xref: /freebsd/sys/netinet/cc/cc_cubic.h (revision 7f9ef5c75fd18fadcd2a74c2c489f218d7611208)
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 	/* The cwnd at the beginning of the current congestion avoidance stage */
98 	uint64_t	cwnd_epoch;
99 	/* various flags */
100 	uint32_t	flags;
101 	/* Minimum observed rtt in usecs. */
102 	int		min_rtt_usecs;
103 	/* Mean observed rtt between congestion epochs. */
104 	int		mean_rtt_usecs;
105 	/* ACKs since last congestion event. */
106 	int		epoch_ack_count;
107 	/* Timestamp (in ticks) at which the current CA epoch started. */
108 	int		t_epoch;
109 	/* Timestamp (in ticks) at which the previous CA epoch started. */
110 	int		undo_t_epoch;
111 	/* Few variables to restore the state after RTO_ERR */
112 	int64_t		undo_K;
113 	uint64_t	undo_W_max;
114 	uint64_t	undo_cwnd_epoch;
115 	uint32_t css_baseline_minrtt;
116 	uint32_t css_current_round_minrtt;
117 	uint32_t css_lastround_minrtt;
118 	uint32_t css_rttsample_count;
119 	uint32_t css_entered_at_round;
120 	uint32_t css_current_round;
121 	uint32_t css_fas_at_css_entry;
122 	uint32_t css_lowrtt_fas;
123 	uint32_t css_last_fas;
124 };
125 #endif
126 
127 /* Userland only bits. */
128 #ifndef _KERNEL
129 
130 extern int hz;
131 
132 /*
133  * Implementation based on the formulae found in the CUBIC Internet Draft
134  * "draft-ietf-tcpm-cubic-04".
135  *
136  */
137 
138 static __inline float
theoretical_cubic_k(double wmax_pkts)139 theoretical_cubic_k(double wmax_pkts)
140 {
141 	double C;
142 
143 	C = 0.4;
144 
145 	return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT));
146 }
147 
148 static __inline unsigned long
theoretical_cubic_cwnd(int ticks_since_epoch,unsigned long wmax,uint32_t smss)149 theoretical_cubic_cwnd(int ticks_since_epoch, unsigned long wmax, uint32_t smss)
150 {
151 	double C, wmax_pkts;
152 
153 	C = 0.4;
154 	wmax_pkts = wmax / (double)smss;
155 
156 	return (smss * (wmax_pkts +
157 	    (C * pow(ticks_since_epoch / (double)hz -
158 	    theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0))));
159 }
160 
161 static __inline unsigned long
theoretical_reno_cwnd(int ticks_since_epoch,int rtt_ticks,unsigned long wmax,uint32_t smss)162 theoretical_reno_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
163     uint32_t smss)
164 {
165 
166 	return ((wmax * 0.5) + ((ticks_since_epoch / (float)rtt_ticks) * smss));
167 }
168 
169 static __inline unsigned long
theoretical_tf_cwnd(int ticks_since_epoch,int rtt_ticks,unsigned long wmax,uint32_t smss)170 theoretical_tf_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax,
171     uint32_t smss)
172 {
173 
174 	return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) *
175 	    (ticks_since_epoch / (float)rtt_ticks) * smss));
176 }
177 
178 #endif /* !_KERNEL */
179 
180 /*
181  * Compute the CUBIC K value used in the cwnd calculation, using an
182  * implementation of eqn 2 in the I-D. The method used
183  * here is adapted from Apple Computer Technical Report #KT-32.
184  */
185 static __inline int64_t
cubic_k(unsigned long wmax_pkts)186 cubic_k(unsigned long wmax_pkts)
187 {
188 	int64_t s, K;
189 	uint16_t p;
190 
191 	K = s = 0;
192 	p = 0;
193 
194 	/* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */
195 	s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR;
196 
197 	/* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */
198 	while (s >= 256) {
199 		s >>= 3;
200 		p++;
201 	}
202 
203 	/*
204 	 * Some magic constants taken from the Apple TR with appropriate
205 	 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 <<
206 	 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT.
207 	 */
208 	K = (((s * 275) >> CUBIC_SHIFT) + 98) -
209 	    (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT);
210 
211 	/* Multiply by 2^p to undo the rebasing of s from above. */
212 	return (K <<= p);
213 }
214 
215 /*
216  * Compute the new cwnd value using an implementation of eqn 1 from the I-D.
217  * Thanks to Kip Macy for help debugging this function.
218  *
219  * XXXLAS: Characterise bounds for overflow.
220  */
221 static __inline unsigned long
cubic_cwnd(int usecs_since_epoch,unsigned long wmax,uint32_t smss,int64_t K)222 cubic_cwnd(int usecs_since_epoch, unsigned long wmax, uint32_t smss, int64_t K)
223 {
224 	int64_t cwnd;
225 
226 	/* K is in fixed point form with CUBIC_SHIFT worth of precision. */
227 
228 	/* t - K, with CUBIC_SHIFT worth of precision. */
229 	cwnd = (((int64_t)usecs_since_epoch << CUBIC_SHIFT) - (K * hz * tick)) /
230 	       (hz * tick);
231 
232 	if (cwnd > CUBED_ROOT_MAX_ULONG)
233 		return INT_MAX;
234 	if (cwnd < -CUBED_ROOT_MAX_ULONG)
235 		return 0;
236 
237 	/* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */
238 	cwnd *= (cwnd * cwnd);
239 
240 	/*
241 	 * C(t - K)^3 + wmax
242 	 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of
243 	 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above,
244 	 * and an extra from multiplying through by CUBIC_C_FACTOR.
245 	 */
246 
247 	cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax;
248 
249 	/*
250 	 * for negative cwnd, limiting to zero as lower bound
251 	 */
252 	return (lmax(0,cwnd));
253 }
254 
255 /*
256  * Compute an approximation of the NewReno cwnd some number of usecs after a
257  * congestion event. RTT should be the average RTT estimate for the path
258  * measured over the previous congestion epoch and wmax is the value of cwnd at
259  * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is
260  * rather tricky to understand and it turns out this function is not required.
261  * It is left here for reference.
262  *
263  * XXX: Not used
264  */
265 static __inline unsigned long
reno_cwnd(int usecs_since_epoch,int rtt_usecs,unsigned long wmax,uint32_t smss)266 reno_cwnd(int usecs_since_epoch, int rtt_usecs, 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) + (((usecs_since_epoch * smss)
276 	    << CUBIC_SHIFT) / rtt_usecs)) >> CUBIC_SHIFT);
277 }
278 
279 /*
280  * Compute the "TCP friendly" cwnd by newreno in congestion avoidance state.
281  */
282 static __inline unsigned long
tf_cwnd(struct cc_var * ccv)283 tf_cwnd(struct cc_var *ccv)
284 {
285 	/* newreno is "TCP friendly" */
286 	return newreno_cc_cwnd_in_cong_avoid(ccv);
287 }
288 
289 #endif /* _NETINET_CC_CUBIC_H_ */
290