1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2008-2010 Lawrence Stewart <lstewart@freebsd.org>
5 * Copyright (c) 2010 The FreeBSD Foundation
6 * All rights reserved.
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 #include <sys/limits.h>
45
46 /* Number of bits of precision for fixed point math calcs. */
47 #define CUBIC_SHIFT 8
48
49 #define CUBIC_SHIFT_4 32
50
51 /* 0.5 << CUBIC_SHIFT. */
52 #define RENO_BETA 128
53
54 /* ~0.7 << CUBIC_SHIFT. */
55 #define CUBIC_BETA 179
56
57 /* ~0.3 << CUBIC_SHIFT. */
58 #define ONE_SUB_CUBIC_BETA 77
59
60 /* 3 * ONE_SUB_CUBIC_BETA. */
61 #define THREE_X_PT3 231
62
63 /* (2 << CUBIC_SHIFT) - ONE_SUB_CUBIC_BETA. */
64 #define TWO_SUB_PT3 435
65
66 /* ~0.4 << CUBIC_SHIFT. */
67 #define CUBIC_C_FACTOR 102
68
69 /* CUBIC fast convergence factor: (1+beta_cubic)/2. */
70 #define CUBIC_FC_FACTOR 217
71
72 /* Don't trust s_rtt until this many rtt samples have been taken. */
73 #define CUBIC_MIN_RTT_SAMPLES 8
74
75 /*
76 * (2^21)^3 is long max. Dividing (2^63) by Cubic_C_factor
77 * and taking cube-root yields 448845 as the effective useful limit
78 */
79 #define CUBED_ROOT_MAX_ULONG 448845
80
81 /* Flags used in the cubic structure */
82 #define CUBICFLAG_CONG_EVENT 0x00000001 /* congestion experienced */
83 #define CUBICFLAG_IN_SLOWSTART 0x00000002 /* in slow start */
84 #define CUBICFLAG_IN_APPLIMIT 0x00000004 /* application limited */
85 #define CUBICFLAG_RTO_EVENT 0x00000008 /* RTO experienced */
86 #define CUBICFLAG_HYSTART_ENABLED 0x00000010 /* Hystart++ is enabled */
87 #define CUBICFLAG_HYSTART_IN_CSS 0x00000020 /* We are in Hystart++ CSS */
88
89 /* Kernel only bits */
90 #ifdef _KERNEL
91 struct cubic {
92 /* CUBIC K in fixed point form with CUBIC_SHIFT worth of precision. */
93 int64_t K;
94 /* Sum of RTT samples across an epoch in ticks. */
95 int64_t sum_rtt_ticks;
96 /* cwnd at the most recent congestion event. */
97 unsigned long max_cwnd;
98 /* cwnd at the previous congestion event. */
99 unsigned long prev_max_cwnd;
100 /* A copy of prev_max_cwnd. Used for CC_RTO_ERR */
101 unsigned long prev_max_cwnd_cp;
102 /* various flags */
103 uint32_t flags;
104 /* Minimum observed rtt in ticks. */
105 int min_rtt_ticks;
106 /* Mean observed rtt between congestion epochs. */
107 int mean_rtt_ticks;
108 /* ACKs since last congestion event. */
109 int epoch_ack_count;
110 /* Timestamp (in ticks) of arriving in congestion avoidance from last
111 * congestion event.
112 */
113 int t_last_cong;
114 /* Timestamp (in ticks) of a previous congestion event. Used for
115 * CC_RTO_ERR.
116 */
117 int t_last_cong_prev;
118 uint32_t css_baseline_minrtt;
119 uint32_t css_current_round_minrtt;
120 uint32_t css_lastround_minrtt;
121 uint32_t css_rttsample_count;
122 uint32_t css_entered_at_round;
123 uint32_t css_current_round;
124 uint32_t css_fas_at_css_entry;
125 uint32_t css_lowrtt_fas;
126 uint32_t css_last_fas;
127 };
128 #endif
129
130 /* Userland only bits. */
131 #ifndef _KERNEL
132
133 extern int hz;
134
135 /*
136 * Implementation based on the formulae found in the CUBIC Internet Draft
137 * "draft-ietf-tcpm-cubic-04".
