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 * $FreeBSD$ 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 usecs. */ 95 int64_t sum_rtt_usecs; 96 /* Size of cwnd just before cwnd was reduced in the last congestion event */ 97 uint64_t W_max; 98 /* An estimate for the congestion window in the Reno-friendly region */ 99 uint64_t W_est; 100 /* The cwnd at the beginning of the current congestion avoidance stage */ 101 uint64_t cwnd_epoch; 102 /* 103 * Size of cwnd at the time of setting ssthresh most recently, 104 * either upon exiting the first slow start, or just before cwnd 105 * was reduced in the last congestion event 106 */ 107 uint64_t cwnd_prior; 108 /* various flags */ 109 uint32_t flags; 110 /* Minimum observed rtt in usecs. */ 111 int min_rtt_usecs; 112 /* Mean observed rtt between congestion epochs. */ 113 int mean_rtt_usecs; 114 /* ACKs since last congestion event. */ 115 int epoch_ack_count; 116 /* Timestamp (in ticks) at which the current CA epoch started. */ 117 int t_epoch; 118 /* Timestamp (in ticks) at which the previous CA epoch started. */ 119 int undo_t_epoch; 120 /* Few variables to restore the state after RTO_ERR */ 121 int64_t undo_K; 122 uint64_t undo_cwnd_prior; 123 uint64_t undo_W_max; 124 uint64_t undo_W_est; 125 uint64_t undo_cwnd_epoch; 126 /* Number of congestion events experienced */ 127 uint64_t num_cong_events; 128 uint32_t css_baseline_minrtt; 129 uint32_t css_current_round_minrtt; 130 uint32_t css_lastround_minrtt; 131 uint32_t css_rttsample_count; 132 uint32_t css_entered_at_round; 133 uint32_t css_current_round; 134 uint32_t css_fas_at_css_entry; 135 uint32_t css_lowrtt_fas; 136 uint32_t css_last_fas; 137 }; 138 #endif 139 140 /* Userland only bits. */ 141 #ifndef _KERNEL 142 143 extern int hz; 144 145 /* 146 * Implementation based on the formulae found in the CUBIC Internet Draft 147 * "draft-ietf-tcpm-cubic-04". 148 * 149 */ 150 151 static __inline float 152 theoretical_cubic_k(double wmax_pkts) 153 { 154 double C; 155 156 C = 0.4; 157 158 return (pow((wmax_pkts * 0.3) / C, (1.0 / 3.0)) * pow(2, CUBIC_SHIFT)); 159 } 160 161 static __inline unsigned long 162 theoretical_cubic_cwnd(int ticks_since_epoch, unsigned long wmax, uint32_t smss) 163 { 164 double C, wmax_pkts; 165 166 C = 0.4; 167 wmax_pkts = wmax / (double)smss; 168 169 return (smss * (wmax_pkts + 170 (C * pow(ticks_since_epoch / (double)hz - 171 theoretical_cubic_k(wmax_pkts) / pow(2, CUBIC_SHIFT), 3.0)))); 172 } 173 174 static __inline unsigned long 175 theoretical_reno_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax, 176 uint32_t smss) 177 { 178 179 return ((wmax * 0.5) + ((ticks_since_epoch / (float)rtt_ticks) * smss)); 180 } 181 182 static __inline unsigned long 183 theoretical_tf_cwnd(int ticks_since_epoch, int rtt_ticks, unsigned long wmax, 184 uint32_t smss) 185 { 186 187 return ((wmax * 0.7) + ((3 * 0.3) / (2 - 0.3) * 188 (ticks_since_epoch / (float)rtt_ticks) * smss)); 189 } 190 191 #endif /* !_KERNEL */ 192 193 /* 194 * Compute the CUBIC K value used in the cwnd calculation, using an 195 * implementation of eqn 2 in the I-D. The method used 196 * here is adapted from Apple Computer Technical Report #KT-32. 