summaryrefslogtreecommitdiffstats
path: root/net/ipv4/tcp_recovery.c
blob: 557363cde58abb87aec526c935323dfea9eb9229 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
#include <linux/tcp.h>
#include <net/tcp.h>

int sysctl_tcp_recovery __read_mostly = TCP_RACK_LOST_RETRANS;

static void tcp_rack_mark_skb_lost(struct sock *sk, struct sk_buff *skb)
{
	struct tcp_sock *tp = tcp_sk(sk);

	tcp_skb_mark_lost_uncond_verify(tp, skb);
	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
		/* Account for retransmits that are lost again */
		TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
		tp->retrans_out -= tcp_skb_pcount(skb);
		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
	}
}

/* Marks a packet lost, if some packet sent later has been (s)acked.
 * The underlying idea is similar to the traditional dupthresh and FACK
 * but they look at different metrics:
 *
 * dupthresh: 3 OOO packets delivered (packet count)
 * FACK: sequence delta to highest sacked sequence (sequence space)
 * RACK: sent time delta to the latest delivered packet (time domain)
 *
 * The advantage of RACK is it applies to both original and retransmitted
 * packet and therefore is robust against tail losses. Another advantage
 * is being more resilient to reordering by simply allowing some
 * "settling delay", instead of tweaking the dupthresh.
 *
 * The current version is only used after recovery starts but can be
 * easily extended to detect the first loss.
 */
static void tcp_rack_detect_loss(struct sock *sk, const struct skb_mstamp *now)
{
	struct tcp_sock *tp = tcp_sk(sk);
	struct sk_buff *skb;
	u32 reo_wnd;

	/* To be more reordering resilient, allow min_rtt/4 settling delay
	 * (lower-bounded to 1000uS). We use min_rtt instead of the smoothed
	 * RTT because reordering is often a path property and less related
	 * to queuing or delayed ACKs.
	 *
	 * TODO: measure and adapt to the observed reordering delay, and
	 * use a timer to retransmit like the delayed early retransmit.
	 */
	reo_wnd = 1000;
	if (tp->rack.reord && tcp_min_rtt(tp) != ~0U)
		reo_wnd = max(tcp_min_rtt(tp) >> 2, reo_wnd);

	tcp_for_write_queue(skb, sk) {
		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);

		if (skb == tcp_send_head(sk))
			break;

		/* Skip ones already (s)acked */
		if (!after(scb->end_seq, tp->snd_una) ||
		    scb->sacked & TCPCB_SACKED_ACKED)
			continue;

		if (skb_mstamp_after(&tp->rack.mstamp, &skb->skb_mstamp)) {
			/* Step 3 in draft-cheng-tcpm-rack-00.txt:
			 * A packet is lost if its elapsed time is beyond
			 * the recent RTT plus the reordering window.
			 */
			if (skb_mstamp_us_delta(now, &skb->skb_mstamp) >
			    tp->rack.rtt_us + reo_wnd) {
				tcp_rack_mark_skb_lost(sk, skb);
			}
		} else if (!(scb->sacked & TCPCB_RETRANS)) {
			/* Original data are sent sequentially so stop early
			 * b/c the rest are all sent after rack_sent
			 */
			break;
		}
	}
}

void tcp_rack_mark_lost(struct sock *sk, const struct skb_mstamp *now)
{
	struct tcp_sock *tp = tcp_sk(sk);

	if (inet_csk(sk)->icsk_ca_state < TCP_CA_Recovery || !tp->rack.advanced)
		return;
	/* Reset the advanced flag to avoid unnecessary queue scanning */
	tp->rack.advanced = 0;
	tcp_rack_detect_loss(sk, now);
}

/* Record the most recently (re)sent time among the (s)acked packets
 * This is "Step 3: Advance RACK.xmit_time and update RACK.RTT" from
 * draft-cheng-tcpm-rack-00.txt
 */
void tcp_rack_advance(struct tcp_sock *tp, u8 sacked,
		      const struct skb_mstamp *xmit_time,
		      const struct skb_mstamp *ack_time)
{
	u32 rtt_us;

	if (tp->rack.mstamp.v64 &&
	    !skb_mstamp_after(xmit_time, &tp->rack.mstamp))
		return;

	rtt_us = skb_mstamp_us_delta(ack_time, xmit_time);
	if (sacked & TCPCB_RETRANS) {
		/* If the sacked packet was retransmitted, it's ambiguous
		 * whether the retransmission or the original (or the prior
		 * retransmission) was sacked.
		 *
		 * If the original is lost, there is no ambiguity. Otherwise
		 * we assume the original can be delayed up to aRTT + min_rtt.
		 * the aRTT term is bounded by the fast recovery or timeout,
		 * so it's at least one RTT (i.e., retransmission is at least
		 * an RTT later).
		 */
		if (rtt_us < tcp_min_rtt(tp))
			return;
	}
	tp->rack.rtt_us = rtt_us;
	tp->rack.mstamp = *xmit_time;
	tp->rack.advanced = 1;
}