xref: /linux/net/rxrpc/rtt.c (revision e814f3fd16acfb7f9966773953de8f740a1e3202)
1 // SPDX-License-Identifier: GPL-2.0
2 /* RTT/RTO calculation.
3  *
4  * Adapted from TCP for AF_RXRPC by David Howells (dhowells@redhat.com)
5  *
6  * https://tools.ietf.org/html/rfc6298
7  * https://tools.ietf.org/html/rfc1122#section-4.2.3.1
8  * http://ccr.sigcomm.org/archive/1995/jan95/ccr-9501-partridge87.pdf
9  */
10 
11 #include <linux/net.h>
12 #include "ar-internal.h"
13 
14 #define RXRPC_RTO_MAX	(120 * USEC_PER_SEC)
15 #define RXRPC_TIMEOUT_INIT ((unsigned int)(1 * USEC_PER_SEC)) /* RFC6298 2.1 initial RTO value */
16 #define rxrpc_jiffies32 ((u32)jiffies)		/* As rxrpc_jiffies32 */
17 
18 static u32 rxrpc_rto_min_us(struct rxrpc_call *call)
19 {
20 	return 200;
21 }
22 
23 static u32 __rxrpc_set_rto(const struct rxrpc_call *call)
24 {
25 	return (call->srtt_us >> 3) + call->rttvar_us;
26 }
27 
28 static u32 rxrpc_bound_rto(u32 rto)
29 {
30 	return clamp(200000, rto + 100000, RXRPC_RTO_MAX);
31 }
32 
33 /*
34  * Called to compute a smoothed rtt estimate. The data fed to this
35  * routine either comes from timestamps, or from segments that were
36  * known _not_ to have been retransmitted [see Karn/Partridge
37  * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
38  * piece by Van Jacobson.
39  * NOTE: the next three routines used to be one big routine.
40  * To save cycles in the RFC 1323 implementation it was better to break
41  * it up into three procedures. -- erics
42  */
43 static void rxrpc_rtt_estimator(struct rxrpc_call *call, long sample_rtt_us)
44 {
45 	long m = sample_rtt_us; /* RTT */
46 	u32 srtt = call->srtt_us;
47 
48 	/*	The following amusing code comes from Jacobson's
49 	 *	article in SIGCOMM '88.  Note that rtt and mdev
50 	 *	are scaled versions of rtt and mean deviation.
51 	 *	This is designed to be as fast as possible
52 	 *	m stands for "measurement".
53 	 *
54 	 *	On a 1990 paper the rto value is changed to:
55 	 *	RTO = rtt + 4 * mdev
56 	 *
57 	 * Funny. This algorithm seems to be very broken.
58 	 * These formulae increase RTO, when it should be decreased, increase
59 	 * too slowly, when it should be increased quickly, decrease too quickly
60 	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
61 	 * does not matter how to _calculate_ it. Seems, it was trap
62 	 * that VJ failed to avoid. 8)
63 	 */
64 	if (srtt != 0) {
65 		m -= (srtt >> 3);	/* m is now error in rtt est */
66 		srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
67 		if (m < 0) {
68 			m = -m;		/* m is now abs(error) */
69 			m -= (call->mdev_us >> 2);   /* similar update on mdev */
70 			/* This is similar to one of Eifel findings.
71 			 * Eifel blocks mdev updates when rtt decreases.
72 			 * This solution is a bit different: we use finer gain
73 			 * for mdev in this case (alpha*beta).
74 			 * Like Eifel it also prevents growth of rto,
75 			 * but also it limits too fast rto decreases,
76 			 * happening in pure Eifel.
77 			 */
78 			if (m > 0)
79 				m >>= 3;
80 		} else {
81 			m -= (call->mdev_us >> 2);   /* similar update on mdev */
82 		}
83 
84 		call->mdev_us += m;		/* mdev = 3/4 mdev + 1/4 new */
85 		if (call->mdev_us > call->mdev_max_us) {
86 			call->mdev_max_us = call->mdev_us;
87 			if (call->mdev_max_us > call->rttvar_us)
88 				call->rttvar_us = call->mdev_max_us;
89 		}
90 	} else {
91 		/* no previous measure. */
92 		srtt = m << 3;		/* take the measured time to be rtt */
93 		call->mdev_us = m << 1;	/* make sure rto = 3*rtt */
94 		call->rttvar_us = umax(call->mdev_us, rxrpc_rto_min_us(call));
95 		call->mdev_max_us = call->rttvar_us;
96 	}
97 
98 	call->srtt_us = umax(srtt, 1);
99 }
100 
101 /*
102  * Calculate rto without backoff.  This is the second half of Van Jacobson's
103  * routine referred to above.
