1 /*
2 * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
3 * Copyright (c) 2010 The FreeBSD Foundation
4 * All rights reserved.
5 * Copyright (c) 2017 by Delphix. All rights reserved.
6 * Copyright 2019 Joyent, Inc.
7 *
8 * This software was developed by Lawrence Stewart while studying at the Centre
9 * for Advanced Internet Architectures, Swinburne University of Technology, made
10 * possible in part by a grant from the Cisco University Research Program Fund
11 * at Community Foundation Silicon Valley.
12 *
13 * Portions of this software were developed at the Centre for Advanced
14 * Internet Architectures, Swinburne University of Technology, Melbourne,
15 * Australia by David Hayes under sponsorship from the FreeBSD Foundation.
16 *
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
19 * are met:
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 *
26 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 *
38 * $FreeBSD$
39 */
40
41 #ifndef _NETINET_CC_CUBIC_H_
42 #define _NETINET_CC_CUBIC_H_
43
44 /* Number of bits of precision for fixed point math calcs. */
45 #define CUBIC_SHIFT 8
46
47 #define CUBIC_SHIFT_4 32
48
49 /* 0.5 << CUBIC_SHIFT. */
50 #define RENO_BETA 128
51
52 /* ~0.8 << CUBIC_SHIFT. */
53 #define CUBIC_BETA 204
54
55 /* ~0.2 << CUBIC_SHIFT. */
56 #define ONE_SUB_CUBIC_BETA 51
57
58 /* 3 * ONE_SUB_CUBIC_BETA. */
59 #define THREE_X_PT2 153
60
61 /* (2 << CUBIC_SHIFT) - ONE_SUB_CUBIC_BETA. */
62 #define TWO_SUB_PT2 461
63
64 /* ~0.4 << CUBIC_SHIFT. */
65 #define CUBIC_C_FACTOR 102
66
67 /* CUBIC fast convergence factor: ~0.9 << CUBIC_SHIFT. */
68 #define CUBIC_FC_FACTOR 230
69
70 /* Don't trust s_rtt until this many rtt samples have been taken. */
71 #define CUBIC_MIN_RTT_SAMPLES 8
72
73 /* Userland only bits. */
74 #ifndef _KERNEL
75
76 extern int hz;
77
78 /*
79 * Implementation based on the formulae found in the CUBIC Internet Draft
80 * "draft-rhee-tcpm-cubic-02".
81 *
82 * Note BETA used in cc_cubic is equal to (1-beta) in the I-D
83 */
84
85 static __inline float
86 theoretical_cubic_k(double wmax_pkts)
87 {
88 double C;
89
90 C = 0.4;
91
92 return (pow((wmax_pkts * 0.2) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
93 }
94
95 static __inline uint32_t
96 theoretical_cubic_cwnd(int ticks_since_cong, uint32_t wmax, uint32_t smss)
97 {
98 double C, wmax_pkts;
99
100 C = 0.4;
101 wmax_pkts = wmax / (double)smss;
102
103 return (smss * (wmax_pkts +
104 (C * pow(ticks_since_cong / (double)hz -
105 theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
106 }
107
108 static __inline uint32_t
109 theoretical_reno_cwnd(int ticks_since_cong, int rtt_ticks, uint32_t wmax,
110 uint32_t smss)
111 {
112
113 return ((wmax * 0.5) + ((ticks_since_cong / (float)rtt_ticks) * smss));
114 }
115
116 static __inline uint32_t
117 theoretical_tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
118 uint32_t smss)
119 {
120
121 return ((wmax * 0.8) + ((3 * 0.2) / (2 - 0.2) *
122 (ticks_since_cong / (float)rtt_ticks) * smss));
123 }
124
125 #endif /* !_KERNEL */
126
127 /*
128 * Compute the CUBIC K value used in the cwnd calculation, using an
129 * implementation of eqn 2 in the I-D. The method used
130 * here is adapted from Apple Computer Technical Report #KT-32.
131 */
132 static __inline int64_t
133 cubic_k(uint32_t wmax_pkts)
134 {
135 int64_t s, K;
136 uint16_t p;
137
138 K = s = 0;
139 p = 0;
140
141 /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
142 s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
143
144 /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
145 while (s >= 256) {
146 s >>= 3;
147 p++;
148 }
149
150 /*
151 * Some magic constants taken from the Apple TR with appropriate
152 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
153 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
154 */
155 K = (((s * 275) >> CUBIC_SHIFT) + 98) -
156 (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
157
158 /* Multiply by 2^p to undo the rebasing of s from above. */
159 return (K <<= p);
160 }
161
162 /*
163 * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
164 * Thanks to Kip Macy for help debugging this function.
165 *
166 * XXXLAS: Characterise bounds for overflow.
167 */
168 static __inline uint32_t
169 cubic_cwnd(hrtime_t nsecs_since_cong, uint32_t wmax, uint32_t smss, int64_t K)
170 {
171 int64_t t, cwnd;
172
173 /*
174 * Convert nsecs_since_cong to milliseconds, with CUBIC_SHIFT worth
175 * of precision.
176 */
177 t = NSEC2MSEC(nsecs_since_cong << CUBIC_SHIFT);
178
179 /*
180 * K is the time period in seconds that it will take to reach wmax. The
181 * value is kept in fixed point form with CUBIC_SHIFT worth of
182 * precision.
183 *
184 * For comparison with t, we convert K to milliseconds, and then convert
185 * the result back to seconds.
186 *
187 * cwnd = t - K, with CUBIC_SHIFT worth of precision.
188 */
189 cwnd = (t - K * MILLISEC) / MILLISEC;
190
191 /* cwnd = (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
192 cwnd *= (cwnd * cwnd);
193
194 /*
195 * C(t - K)^3 + wmax
196 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
197 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
198 * and an extra from multiplying through by CUBIC_C_FACTOR.
199 */
200 cwnd = ((cwnd * CUBIC_C_FACTOR * smss) >> CUBIC_SHIFT_4) + wmax;
201
202 return ((uint32_t)cwnd);
203 }
204
205 /*
206 * Compute an approximation of the "TCP friendly" cwnd some number of
207 * nanoseconds after a congestion event that is designed to yield the same
208 * average cwnd as NewReno while using CUBIC's beta of 0.8. RTT should be the
209 * average RTT estimate for the path measured over the previous congestion
210 * epoch and wmax is the value of cwnd at the last congestion event.
211 */
212 static __inline uint32_t
213 tf_cwnd(hrtime_t nsecs_since_cong, hrtime_t rtt_nsecs, uint32_t wmax,
214 uint32_t smss)
215 {
216
217 /* Equation 4 of I-D. */
218 return (((wmax * CUBIC_BETA) + (((THREE_X_PT2 * nsecs_since_cong *
219 smss) << CUBIC_SHIFT) / TWO_SUB_PT2 / rtt_nsecs)) >> CUBIC_SHIFT);
220 }
221
222 #endif /* _NETINET_CC_CUBIC_H_ */