138 *
139 */
140
141 static __inline float
142 theoretical_cubic_k(double wmax_pkts)
143 {
144 double C;
145
146 C = 0.4;
147
148 return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
149 }
150
151 static __inline unsigned long
152 theoretical_cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss)
153 {
154 double C, wmax_pkts;
155
156 C = 0.4;
157 wmax_pkts = wmax / (double)smss;
158
159 return (smss * (wmax_pkts +
160 (C * pow(ticks_since_cong / (double)hz -
161 theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
162 }
163
164 static __inline unsigned long
165 theoretical_reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
166 uint32_t smss)
167 {
168
169 return ((wmax * 0.5) + ((ticks_since_cong / (float)rtt_ticks) * smss));
170 }
171
172 static __inline unsigned long
173 theoretical_tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
174 uint32_t smss)
175 {
176
177 return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) *
178 (ticks_since_cong / (float)rtt_ticks) * smss));
179 }
180
181 #endif /* !_KERNEL */
182
183 /*
184 * Compute the CUBIC K value used in the cwnd calculation, using an
185 * implementation of eqn 2 in the I-D. The method used
186 * here is adapted from Apple Computer Technical Report #KT-32.
187 */
188 static __inline int64_t
189 cubic_k(unsigned long wmax_pkts)
190 {
191 int64_t s, K;
192 uint16_t p;
193
194 K = s = 0;
195 p = 0;
196
197 /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
198 s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
199
200 /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
201 while (s >= 256) {
202 s >>= 3;
203 p++;
204 }
205
206 /*
207 * Some magic constants taken from the Apple TR with appropriate
208 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
209 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
210 */
211 K = (((s * 275) >> CUBIC_SHIFT) + 98) -
212 (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
213
214 /* Multiply by 2^p to undo the rebasing of s from above. */
215 return (K <<= p);
216 }
217
218 /*
219 * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
220 * Thanks to Kip Macy for help debugging this function.
221 *
222 * XXXLAS: Characterise bounds for overflow.
223 */
224 static __inline unsigned long
225 cubic_cwnd(int ticks_since_cong, unsigned long wmax, uint32_t smss, int64_t K)
226 {
227 int64_t cwnd;
228
229 /* K is in fixed point form with CUBIC_SHIFT worth of precision. */
230
231 /* t - K, with CUBIC_SHIFT worth of precision. */
232 cwnd = (((int64_t)ticks_since_cong << CUBIC_SHIFT) - (K * hz)) / hz;
233
234 if (cwnd > CUBED_ROOT_MAX_ULONG)
235 return INT_MAX;
236 if (cwnd < -CUBED_ROOT_MAX_ULONG)
237 return 0;
238
239 /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
240 cwnd *= (cwnd * cwnd);
241
242 /*
243 * C(t - K)^3 + wmax
244 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
245 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
246 * and an extra from multiplying through by CUBIC_C_FACTOR.
247 */
248
249 cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax;
250
251 /*
252 * for negative cwnd, limiting to zero as lower bound
253 */
254 return (lmax(0,cwnd));
255 }
256
257 /*
258 * Compute an approximation of the NewReno cwnd some number of ticks after a
259 * congestion event. RTT should be the average RTT estimate for the path
260 * measured over the previous congestion epoch and wmax is the value of cwnd at
261 * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is
262 * rather tricky to understand and it turns out this function is not required.
263 * It is left here for reference.
264 */
265 static __inline unsigned long
266 reno_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
267 uint32_t smss)
268 {
269
270 /*
271 * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT
272 * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in
273 * bytes, we have to multiply by smss.
274 */
275 return (((wmax * RENO_BETA) + (((ticks_since_cong * smss)
276 << CUBIC_SHIFT) / rtt_ticks)) >> CUBIC_SHIFT);
277 }
278
279 /*
280 * Compute an approximation of the "TCP friendly" cwnd some number of ticks
281 * after a congestion event that is designed to yield the same average cwnd as
282 * NewReno while using CUBIC's beta of 0.7. RTT should be the average RTT
283 * estimate for the path measured over the previous congestion epoch and wmax is
284 * the value of cwnd at the last congestion event.
285 */
286 static __inline unsigned long
287 tf_cwnd(int ticks_since_cong, int rtt_ticks, unsigned long wmax,
288 uint32_t smss)
289 {
290
291 /* Equation 4 of I-D. */
292 return (((wmax * CUBIC_BETA) +
293 (((THREE_X_PT3 * (unsigned long)ticks_since_cong *
294 (unsigned long)smss) << CUBIC_SHIFT) / (TWO_SUB_PT3 * rtt_ticks)))
295 >> CUBIC_SHIFT);
296 }
297
298 #endif /* _NETINET_CC_CUBIC_H_ */
Cache object: f42688650f950bc78c5904a72ff7d110
|