104  */
105 static void rxrpc_set_rto(struct rxrpc_call *call)
106 {
107 	u32 rto;
108 
109 	/* 1. If rtt variance happened to be less 50msec, it is hallucination.
110 	 *    It cannot be less due to utterly erratic ACK generation made
111 	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
112 	 *    to do with delayed acks, because at cwnd>2 true delack timeout
113 	 *    is invisible. Actually, Linux-2.4 also generates erratic
114 	 *    ACKs in some circumstances.
115 	 */
116 	rto = __rxrpc_set_rto(call);
117 
118 	/* 2. Fixups made earlier cannot be right.
119 	 *    If we do not estimate RTO correctly without them,
120 	 *    all the algo is pure shit and should be replaced
121 	 *    with correct one. It is exactly, which we pretend to do.
122 	 */
123 
124 	/* NOTE: clamping at RXRPC_RTO_MIN is not required, current algo
125 	 * guarantees that rto is higher.
126 	 */
127 	call->rto_us = rxrpc_bound_rto(rto);
128 }
129 
130 static void rxrpc_update_rtt_min(struct rxrpc_call *call, ktime_t resp_time, long rtt_us)
131 {
132 	/* Window size 5mins in approx usec (ipv4.sysctl_tcp_min_rtt_wlen) */
133 	u32 wlen_us = 5ULL * NSEC_PER_SEC / 1024;
134 
135 	minmax_running_min(&call->min_rtt, wlen_us, resp_time / 1024,
136 			   (u32)rtt_us ? : jiffies_to_usecs(1));
137 }
138 
139 static void rxrpc_ack_update_rtt(struct rxrpc_call *call, ktime_t resp_time, long rtt_us)
140 {
141 	if (rtt_us < 0)
142 		return;
143 
144 	/* Update RACK min RTT [RFC8985 6.1 Step 1]. */
145 	rxrpc_update_rtt_min(call, resp_time, rtt_us);
146 
147 	rxrpc_rtt_estimator(call, rtt_us);
148 	rxrpc_set_rto(call);
149 
150 	/* Only reset backoff on valid RTT measurement [RFC6298]. */
151 	call->backoff = 0;
152 }
153 
154 /*
155  * Add RTT information to cache.  This is called in softirq mode and has
156  * exclusive access to the call RTT data.
157  */
158 void rxrpc_call_add_rtt(struct rxrpc_call *call, enum rxrpc_rtt_rx_trace why,
159 			int rtt_slot,
160 			rxrpc_serial_t send_serial, rxrpc_serial_t resp_serial,
161 			ktime_t send_time, ktime_t resp_time)
162 {
163 	s64 rtt_us;
164 
165 	rtt_us = ktime_to_us(ktime_sub(resp_time, send_time));
166 	if (rtt_us < 0)
167 		return;
168 
169 	rxrpc_ack_update_rtt(call, resp_time, rtt_us);
170 	if (call->rtt_count < 3)
171 		call->rtt_count++;
172 	call->rtt_taken++;
173 
174 	WRITE_ONCE(call->peer->recent_srtt_us, call->srtt_us / 8);
175 	WRITE_ONCE(call->peer->recent_rto_us, call->rto_us);
176 
177 	trace_rxrpc_rtt_rx(call, why, rtt_slot, send_serial, resp_serial,
178 			   rtt_us, call->srtt_us, call->rto_us);
179 }
180 
181 /*
182  * Get the retransmission timeout to set in nanoseconds, backing it off each
183  * time we retransmit.
184  */
185 ktime_t rxrpc_get_rto_backoff(struct rxrpc_call *call, bool retrans)
186 {
187 	u64 timo_us;
188 	u32 backoff = READ_ONCE(call->backoff);
189 
190 	timo_us = call->rto_us;
191 	timo_us <<= backoff;
192 	if (retrans && timo_us * 2 <= RXRPC_RTO_MAX)
193 		WRITE_ONCE(call->backoff, backoff + 1);
194 
195 	if (timo_us < 1)
196 		timo_us = 1;
197 
198 	return ns_to_ktime(timo_us * NSEC_PER_USEC);
199 }
200 
201 void rxrpc_call_init_rtt(struct rxrpc_call *call)
202 {
203 	call->rtt_last_req = KTIME_MIN;
204 	call->rto_us	= RXRPC_TIMEOUT_INIT;
205 	call->mdev_us	= RXRPC_TIMEOUT_INIT;
206 	call->backoff	= 0;
207 	//minmax_reset(&call->rtt_min, rxrpc_jiffies32, ~0U);
208 }
209