1 // SPDX-License-Identifier: GPL-2.0
5 static u32 tcp_rack_reo_wnd(const struct sock *sk)
7 struct tcp_sock *tp = tcp_sk(sk);
10 /* If reordering has not been observed, be aggressive during
11 * the recovery or starting the recovery by DUPACK threshold.
13 if (inet_csk(sk)->icsk_ca_state >= TCP_CA_Recovery)
16 if (tp->sacked_out >= tp->reordering &&
17 !(sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_NO_DUPTHRESH))
21 /* To be more reordering resilient, allow min_rtt/4 settling delay.
22 * Use min_rtt instead of the smoothed RTT because reordering is
23 * often a path property and less related to queuing or delayed ACKs.
24 * Upon receiving DSACKs, linearly increase the window up to the
27 return min((tcp_min_rtt(tp) >> 2) * tp->rack.reo_wnd_steps,
31 s32 tcp_rack_skb_timeout(struct tcp_sock *tp, struct sk_buff *skb, u32 reo_wnd)
33 return tp->rack.rtt_us + reo_wnd -
34 tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(skb));
37 /* RACK loss detection (IETF draft draft-ietf-tcpm-rack-01):
39 * Marks a packet lost, if some packet sent later has been (s)acked.
40 * The underlying idea is similar to the traditional dupthresh and FACK
41 * but they look at different metrics:
43 * dupthresh: 3 OOO packets delivered (packet count)
44 * FACK: sequence delta to highest sacked sequence (sequence space)
45 * RACK: sent time delta to the latest delivered packet (time domain)
47 * The advantage of RACK is it applies to both original and retransmitted
48 * packet and therefore is robust against tail losses. Another advantage
49 * is being more resilient to reordering by simply allowing some
50 * "settling delay", instead of tweaking the dupthresh.
52 * When tcp_rack_detect_loss() detects some packets are lost and we
53 * are not already in the CA_Recovery state, either tcp_rack_reo_timeout()
54 * or tcp_time_to_recover()'s "Trick#1: the loss is proven" code path will
55 * make us enter the CA_Recovery state.
57 static void tcp_rack_detect_loss(struct sock *sk, u32 *reo_timeout)
59 struct tcp_sock *tp = tcp_sk(sk);
60 struct sk_buff *skb, *n;
64 reo_wnd = tcp_rack_reo_wnd(sk);
65 list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue,
67 struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
70 /* Skip ones marked lost but not yet retransmitted */
71 if ((scb->sacked & TCPCB_LOST) &&
72 !(scb->sacked & TCPCB_SACKED_RETRANS))
75 if (!tcp_skb_sent_after(tp->rack.mstamp,
76 tcp_skb_timestamp_us(skb),
77 tp->rack.end_seq, scb->end_seq))
80 /* A packet is lost if it has not been s/acked beyond
81 * the recent RTT plus the reordering window.
83 remaining = tcp_rack_skb_timeout(tp, skb, reo_wnd);
85 tcp_mark_skb_lost(sk, skb);
86 list_del_init(&skb->tcp_tsorted_anchor);
88 /* Record maximum wait time */
89 *reo_timeout = max_t(u32, *reo_timeout, remaining);
94 bool tcp_rack_mark_lost(struct sock *sk)
96 struct tcp_sock *tp = tcp_sk(sk);
99 if (!tp->rack.advanced)
102 /* Reset the advanced flag to avoid unnecessary queue scanning */
103 tp->rack.advanced = 0;
104 tcp_rack_detect_loss(sk, &timeout);
106 timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN;
107 inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
108 timeout, inet_csk(sk)->icsk_rto);
113 /* Record the most recently (re)sent time among the (s)acked packets
114 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
115 * draft-cheng-tcpm-rack-00.txt
117 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
122 rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
123 if (rtt_us < tcp_min_rtt(tp) && (sacked & TCPCB_RETRANS)) {
124 /* If the sacked packet was retransmitted, it's ambiguous
125 * whether the retransmission or the original (or the prior
126 * retransmission) was sacked.
