master
1/*-
2 * SPDX-License-Identifier: BSD-2-Clause
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
39#ifndef _NETINET_CC_CUBIC_H_
40#define _NETINET_CC_CUBIC_H_
41
42#include <sys/limits.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/* ~0.7 << CUBIC_SHIFT. */
53#define CUBIC_BETA 179
54
55/* ~0.3 << CUBIC_SHIFT. */
56#define ONE_SUB_CUBIC_BETA 77
57
58/* 3 * ONE_SUB_CUBIC_BETA. */
59#define THREE_X_PT3 231
60
61/* (2 << CUBIC_SHIFT) - ONE_SUB_CUBIC_BETA. */
62#define TWO_SUB_PT3 435
63
64/* ~0.4 << CUBIC_SHIFT. */
65#define CUBIC_C_FACTOR 102
66
67/* CUBIC fast convergence factor: (1+beta_cubic)/2. */
68#define CUBIC_FC_FACTOR 217
69
70/* Don't trust s_rtt until this many rtt samples have been taken. */
71#define CUBIC_MIN_RTT_SAMPLES 8
72
73/*
74 * (2^21)^3 is long max. Dividing (2^63) by Cubic_C_factor
75 * and taking cube-root yields 448845 as the effective useful limit
76 */
77#define CUBED_ROOT_MAX_ULONG 448845
78
79/* Flags used in the cubic structure */
80#define CUBICFLAG_CONG_EVENT 0x00000001 /* congestion experienced */
81#define CUBICFLAG_IN_SLOWSTART 0x00000002 /* in slow start */
82#define CUBICFLAG_IN_APPLIMIT 0x00000004 /* application limited */
83#define CUBICFLAG_RTO_EVENT 0x00000008 /* RTO experienced */
84#define CUBICFLAG_HYSTART_ENABLED 0x00000010 /* Hystart++ is enabled */
85#define CUBICFLAG_HYSTART_IN_CSS 0x00000020 /* We are in Hystart++ CSS */
86
87/* Kernel only bits */
88#ifdef _KERNEL
89struct cubic {
90 /* CUBIC K in fixed point form with CUBIC_SHIFT worth of precision. */
91 int64_t K;
92 /* Sum of RTT samples across an epoch in usecs. */
93 int64_t sum_rtt_usecs;
94 /* Size of cwnd just before cwnd was reduced in the last congestion event */
95 uint64_t W_max;
96 /* An estimate for the congestion window in the Reno-friendly region */
97 uint64_t W_est;
98 /* The cwnd at the beginning of the current congestion avoidance stage */
99 uint64_t cwnd_epoch;
100 /*
101 * Size of cwnd at the time of setting ssthresh most recently,
102 * either upon exiting the first slow start, or just before cwnd
103 * was reduced in the last congestion event
104 */
105 uint64_t cwnd_prior;
106 /* various flags */
107 uint32_t flags;
108 /* Minimum observed rtt in usecs. */
109 int min_rtt_usecs;
110 /* Mean observed rtt between congestion epochs. */
111 int mean_rtt_usecs;
112 /* ACKs since last congestion event. */
113 int epoch_ack_count;
114 /* Timestamp (in ticks) at which the current CA epoch started. */
115 int t_epoch;
116 /* Timestamp (in ticks) at which the previous CA epoch started. */
117 int undo_t_epoch;
118 /* Few variables to restore the state after RTO_ERR */
119 int64_t undo_K;
120 uint64_t undo_cwnd_prior;
121 uint64_t undo_W_max;
122 uint64_t undo_W_est;
123 uint64_t undo_cwnd_epoch;
124 /* Number of congestion events experienced */
125 uint64_t num_cong_events;
126 uint32_t css_baseline_minrtt;
127 uint32_t css_current_round_minrtt;
128 uint32_t css_lastround_minrtt;
129 uint32_t css_rttsample_count;
130 uint32_t css_entered_at_round;
131 uint32_t css_current_round;
132 uint32_t css_fas_at_css_entry;
133 uint32_t css_lowrtt_fas;
134 uint32_t css_last_fas;
135};
136#endif
137
138/* Userland only bits. */
139#ifndef _KERNEL
140
141extern int hz;
142
143/*
144 * Implementation based on the formulae found in the CUBIC Internet Draft
145 * "draft-ietf-tcpm-cubic-04".
