xref: /openbmc/linux/net/ipv4/tcp_recovery.c (revision e639c869)
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 || !tp->lost_out) && min_rtt != ~0U) {
59 		reo_wnd = max((min_rtt >> 2) * tp->rack.reo_wnd_steps, reo_wnd);
60 		reo_wnd = min(reo_wnd, tp->srtt_us >> 3);
61 	}
62 
63 	list_for_each_entry_safe(skb, n, &tp->tsorted_sent_queue,
64 				 tcp_tsorted_anchor) {
65 		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
66 		s32 remaining;
67 
68 		/* Skip ones marked lost but not yet retransmitted */
69 		if ((scb->sacked & TCPCB_LOST) &&
70 		    !(scb->sacked & TCPCB_SACKED_RETRANS))
71 			continue;
72 
73 		if (!tcp_rack_sent_after(tp->rack.mstamp, skb->skb_mstamp,
74 					 tp->rack.end_seq, scb->end_seq))
75 			break;
76 
77 		/* A packet is lost if it has not been s/acked beyond
78 		 * the recent RTT plus the reordering window.
79 		 */
80 		remaining = tp->rack.rtt_us + reo_wnd -
81 			    tcp_stamp_us_delta(tp->tcp_mstamp, skb->skb_mstamp);
82 		if (remaining < 0) {
83 			tcp_rack_mark_skb_lost(sk, skb);
84 			list_del_init(&skb->tcp_tsorted_anchor);
85 		} else {
86 			/* Record maximum wait time (+1 to avoid 0) */
87 			*reo_timeout = max_t(u32, *reo_timeout, 1 + remaining);
88 		}
89 	}
90 }
91 
92 void tcp_rack_mark_lost(struct sock *sk)
93 {
94 	struct tcp_sock *tp = tcp_sk(sk);
95 	u32 timeout;
96 
97 	if (!tp->rack.advanced)
98 		return;
99 
100 	/* Reset the advanced flag to avoid unnecessary queue scanning */
101 	tp->rack.advanced = 0;
102 	tcp_rack_detect_loss(sk, &timeout);
103 	if (timeout) {
104 		timeout = usecs_to_jiffies(timeout) + TCP_TIMEOUT_MIN;
105 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_REO_TIMEOUT,
106 					  timeout, inet_csk(sk)->icsk_rto);
107 	}
108 }
109 
110 /* Record the most recently (re)sent time among the (s)acked packets
111  * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
112  * draft-cheng-tcpm-rack-00.txt
113  */
114 void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq,
115 		      u64 xmit_time)
116 {
117 	u32 rtt_us;
118 
119 	if (tp->rack.mstamp &&
120 	    !tcp_rack_sent_after(xmit_time, tp->rack.mstamp,
121 				 end_seq, tp->rack.end_seq))
122 		return;
123 
124 	rtt_us = tcp_stamp_us_delta(tp->tcp_mstamp, xmit_time);
125 	if (sacked & TCPCB_RETRANS) {
126 		/* If the sacked packet was retransmitted, it's ambiguous
127 		 * whether the retransmission or the original (or the prior
128 		 * retransmission) was sacked.
129 		 *
130 		 * If the original is lost, there is no ambiguity. Otherwise
131 		 * we assume the original can be delayed up to aRTT + min_rtt.
132 		 * the aRTT term is bounded by the fast recovery or timeout,
133 		 * so it's at least one RTT (i.e., retransmission is at least
134 		 * an RTT later).
135 		 */
136 		if (rtt_us < tcp_min_rtt(tp))
137 			return;
138 	}
139 	tp->rack.rtt_us = rtt_us;
140 	tp->rack.mstamp = xmit_time;
141 	tp->rack.end_seq = end_seq;
142 	tp->rack.advanced = 1;
143 }
144 
145 /* We have waited long enough to accommodate reordering. Mark the expired
146  * packets lost and retransmit them.
147  */
148 void tcp_rack_reo_timeout(struct sock *sk)
149 {
150 	struct tcp_sock *tp = tcp_sk(sk);
151 	u32 timeout, prior_inflight;
152 
153 	prior_inflight = tcp_packets_in_flight(tp);
154 	tcp_rack_detect_loss(sk, &timeout);
155 	if (prior_inflight != tcp_packets_in_flight(tp)) {
156 		if (inet_csk(sk)->icsk_ca_state != TCP_CA_Recovery) {
157 			tcp_enter_recovery(sk, false);
158 			if (!inet_csk(sk)->icsk_ca_ops->cong_control)
159 				tcp_cwnd_reduction(sk, 1, 0);
160 		}
161 		tcp_xmit_retransmit_queue(sk);
162 	}
163 	if (inet_csk(sk)->icsk_pending != ICSK_TIME_RETRANS)
164 		tcp_rearm_rto(sk);
165 }
166 
167 /* Updates the RACK's reo_wnd based on DSACK and no. of recoveries.
168  *
169  * If DSACK is received, increment reo_wnd by min_rtt/4 (upper bounded
170  * by srtt), since there is possibility that spurious retransmission was
171  * due to reordering delay longer than reo_wnd.
172  *
173  * Persist the current reo_wnd value for TCP_RACK_RECOVERY_THRESH (16)
174  * no. of successful recoveries (accounts for full DSACK-based loss
175  * recovery undo). After that, reset it to default (min_rtt/4).
176  *
177  * At max, reo_wnd is incremented only once per rtt. So that the new
178  * DSACK on which we are reacting, is due to the spurious retx (approx)
179  * after the reo_wnd has been updated last time.
180  *
181  * reo_wnd is tracked in terms of steps (of min_rtt/4), rather than
182  * absolute value to account for change in rtt.
183  */
184 void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs)
185 {
186 	struct tcp_sock *tp = tcp_sk(sk);
187 
188 	if (sock_net(sk)->ipv4.sysctl_tcp_recovery & TCP_RACK_STATIC_REO_WND ||
189 	    !rs->prior_delivered)
190 		return;
191 
192 	/* Disregard DSACK if a rtt has not passed since we adjusted reo_wnd */
193 	if (before(rs->prior_delivered, tp->rack.last_delivered))
194 		tp->rack.dsack_seen = 0;
195 
196 	/* Adjust the reo_wnd if update is pending */
197 	if (tp->rack.dsack_seen) {
198 		tp->rack.reo_wnd_steps = min_t(u32, 0xFF,
199 					       tp->rack.reo_wnd_steps + 1);
200 		tp->rack.dsack_seen = 0;
201 		tp->rack.last_delivered = tp->delivered;
202 		tp->rack.reo_wnd_persist = TCP_RACK_RECOVERY_THRESH;
203 	} else if (!tp->rack.reo_wnd_persist) {
204 		tp->rack.reo_wnd_steps = 1;
205 	}
206 }
207