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 39 /* 40 * An implementation of the CUBIC congestion control algorithm for FreeBSD, 41 * based on the Internet Draft "draft-rhee-tcpm-cubic-02" by Rhee, Xu and Ha. 42 * Originally released as part of the NewTCP research project at Swinburne 43 * University of Technology's Centre for Advanced Internet Architectures, 44 * Melbourne, Australia, which was made possible in part by a grant from the 45 * Cisco University Research Program Fund at Community Foundation Silicon 46 * Valley. More details are available at: 47 * http://caia.swin.edu.au/urp/newtcp/ 48 */ 49 50 #include <sys/cdefs.h> 51 __FBSDID("$FreeBSD$"); 52 53 #include <sys/param.h> 54 #include <sys/kernel.h> 55 #include <sys/malloc.h> 56 #include <sys/module.h> 57 #include <sys/socket.h> 58 #include <sys/socketvar.h> 59 #include <sys/sysctl.h> 60 #include <sys/systm.h> 61 62 #include <net/vnet.h> 63 64 #include <netinet/tcp.h> 65 #include <netinet/tcp_seq.h> 66 #include <netinet/tcp_timer.h> 67 #include <netinet/tcp_var.h> 68 #include <netinet/cc/cc.h> 69 #include <netinet/cc/cc_cubic.h> 70 #include <netinet/cc/cc_module.h> 71 72 static void cubic_ack_received(struct cc_var *ccv, uint16_t type); 73 static void cubic_cb_destroy(struct cc_var *ccv); 74 static int cubic_cb_init(struct cc_var *ccv); 75 static void cubic_cong_signal(struct cc_var *ccv, uint32_t type); 76 static void cubic_conn_init(struct cc_var *ccv); 77 static int cubic_mod_init(void); 78 static void cubic_post_recovery(struct cc_var *ccv); 79 static void cubic_record_rtt(struct cc_var *ccv); 80 static void cubic_ssthresh_update(struct cc_var *ccv); 81 82 struct cubic { 83 /* Cubic K in fixed point form with CUBIC_SHIFT worth of precision. */ 84 int64_t K; 85 /* Sum of RTT samples across an epoch in ticks. */ 86 int64_t sum_rtt_ticks; 87 /* cwnd at the most recent congestion event. */ 88 unsigned long max_cwnd; 89 /* cwnd at the previous congestion event. */ 90 unsigned long prev_max_cwnd; 91 /* Number of congestion events. */ 92 uint32_t num_cong_events; 93 /* Minimum observed rtt in ticks. */ 94 int min_rtt_ticks; 95 /* Mean observed rtt between congestion epochs. */ 96 int mean_rtt_ticks; 97 /* ACKs since last congestion event. */ 98 int epoch_ack_count; 99 /* Time of last congestion event in ticks. */ 100 int t_last_cong; 101 }; 102 103 static MALLOC_DEFINE(M_CUBIC, "cubic data", 104 "Per connection data required for the CUBIC congestion control algorithm"); 105 106 struct cc_algo cubic_cc_algo = { 107 .name = "cubic", 108 .ack_received = cubic_ack_received, 109 .cb_destroy = cubic_cb_destroy, 110 .cb_init = cubic_cb_init, 111 .cong_signal = cubic_cong_signal, 112 .conn_init = cubic_conn_init, 113 .mod_init = cubic_mod_init, 114 .post_recovery = cubic_post_recovery, 115 }; 116 117 static void 118 cubic_ack_received(struct cc_var *ccv, uint16_t type) 119 { 120 struct cubic *cubic_data; 121 unsigned long w_tf, w_cubic_next; 122 int ticks_since_cong; 123 124 cubic_data = ccv->cc_data; 125 cubic_record_rtt(ccv); 126 127 /* 128 * Regular ACK and we're not in cong/fast recovery and we're cwnd 129 * limited and we're either not doing ABC or are slow starting or are 130 * doing ABC and we've sent a cwnd's worth of bytes. 131 */ 132 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 133 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 || 134 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 135 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) { 136 /* Use the logic in NewReno ack_received() for slow start. */ 137 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 138 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE) 139 newreno_cc_algo.ack_received(ccv, type); 140 else { 141 ticks_since_cong = ticks - cubic_data->t_last_cong; 142 143 /* 144 * The mean RTT is used to best reflect the equations in 145 * the I-D. Using min_rtt in the tf_cwnd calculation 146 * causes w_tf to grow much faster than it should if the 147 * RTT is dominated by network buffering rather than 148 * propagation delay. 