128 * If the original is lost, there is no ambiguity. Otherwise
129 * we assume the original can be delayed up to aRTT + min_rtt.
130 * the aRTT term is bounded by the fast recovery or timeout,
131 * so it's at least one RTT (i.e., retransmission is at least
136 tp->rack.advanced = 1;
137 tp->rack.rtt_us = rtt_us;
138 if (tcp_skb_sent_after(xmit_time, tp->rack.mstamp,
139 end_seq, tp->rack.end_seq)) {
140 tp->rack.mstamp = xmit_time;
141 tp->rack.end_seq = end_seq;
145 /* We have waited long enough to accommodate reordering. Mark the expired
146 * packets lost and retransmit them.
148 void tcp_rack_reo_timeout(struct sock *sk)
150 struct tcp_sock *tp = tcp_sk(sk);
151 u32 timeout, prior_inflight;
154 prior_inflight = tcp_packets_in_flight(tp);
155 tcp_rack_detect_loss(sk, &timeout);
156 if (prior_inflight != tcp_packets_in_flight(tp)) {
157 if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
158 tcp_enter_recovery(sk, false);
159 if (!inet_csk(sk)->icsk_ca_ops->cong_control)
160 tcp_cwnd_reduction(sk, 1, tp->lost - lost, 0);
162 tcp_xmit_retransmit_queue(sk);
164 if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
168 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries.
170 * If a DSACK is received that seems like it may have been due to reordering
171 * triggering fast recovery, increment reo_wnd by min_rtt/4 (upper bounded
172 * by srtt), since there is possibility that spurious retransmission was
173 * due to reordering delay longer than reo_wnd.
175 * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16)
176 * no. of successful recoveries (accounts for full DSACK-based loss
177 * recovery undo). After that, reset it to default (min_rtt/4).
179 * At max, reo_wnd is incremented only once per rtt. So that the new
180 * DSACK on which we are reacting, is due to the spurious retx (approx)
181 * after the reo_wnd has been updated last time.
183 * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than
184 * absolute value to account for change in rtt.
186 void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs)
188 struct tcp_sock *tp = tcp_sk(sk);
190 if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_STATIC_REO_WND ||
191 !rs->prior_delivered)
194 /* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */
195 if (before(rs->prior_delivered, tp->rack.last_delivered))
196 tp->rack.dsack_seen = 0;
198 /* Adjust the reo_wnd if update is pending */
199 if (tp->rack.dsack_seen) {
200 tp->rack.reo_wnd_steps = min_t(u32, 0xFF,
201 tp->rack.reo_wnd_steps + 1);
202 tp->rack.dsack_seen = 0;
203 tp->rack.last_delivered = tp->delivered;
204 tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH;
205 } else if (!tp->rack.reo_wnd_persist) {
206 tp->rack.reo_wnd_steps = 1;
210 /* RFC6582 NewReno recovery for non-SACK connection. It simply retransmits
211 * the next unacked packet upon receiving
212 * a) three or more DUPACKs to start the fast recovery
213 * b) an ACK acknowledging new data during the fast recovery.
215 void tcp_newreno_mark_lost(struct sock *sk, bool snd_una_advanced)
217 const u8 state = inet_csk(sk)->icsk_ca_state;
218 struct tcp_sock *tp = tcp_sk(sk);
220 if ((state < TCP_CA_Recovery && tp->sacked_out >= tp->reordering) ||
221 (state == TCP_CA_Recovery && snd_una_advanced)) {
222 struct sk_buff *skb = tcp_rtx_queue_head(sk);
225 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
228 mss = tcp_skb_mss(skb);
229 if (tcp_skb_pcount(skb) > 1 && skb->len > mss)
230 tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
231 mss, mss, GFP_ATOMIC);
233 tcp_mark_skb_lost(sk, skb);