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/limits.h> 56 #include <sys/malloc.h> 57 #include <sys/module.h> 58 #include <sys/socket.h> 59 #include <sys/socketvar.h> 60 #include <sys/sysctl.h> 61 #include <sys/systm.h> 62 63 #include <net/vnet.h> 64 65 #include <net/route.h> 66 #include <net/route/nhop.h> 67 68 #include <netinet/in_pcb.h> 69 #include <netinet/tcp.h> 70 #include <netinet/tcp_seq.h> 71 #include <netinet/tcp_timer.h> 72 #include <netinet/tcp_var.h> 73 #include <netinet/tcp_log_buf.h> 74 #include <netinet/tcp_hpts.h> 75 #include <netinet/cc/cc.h> 76 #include <netinet/cc/cc_cubic.h> 77 #include <netinet/cc/cc_module.h> 78 79 static void cubic_ack_received(struct cc_var *ccv, uint16_t type); 80 static void cubic_cb_destroy(struct cc_var *ccv); 81 static int cubic_cb_init(struct cc_var *ccv, void *ptr); 82 static void cubic_cong_signal(struct cc_var *ccv, uint32_t type); 83 static void cubic_conn_init(struct cc_var *ccv); 84 static int cubic_mod_init(void); 85 static void cubic_post_recovery(struct cc_var *ccv); 86 static void cubic_record_rtt(struct cc_var *ccv); 87 static void cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg); 88 static void cubic_after_idle(struct cc_var *ccv); 89 static size_t cubic_data_sz(void); 90 static void cubic_newround(struct cc_var *ccv, uint32_t round_cnt); 91 static void cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, 92 uint32_t rxtcnt, uint32_t fas); 93 94 struct cc_algo cubic_cc_algo = { 95 .name = "cubic", 96 .ack_received = cubic_ack_received, 97 .cb_destroy = cubic_cb_destroy, 98 .cb_init = cubic_cb_init, 99 .cong_signal = cubic_cong_signal, 100 .conn_init = cubic_conn_init, 101 .mod_init = cubic_mod_init, 102 .post_recovery = cubic_post_recovery, 103 .after_idle = cubic_after_idle, 104 .cc_data_sz = cubic_data_sz, 105 .rttsample = cubic_rttsample, 106 .newround = cubic_newround 107 }; 108 109 static void 110 cubic_log_hystart_event(struct cc_var *ccv, struct cubic *cubicd, uint8_t mod, uint32_t flex1) 111 { 112 /* 113 * Types of logs (mod value) 114 * 1 - rtt_thresh in flex1, checking to see if RTT is to great. 115 * 2 - rtt is too great, rtt_thresh in flex1. 116 * 3 - CSS is active incr in flex1 117 * 4 - A new round is beginning flex1 is round count 118 * 5 - A new RTT measurement flex1 is the new measurement. 119 * 6 - We enter CA ssthresh is also in flex1. 120 * 7 - Socket option to change hystart executed opt.val in flex1. 121 * 8 - Back out of CSS into SS, flex1 is the css_baseline_minrtt 122 * 9 - We enter CA, via an ECN mark. 123 * 10 - We enter CA, via a loss. 124 * 11 - We have slipped out of SS into CA via cwnd growth. 125 * 12 - After idle has re-enabled hystart++ 126 */ 127 struct tcpcb *tp; 128 129 if (hystart_bblogs == 0) 130 return; 131 tp = ccv->ccvc.tcp; 132 if (tcp_bblogging_on(tp)) { 133 union tcp_log_stackspecific log; 134 struct timeval tv; 135 136 memset(&log, 0, sizeof(log)); 137 log.u_bbr.flex1 = flex1; 138 log.u_bbr.flex2 = cubicd->css_current_round_minrtt; 139 log.u_bbr.flex3 = cubicd->css_lastround_minrtt; 140 log.u_bbr.flex4 = cubicd->css_rttsample_count; 141 log.u_bbr.flex5 = cubicd->css_entered_at_round; 142 log.u_bbr.flex6 = cubicd->css_baseline_minrtt; 143 /* We only need bottom 16 bits of flags */ 144 log.u_bbr.flex7 = cubicd->flags & 0x0000ffff; 145 log.u_bbr.flex8 = mod; 146 log.u_bbr.epoch = cubicd->css_current_round; 