149 */ 150 w_tf = tf_cwnd(ticks_since_cong, 151 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd, 152 CCV(ccv, t_maxseg)); 153 154 w_cubic_next = cubic_cwnd(ticks_since_cong + 155 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd, 156 CCV(ccv, t_maxseg), cubic_data->K); 157 158 ccv->flags &= ~CCF_ABC_SENTAWND; 159 160 if (w_cubic_next < w_tf) 161 /* 162 * TCP-friendly region, follow tf 163 * cwnd growth. 164 */ 165 CCV(ccv, snd_cwnd) = w_tf; 166 167 else if (CCV(ccv, snd_cwnd) < w_cubic_next) { 168 /* 169 * Concave or convex region, follow CUBIC 170 * cwnd growth. 171 */ 172 if (V_tcp_do_rfc3465) 173 CCV(ccv, snd_cwnd) = w_cubic_next; 174 else 175 CCV(ccv, snd_cwnd) += ((w_cubic_next - 176 CCV(ccv, snd_cwnd)) * 177 CCV(ccv, t_maxseg)) / 178 CCV(ccv, snd_cwnd); 179 } 180 181 /* 182 * If we're not in slow start and we're probing for a 183 * new cwnd limit at the start of a connection 184 * (happens when hostcache has a relevant entry), 185 * keep updating our current estimate of the 186 * max_cwnd. 187 */ 188 if (cubic_data->num_cong_events == 0 && 189 cubic_data->max_cwnd < CCV(ccv, snd_cwnd)) 190 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 191 } 192 } 193 } 194 195 static void 196 cubic_cb_destroy(struct cc_var *ccv) 197 { 198 free(ccv->cc_data, M_CUBIC); 199 } 200 201 static int 202 cubic_cb_init(struct cc_var *ccv) 203 { 204 struct cubic *cubic_data; 205 206 cubic_data = malloc(sizeof(struct cubic), M_CUBIC, M_NOWAIT|M_ZERO); 207 208 if (cubic_data == NULL) 209 return (ENOMEM); 210 211 /* Init some key variables with sensible defaults. */ 212 cubic_data->t_last_cong = ticks; 213 cubic_data->min_rtt_ticks = TCPTV_SRTTBASE; 214 cubic_data->mean_rtt_ticks = 1; 215 216 ccv->cc_data = cubic_data; 217 218 return (0); 219 } 220 221 /* 222 * Perform any necessary tasks before we enter congestion recovery. 223 */ 224 static void 225 cubic_cong_signal(struct cc_var *ccv, uint32_t type) 226 { 227 struct cubic *cubic_data; 228 229 cubic_data = ccv->cc_data; 230 231 switch (type) { 232 case CC_NDUPACK: 233 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 234 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 235 cubic_ssthresh_update(ccv); 236 cubic_data->num_cong_events++; 237 cubic_data->prev_max_cwnd = cubic_data->max_cwnd; 238 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 239 } 240 ENTER_RECOVERY(CCV(ccv, t_flags)); 241 } 242 break; 243 244 case CC_ECN: 245 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 246 cubic_ssthresh_update(ccv); 247 cubic_data->num_cong_events++; 248 cubic_data->prev_max_cwnd = cubic_data->max_cwnd; 249 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 250 cubic_data->t_last_cong = ticks; 251 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 252 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 253 } 254 break; 255 256 case CC_RTO: 257 /* 258 * Grab the current time and record it so we know when the 259 * most recent congestion event was. Only record it when the 260 * timeout has fired more than once, as there is a reasonable 261 * chance the first one is a false alarm and may not indicate 262 * congestion. 263 */ 264 if (CCV(ccv, t_rxtshift) >= 2) { 265 cubic_data->num_cong_events++; 266 cubic_data->t_last_cong = ticks; 267 } 268 break; 269 } 270 } 271 272 static void 273 cubic_conn_init(struct cc_var *ccv) 274 { 275 struct cubic *cubic_data; 276 277 cubic_data = ccv->cc_data; 278 279 /* 280 * Ensure we have a sane initial value for max_cwnd recorded. Without 281 * this here bad things happen when entries from the TCP hostcache 282 * get used. 283 */ 284 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 285 } 286 287 static int 288 cubic_mod_init(void) 289 { 290 291 cubic_cc_algo.after_idle = newreno_cc_algo.after_idle; 292 293 return (0); 294 } 295 296 /* 297 * Perform any necessary tasks before we exit congestion recovery. 