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 #define CUBICFLAG_IN_TF 0x00000040 /* We are in TCP friendly region */
87
88 /* Kernel only bits */
89 #ifdef _KERNEL
90 struct cubic {
91 /* CUBIC K in fixed point form with CUBIC_SHIFT worth of precision. */
92 int64_t K;
93 /* Sum of RTT samples across an epoch in usecs. */
94 int64_t sum_rtt_usecs;
95 /* Size of cwnd just before cwnd was reduced in the last congestion event */
96 uint64_t W_max;
97 /* An estimate for the congestion window in the Reno-friendly region */
98 uint64_t W_est;
99 /* The cwnd at the beginning of the current congestion avoidance stage */
100 uint64_t cwnd_epoch;
101 /*
102 * Size of cwnd at the time of setting ssthresh most recently,
103 * either upon exiting the first slow start, or just before cwnd
104 * was reduced in the last congestion event
105 */
106 uint64_t cwnd_prior;
107 /* various flags */
108 uint32_t flags;
109 /* Minimum observed rtt in usecs. */
110 int min_rtt_usecs;
111 /* Mean observed rtt between congestion epochs. */
112 int mean_rtt_usecs;
113 /* ACKs since last congestion event. */
114 int epoch_ack_count;
115 /* Timestamp (in ticks) at which the current CA epoch started. */
116 int t_epoch;
117 /* Timestamp (in ticks) at which the previous CA epoch started. */
118 int undo_t_epoch;
119 /* Few variables to restore the state after RTO_ERR */
120 int64_t undo_K;
121 uint64_t undo_cwnd_prior;
122 uint64_t undo_W_max;
123 uint64_t undo_W_est;
124 uint64_t undo_cwnd_epoch;
125 uint32_t css_baseline_minrtt;
126 uint32_t css_current_round_minrtt;
127 uint32_t css_lastround_minrtt;
128 uint32_t css_rttsample_count;
129 uint32_t css_entered_at_round;
130 uint32_t css_current_round;
131 uint32_t css_fas_at_css_entry;
132 uint32_t css_lowrtt_fas;
133 uint32_t css_last_fas;
134 };
135 #endif
136
137 /* Userland only bits. */
138 #ifndef _KERNEL
139
140 extern int hz;
141
142 /*
143 * Implementation based on the formulae found in the CUBIC Internet Draft
144 * "draft-ietf-tcpm-cubic-04".
145 *
146 */
147
148 static __inline float
theoretical_cubic_k(double wmax_pkts)149 theoretical_cubic_k(double wmax_pkts)
150 {
151 double C;
152
153 C = 0.4;
154
155 return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
156 }
157
158 static __inline unsigned long
theoretical_cubic_cwnd(int ticks_since_epoch,unsigned long wmax,uint32_t smss)159 theoretical_cubic_cwnd(int ticks_since_epoch, unsigned long wmax, uint32_t smss)
160 {
161 double C, wmax_pkts;
162
163 C = 0.4;
164 wmax_pkts = wmax / (double)smss;
165
166 return (smss * (wmax_pkts +
167 (C * pow(ticks_since_epoch / (double)hz -
168 theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
169 }
170
171 static __inline unsigned long
theoretical_reno_cwnd(int ticks_since_epoch,int rtt_ticks,unsigned long wmax,uint32_t smss)172 theoretical_reno_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
173 uint32_t smss)
174 {
175
176 return ((wmax * 0.5) + ((ticks_since_epoch / (float)rtt_ticks) * smss));
177 }
178
179 static __inline unsigned long
theoretical_tf_cwnd(int ticks_since_epoch,int rtt_ticks,unsigned long wmax,uint32_t smss)180 theoretical_tf_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
181 uint32_t smss)
182 {
183
184 return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) *
185 (ticks_since_epoch / (float)rtt_ticks) * smss));
186 }
187
188 #endif /* !_KERNEL */
189
190 /*
191 * Compute the CUBIC K value used in the cwnd calculation, using an
192 * implementation of eqn 2 in the I-D. The method used
193 * here is adapted from Apple Computer Technical Report #KT-32.
194 */
195 static __inline int64_t
cubic_k(unsigned long wmax_pkts)196 cubic_k(unsigned long wmax_pkts)
197 {
198 int64_t s, K;
199 uint16_t p;
200
201 K = s = 0;
202 p = 0;
203
204 /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
205 s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
206
207 /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
208 while (s >= 256) {
209 s >>= 3;
210 p++;
211 }
212
213 /*
214 * Some magic constants taken from the Apple TR with appropriate
215 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
216 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
217 */
218 K = (((s * 275) >> CUBIC_SHIFT) + 98) -
219 (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
220
221 /* Multiply by 2^p to undo the rebasing of s from above. */
222 return (K <<= p);
223 }
224
225 /*
226 * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
227 * Thanks to Kip Macy for help debugging this function.
228 *
229 * XXXLAS: Characterise bounds for overflow.
230 */
231 static __inline unsigned long
cubic_cwnd(int usecs_since_epoch,unsigned long wmax,uint32_t smss,int64_t K)232 cubic_cwnd(int usecs_since_epoch, unsigned long wmax, uint32_t smss, int64_t K)
233 {
234 int64_t cwnd;
235
236 /* K is in fixed point form with CUBIC_SHIFT worth of precision. */
237
238 /* t - K, with CUBIC_SHIFT worth of precision. */
239 cwnd = (((int64_t)usecs_since_epoch << CUBIC_SHIFT) - (K * hz * tick)) /
240 (hz * tick);
241
242 if (cwnd > CUBED_ROOT_MAX_ULONG)
243 return INT_MAX;
244 if (cwnd < -CUBED_ROOT_MAX_ULONG)
245 return 0;
246
247 /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
248 cwnd *= (cwnd * cwnd);
249
250 /*
251 * C(t - K)^3 + wmax
252 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
253 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
254 * and an extra from multiplying through by CUBIC_C_FACTOR.
255 */
256
257 cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax;
258
259 /*
260 * for negative cwnd, limiting to zero as lower bound
261 */
262 return (lmax(0,cwnd));
263 }
264
265 /*
266 * Compute an approximation of the NewReno cwnd some number of usecs after a
267 * congestion event. RTT should be the average RTT estimate for the path
268 * measured over the previous congestion epoch and wmax is the value of cwnd at
269 * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is
270 * rather tricky to understand and it turns out this function is not required.
271 * It is left here for reference.
272 *
273 * XXX: Not used
274 */
275 static __inline unsigned long
reno_cwnd(int usecs_since_epoch,int rtt_usecs,unsigned long wmax,uint32_t smss)276 reno_cwnd(int usecs_since_epoch, int rtt_usecs, unsigned long wmax,
277 uint32_t smss)
278 {
279
280 /*
281 * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT
282 * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in
283 * bytes, we have to multiply by smss.
284 */
285 return (((wmax * RENO_BETA) + (((usecs_since_epoch * smss)
286 << CUBIC_SHIFT) / rtt_usecs)) >> CUBIC_SHIFT);
287 }
288
289 /*
290 * Compute the "TCP friendly" cwnd by newreno in congestion avoidance state.
291 */
292 static __inline unsigned long
tf_cwnd(struct cc_var * ccv)293 tf_cwnd(struct cc_var *ccv)
294 {
295 /* newreno is "TCP friendly" */
296 return newreno_cc_cwnd_in_cong_avoid(ccv);
297 }
298
299 #endif /* _NETINET_CC_CUBIC_H_ */
300