147 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 148 log.u_bbr.lt_epoch = cubicd->css_fas_at_css_entry; 149 log.u_bbr.pkts_out = cubicd->css_last_fas; 150 log.u_bbr.delivered = cubicd->css_lowrtt_fas; 151 log.u_bbr.pkt_epoch = ccv->flags; 152 TCP_LOG_EVENTP(tp, NULL, 153 &tptosocket(tp)->so_rcv, 154 &tptosocket(tp)->so_snd, 155 TCP_HYSTART, 0, 156 0, &log, false, &tv); 157 } 158 } 159 160 static void 161 cubic_does_slow_start(struct cc_var *ccv, struct cubic *cubicd) 162 { 163 /* 164 * In slow-start with ABC enabled and no RTO in sight? 165 * (Must not use abc_l_var > 1 if slow starting after 166 * an RTO. On RTO, snd_nxt = snd_una, so the 167 * snd_nxt == snd_max check is sufficient to 168 * handle this). 169 * 170 * XXXLAS: Find a way to signal SS after RTO that 171 * doesn't rely on tcpcb vars. 172 */ 173 u_int cw = CCV(ccv, snd_cwnd); 174 u_int incr = CCV(ccv, t_maxseg); 175 uint16_t abc_val; 176 177 cubicd->flags |= CUBICFLAG_IN_SLOWSTART; 178 if (ccv->flags & CCF_USE_LOCAL_ABC) 179 abc_val = ccv->labc; 180 else 181 abc_val = V_tcp_abc_l_var; 182 if ((ccv->flags & CCF_HYSTART_ALLOWED) && 183 (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) && 184 ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) == 0)) { 185 /* 186 * Hystart is allowed and still enabled and we are not yet 187 * in CSS. Lets check to see if we can make a decision on 188 * if we need to go into CSS. 189 */ 190 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 191 (cubicd->css_current_round_minrtt != 0xffffffff) && 192 (cubicd->css_lastround_minrtt != 0xffffffff)) { 193 uint32_t rtt_thresh; 194 195 /* Clamp (minrtt_thresh, lastround/8, maxrtt_thresh) */ 196 rtt_thresh = (cubicd->css_lastround_minrtt >> 3); 197 if (rtt_thresh < hystart_minrtt_thresh) 198 rtt_thresh = hystart_minrtt_thresh; 199 if (rtt_thresh > hystart_maxrtt_thresh) 200 rtt_thresh = hystart_maxrtt_thresh; 201 cubic_log_hystart_event(ccv, cubicd, 1, rtt_thresh); 202 203 if (cubicd->css_current_round_minrtt >= (cubicd->css_lastround_minrtt + rtt_thresh)) { 204 /* Enter CSS */ 205 cubicd->flags |= CUBICFLAG_HYSTART_IN_CSS; 206 cubicd->css_fas_at_css_entry = cubicd->css_lowrtt_fas; 207 /* 208 * The draft (v4) calls for us to set baseline to css_current_round_min 209 * but that can cause an oscillation. We probably shoudl be using 210 * css_lastround_minrtt, but the authors insist that will cause 211 * issues on exiting early. We will leave the draft version for now 212 * but I suspect this is incorrect. 213 */ 214 cubicd->css_baseline_minrtt = cubicd->css_current_round_minrtt; 215 cubicd->css_entered_at_round = cubicd->css_current_round; 216 cubic_log_hystart_event(ccv, cubicd, 2, rtt_thresh); 217 } 218 } 219 } 220 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) 221 incr = min(ccv->bytes_this_ack, 222 ccv->nsegs * abc_val * 223 CCV(ccv, t_maxseg)); 224 else 225 incr = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg)); 226 227 /* Only if Hystart is enabled will the flag get set */ 228 if (cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) { 229 incr /= hystart_css_growth_div; 230 cubic_log_hystart_event(ccv, cubicd, 3, incr); 231 } 232 /* ABC is on by default, so incr equals 0 frequently. */ 233 if (incr > 0) 234 CCV(ccv, snd_cwnd) = min((cw + incr), 235 TCP_MAXWIN << CCV(ccv, snd_scale)); 236 } 237 238 static void 239 