DUPACKs to start the fast recovery
232 * b) an ACK acknowledging new data during the fast recovery.
234 void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced)
236 const u8 state = inet_csk(sk)->icsk_ca_state;
237 struct tcp_sock *tp = tcp_sk(sk);
239 if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) ||
240 (state == TCP_CA_Recovery && snd_una_advanced)) {
241 struct sk_buff *skb = tcp_rtx_queue_head(sk);
244 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
247 mss = tcp_skb_mss(skb);
248 if (tcp_skb_pcount(skb) > 1 && skb->len > mss)
249 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
250 mss, mss, GFP_ATOMIC);
252 tcp_skb_mark_lost_uncond_verify(tp, skb);