298 */ 299 static void 300 cubic_post_recovery(struct cc_var *ccv) 301 { 302 struct cubic *cubic_data; 303 int pipe; 304 305 cubic_data = ccv->cc_data; 306 pipe = 0; 307 308 /* Fast convergence heuristic. */ 309 if (cubic_data->max_cwnd < cubic_data->prev_max_cwnd) 310 cubic_data->max_cwnd = (cubic_data->max_cwnd * CUBIC_FC_FACTOR) 311 >> CUBIC_SHIFT; 312 313 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 314 /* 315 * If inflight data is less than ssthresh, set cwnd 316 * conservatively to avoid a burst of data, as suggested in 317 * the NewReno RFC. Otherwise, use the CUBIC method. 318 * 319 * XXXLAS: Find a way to do this without needing curack 320 */ 321 if (V_tcp_do_rfc6675_pipe) 322 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 323 else 324 pipe = CCV(ccv, snd_max) - ccv->curack; 325 326 if (pipe < CCV(ccv, snd_ssthresh)) 327 /* 328 * Ensure that cwnd does not collapse to 1 MSS under 329 * adverse conditions. Implements RFC6582 330 */ 331 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) + 332 CCV(ccv, t_maxseg); 333 else 334 /* Update cwnd based on beta and adjusted max_cwnd. */ 335 CCV(ccv, snd_cwnd) = max(1, ((CUBIC_BETA * 336 cubic_data->max_cwnd) >> CUBIC_SHIFT)); 337 } 338 cubic_data->t_last_cong = ticks; 339 340 /* Calculate the average RTT between congestion epochs. */ 341 if (cubic_data->epoch_ack_count > 0 && 342 cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) { 343 cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks / 344 cubic_data->epoch_ack_count); 345 } 346 347 cubic_data->epoch_ack_count = 0; 348 cubic_data->sum_rtt_ticks = 0; 349 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg)); 350 } 351 352 /* 353 * Record the min RTT and sum samples for the epoch average RTT calculation. 354 */ 355 static void 356 cubic_record_rtt(struct cc_var *ccv) 357 { 358 struct cubic *cubic_data; 359 int t_srtt_ticks; 360 361 /* Ignore srtt until a min number of samples have been taken. */ 362 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) { 363 cubic_data = ccv->cc_data; 364 t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE; 365 366 /* 367 * Record the current SRTT as our minrtt if it's the smallest 368 * we've seen or minrtt is currently equal to its initialised 369 * value. 370 * 371 * XXXLAS: Should there be some hysteresis for minrtt? 372 */ 373 if ((t_srtt_ticks < cubic_data->min_rtt_ticks || 374 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) { 375 cubic_data->min_rtt_ticks = max(1, t_srtt_ticks); 376 377 /* 378 * If the connection is within its first congestion 379 * epoch, ensure we prime mean_rtt_ticks with a 380 * reasonable value until the epoch average RTT is 381 * calculated in cubic_post_recovery(). 382 */ 383 if (cubic_data->min_rtt_ticks > 384 cubic_data->mean_rtt_ticks) 385 cubic_data->mean_rtt_ticks = 386 cubic_data->min_rtt_ticks; 387 } 388 389 /* Sum samples for epoch average RTT calculation. */ 390 cubic_data->sum_rtt_ticks += t_srtt_ticks; 391 cubic_data->epoch_ack_count++; 392 } 393 } 394 395 /* 396 * Update the ssthresh in the event of congestion. 397 */ 398 static void 399 cubic_ssthresh_update(struct cc_var *ccv) 400 { 401 struct cubic *cubic_data; 402 403 cubic_data = ccv->cc_data; 404 405 /* 406 * On the first congestion event, set ssthresh to cwnd * 0.5, on 407 * subsequent congestion events, set it to cwnd * beta. 408 */ 409 if (cubic_data->num_cong_events == 0) 410 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd) >> 1; 411 else 412 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 413 CUBIC_BETA) >> CUBIC_SHIFT; 414 } 415 416 417 DECLARE_CC_MODULE(cubic, &cubic_cc_algo); 418