cubic_ack_received(struct cc_var *ccv, uint16_t type) 240 { 241 struct cubic *cubic_data; 242 unsigned long w_tf, w_cubic_next; 243 int ticks_since_cong; 244 245 cubic_data = ccv->cc_data; 246 cubic_record_rtt(ccv); 247 248 /* 249 * For a regular ACK and we're not in cong/fast recovery and 250 * we're cwnd limited, always recalculate cwnd. 251 */ 252 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 253 (ccv->flags & CCF_CWND_LIMITED)) { 254 /* Use the logic in NewReno ack_received() for slow start. */ 255 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 256 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE) { 257 cubic_does_slow_start(ccv, cubic_data); 258 } else { 259 if (cubic_data->flags & CUBICFLAG_HYSTART_IN_CSS) { 260 /* 261 * We have slipped into CA with 262 * CSS active. Deactivate all. 263 */ 264 /* Turn off the CSS flag */ 265 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 266 /* Disable use of CSS in the future except long idle */ 267 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 268 cubic_log_hystart_event(ccv, cubic_data, 11, CCV(ccv, snd_ssthresh)); 269 } 270 if ((cubic_data->flags & CUBICFLAG_RTO_EVENT) && 271 (cubic_data->flags & CUBICFLAG_IN_SLOWSTART)) { 272 /* RFC8312 Section 4.7 */ 273 cubic_data->flags &= ~(CUBICFLAG_RTO_EVENT | 274 CUBICFLAG_IN_SLOWSTART); 275 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 276 cubic_data->K = 0; 277 } else if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART | 278 CUBICFLAG_IN_APPLIMIT)) { 279 cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART | 280 CUBICFLAG_IN_APPLIMIT); 281 cubic_data->t_last_cong = ticks; 282 cubic_data->K = cubic_k(cubic_data->max_cwnd / 283 CCV(ccv, t_maxseg)); 284 } 285 if ((ticks_since_cong = 286 ticks - cubic_data->t_last_cong) < 0) { 287 /* 288 * dragging t_last_cong along 289 */ 290 ticks_since_cong = INT_MAX; 291 cubic_data->t_last_cong = ticks - INT_MAX; 292 } 293 /* 294 * The mean RTT is used to best reflect the equations in 295 * the I-D. Using min_rtt in the tf_cwnd calculation 296 * causes w_tf to grow much faster than it should if the 297 * RTT is dominated by network buffering rather than 298 * propagation delay. 299 */ 300 w_tf = tf_cwnd(ticks_since_cong, 301 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd, 302 CCV(ccv, t_maxseg)); 303 304 w_cubic_next = cubic_cwnd(ticks_since_cong + 305 cubic_data->mean_rtt_ticks, cubic_data->max_cwnd, 306 CCV(ccv, t_maxseg), cubic_data->K); 307 308 ccv->flags &= ~CCF_ABC_SENTAWND; 309 310 if (w_cubic_next < w_tf) { 311 /* 312 * TCP-friendly region, follow tf 313 * cwnd growth. 314 */ 315 if (CCV(ccv, snd_cwnd) < w_tf) 316 CCV(ccv, snd_cwnd) = ulmin(w_tf, INT_MAX); 317 } else if (CCV(ccv, snd_cwnd) < w_cubic_next) { 318 /* 319 * Concave or convex region, follow CUBIC 320 * cwnd growth. 321 * Only update snd_cwnd, if it doesn't shrink. 322 */ 323 CCV(ccv, snd_cwnd) = ulmin(w_cubic_next, 324 INT_MAX); 325 } 326 327 /* 328 * If we're not in slow start and we're probing for a 329 * new cwnd limit at the start of a connection 330 * (happens when hostcache has a relevant entry), 331 * keep updating our current estimate of the 332 * max_cwnd. 333 */ 334 if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) && 335 cubic_data->max_cwnd < CCV(ccv, snd_cwnd)) { 336 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 337 cubic_data->K = cubic_k(cubic_data->max_cwnd / 338 CCV(ccv, t_maxseg)); 339 } 340 } 341 } else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 342 !