146 *
147 */
148
149static __inline float
150theoretical_cubic_k(double wmax_pkts)
151{
152 double C;
153
154 C = 0.4;
155
156 return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
157}
158
159static __inline unsigned long
160theoretical_cubic_cwnd(int ticks_since_epoch, unsigned long wmax, uint32_t smss)
161{
162 double C, wmax_pkts;
163
164 C = 0.4;
165 wmax_pkts = wmax / (double)smss;
166
167 return (smss * (wmax_pkts +
168 (C * pow(ticks_since_epoch / (double)hz -
169 theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
170}
171
172static __inline unsigned long
173theoretical_reno_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
174 uint32_t smss)
175{
176
177 return ((wmax * 0.5) + ((ticks_since_epoch / (float)rtt_ticks) * smss));
178}
179
180static __inline unsigned long
181theoretical_tf_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
182 uint32_t smss)
183{
184
185 return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) *
186 (ticks_since_epoch / (float)rtt_ticks) * smss));
187}
188
189#endif /* !_KERNEL */
190
191/*
192 * Compute the CUBIC K value used in the cwnd calculation, using an
193 * implementation of eqn 2 in the I-D. The method used
194 * here is adapted from Apple Computer Technical Report #KT-32.
195 */
196static __inline int64_t
197cubic_k(unsigned long wmax_pkts)
198{
199 int64_t s, K;
200 uint16_t p;
201
202 K = s = 0;
203 p = 0;
204
205 /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
206 s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
207
208 /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
209 while (s >= 256) {
210 s >>= 3;
211 p++;
212 }
213
214 /*
215 * Some magic constants taken from the Apple TR with appropriate
216 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
217 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
218 */
219 K = (((s * 275) >> CUBIC_SHIFT) + 98) -
220 (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
221
222 /* Multiply by 2^p to undo the rebasing of s from above. */
223 return (K <<= p);
224}
225
226/*
227 * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
228 * Thanks to Kip Macy for help debugging this function.
229 *
230 * XXXLAS: Characterise bounds for overflow.
231 */
232static __inline unsigned long
233cubic_cwnd(int usecs_since_epoch, unsigned long wmax, uint32_t smss, int64_t K)
234{
235 int64_t cwnd;
236
237 /* K is in fixed point form with CUBIC_SHIFT worth of precision. */
238
239 /* t - K, with CUBIC_SHIFT worth of precision. */
240 cwnd = (((int64_t)usecs_since_epoch << CUBIC_SHIFT) - (K * hz * tick)) /
241 (hz * tick);
242
243 if (cwnd > CUBED_ROOT_MAX_ULONG)
244 return INT_MAX;
245 if (cwnd < -CUBED_ROOT_MAX_ULONG)
246 return 0;
247
248 /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
249 cwnd *= (cwnd * cwnd);
250
251 /*
252 * C(t - K)^3 + wmax
253 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
254 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
255 * and an extra from multiplying through by CUBIC_C_FACTOR.
256 */
257
258 cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax;
259
260 /*
261 * for negative cwnd, limiting to zero as lower bound
262 */
263 return (lmax(0,cwnd));
264}
265
266/*
267 * Compute an approximation of the NewReno cwnd some number of usecs after a
268 * congestion event. RTT should be the average RTT estimate for the path
269 * measured over the previous congestion epoch and wmax is the value of cwnd at
270 * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is
271 * rather tricky to understand and it turns out this function is not required.
272 * It is left here for reference.
273 *
274 * XXX: Not used
275 */
276static __inline unsigned long
277reno_cwnd(int usecs_since_epoch, int rtt_usecs, unsigned long wmax,
278 uint32_t smss)
279{
280
281 /*
282 * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT
283 * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in
284 * bytes, we have to multiply by smss.
285 */
286 return (((wmax * RENO_BETA) + (((usecs_since_epoch * smss)
287 << CUBIC_SHIFT) / rtt_usecs)) >> CUBIC_SHIFT);
288}
289
290/*
291 * Compute an approximation of the "TCP friendly" cwnd some number of usecs
292 * after a congestion event that is designed to yield the same average cwnd as
293 * NewReno while using CUBIC's beta of 0.7. RTT should be the average RTT
294 * estimate for the path measured over the previous congestion epoch and wmax is
295 * the value of cwnd at the last congestion event.
296 */
297static __inline unsigned long
298tf_cwnd(int usecs_since_epoch, int rtt_usecs, unsigned long wmax,
299 uint32_t smss)
300{
301
302 /* Equation 4 of I-D. */
303 return (((wmax * CUBIC_BETA) +
304 (((THREE_X_PT3 * (unsigned long)usecs_since_epoch *
305 (unsigned long)smss) << CUBIC_SHIFT) / (TWO_SUB_PT3 * rtt_usecs)))
306 >> CUBIC_SHIFT);
307}
308
309#endif /* _NETINET_CC_CUBIC_H_ */