197 */ 198 static __inline int64_t 199 cubic_k(unsigned long wmax_pkts) 200 { 201 int64_t s, K; 202 uint16_t p; 203 204 K = s = 0; 205 p = 0; 206 207 /* (wmax * beta)/C with CUBIC_SHIFT worth of precision. */ 208 s = ((wmax_pkts * ONE_SUB_CUBIC_BETA) << CUBIC_SHIFT) / CUBIC_C_FACTOR; 209 210 /* Rebase s to be between 1 and 1/8 with a shift of CUBIC_SHIFT. */ 211 while (s >= 256) { 212 s >>= 3; 213 p++; 214 } 215 216 /* 217 * Some magic constants taken from the Apple TR with appropriate 218 * shifts: 275 == 1.072302 << CUBIC_SHIFT, 98 == 0.3812513 << 219 * CUBIC_SHIFT, 120 == 0.46946116 << CUBIC_SHIFT. 220 */ 221 K = (((s * 275) >> CUBIC_SHIFT) + 98) - 222 (((s * s * 120) >> CUBIC_SHIFT) >> CUBIC_SHIFT); 223 224 /* Multiply by 2^p to undo the rebasing of s from above. */ 225 return (K <<= p); 226 } 227 228 /* 229 * Compute the new cwnd value using an implementation of eqn 1 from the I-D. 230 * Thanks to Kip Macy for help debugging this function. 231 * 232 * XXXLAS: Characterise bounds for overflow. 233 */ 234 static __inline unsigned long 235 cubic_cwnd(int usecs_since_epoch, unsigned long wmax, uint32_t smss, int64_t K) 236 { 237 int64_t cwnd; 238 239 /* K is in fixed point form with CUBIC_SHIFT worth of precision. */ 240 241 /* t - K, with CUBIC_SHIFT worth of precision. */ 242 cwnd = (((int64_t)usecs_since_epoch << CUBIC_SHIFT) - (K * hz * tick)) / 243 (hz * tick); 244 245 if (cwnd > CUBED_ROOT_MAX_ULONG) 246 return INT_MAX; 247 if (cwnd < -CUBED_ROOT_MAX_ULONG) 248 return 0; 249 250 /* (t - K)^3, with CUBIC_SHIFT^3 worth of precision. */ 251 cwnd *= (cwnd * cwnd); 252 253 /* 254 * C(t - K)^3 + wmax 255 * The down shift by CUBIC_SHIFT_4 is because cwnd has 4 lots of 256 * CUBIC_SHIFT included in the value. 3 from the cubing of cwnd above, 257 * and an extra from multiplying through by CUBIC_C_FACTOR. 258 */ 259 260 cwnd = ((cwnd * CUBIC_C_FACTOR) >> CUBIC_SHIFT_4) * smss + wmax; 261 262 /* 263 * for negative cwnd, limiting to zero as lower bound 264 */ 265 return (lmax(0,cwnd)); 266 } 267 268 /* 269 * Compute an approximation of the NewReno cwnd some number of usecs after a 270 * congestion event. RTT should be the average RTT estimate for the path 271 * measured over the previous congestion epoch and wmax is the value of cwnd at 272 * the last congestion event. The "TCP friendly" concept in the CUBIC I-D is 273 * rather tricky to understand and it turns out this function is not required. 274 * It is left here for reference. 275 * 276 * XXX: Not used 277 */ 278 static __inline unsigned long 279 reno_cwnd(int usecs_since_epoch, int rtt_usecs, unsigned long wmax, 280 uint32_t smss) 281 { 282 283 /* 284 * For NewReno, beta = 0.5, therefore: W_tcp(t) = wmax*0.5 + t/RTT 285 * W_tcp(t) deals with cwnd/wmax in pkts, so because our cwnd is in 286 * bytes, we have to multiply by smss. 287 */ 288 return (((wmax * RENO_BETA) + (((usecs_since_epoch * smss) 289 << CUBIC_SHIFT) / rtt_usecs)) >> CUBIC_SHIFT); 290 } 291 292 /* 293 * Compute an approximation of the "TCP friendly" cwnd some number of usecs 294 * after a congestion event that is designed to yield the same average cwnd as 295 * NewReno while using CUBIC's beta of 0.7. RTT should be the average RTT 296 * estimate for the path measured over the previous congestion epoch and wmax is 297 * the value of cwnd at the last congestion event. 298 */ 299 static __inline unsigned long 300 tf_cwnd(int usecs_since_epoch, int rtt_usecs, unsigned long wmax, 301 uint32_t smss) 302 { 303 304 /* Equation 4 of I-D. */ 305 return (((wmax * CUBIC_BETA) + 306 (((THREE_X_PT3 * (unsigned long)usecs_since_epoch * 307 (unsigned long)smss) << CUBIC_SHIFT) / (TWO_SUB_PT3 * rtt_usecs))) 308 >> CUBIC_SHIFT); 309 } 310 311 #endif /* _NETINET_CC_CUBIC_H_ */ 312