(ccv->flags & CCF_CWND_LIMITED)) { 343 cubic_data->flags |= CUBICFLAG_IN_APPLIMIT; 344 } 345 } 346 347 /* 348 * This is a CUBIC specific implementation of after_idle. 349 * - Reset cwnd by calling New Reno implementation of after_idle. 350 * - Reset t_last_cong. 351 */ 352 static void 353 cubic_after_idle(struct cc_var *ccv) 354 { 355 struct cubic *cubic_data; 356 357 cubic_data = ccv->cc_data; 358 359 cubic_data->max_cwnd = ulmax(cubic_data->max_cwnd, CCV(ccv, snd_cwnd)); 360 cubic_data->K = cubic_k(cubic_data->max_cwnd / CCV(ccv, t_maxseg)); 361 if ((cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) == 0) { 362 /* 363 * Re-enable hystart if we have been idle. 364 */ 365 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 366 cubic_data->flags |= CUBICFLAG_HYSTART_ENABLED; 367 cubic_log_hystart_event(ccv, cubic_data, 12, CCV(ccv, snd_ssthresh)); 368 } 369 newreno_cc_after_idle(ccv); 370 cubic_data->t_last_cong = ticks; 371 } 372 373 static void 374 cubic_cb_destroy(struct cc_var *ccv) 375 { 376 free(ccv->cc_data, M_CC_MEM); 377 } 378 379 static size_t 380 cubic_data_sz(void) 381 { 382 return (sizeof(struct cubic)); 383 } 384 385 static int 386 cubic_cb_init(struct cc_var *ccv, void *ptr) 387 { 388 struct cubic *cubic_data; 389 390 INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); 391 if (ptr == NULL) { 392 cubic_data = malloc(sizeof(struct cubic), M_CC_MEM, M_NOWAIT|M_ZERO); 393 if (cubic_data == NULL) 394 return (ENOMEM); 395 } else 396 cubic_data = ptr; 397 398 /* Init some key variables with sensible defaults. */ 399 cubic_data->t_last_cong = ticks; 400 cubic_data->min_rtt_ticks = TCPTV_SRTTBASE; 401 cubic_data->mean_rtt_ticks = 1; 402 403 ccv->cc_data = cubic_data; 404 cubic_data->flags = CUBICFLAG_HYSTART_ENABLED; 405 /* At init set both to infinity */ 406 cubic_data->css_lastround_minrtt = 0xffffffff; 407 cubic_data->css_current_round_minrtt = 0xffffffff; 408 cubic_data->css_current_round = 0; 409 cubic_data->css_baseline_minrtt = 0xffffffff; 410 cubic_data->css_rttsample_count = 0; 411 cubic_data->css_entered_at_round = 0; 412 cubic_data->css_fas_at_css_entry = 0; 413 cubic_data->css_lowrtt_fas = 0; 414 cubic_data->css_last_fas = 0; 415 416 return (0); 417 } 418 419 /* 420 * Perform any necessary tasks before we enter congestion recovery. 421 */ 422 static void 423 cubic_cong_signal(struct cc_var *ccv, uint32_t type) 424 { 425 struct cubic *cubic_data; 426 u_int mss; 427 428 cubic_data = ccv->cc_data; 429 mss = tcp_maxseg(ccv->ccvc.tcp); 430 431 switch (type) { 432 case CC_NDUPACK: 433 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 434 /* Make sure the flags are all off we had a loss */ 435 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 436 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 437 cubic_log_hystart_event(ccv, cubic_data, 10, CCV(ccv, snd_ssthresh)); 438 } 439 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 440 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 441 cubic_ssthresh_update(ccv, mss); 442 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 443 cubic_data->t_last_cong = ticks; 444 cubic_data->K = cubic_k(cubic_data->max_cwnd / mss); 445 } 446 ENTER_RECOVERY(CCV(ccv, t_flags)); 447 } 448 break; 449 450 case CC_ECN: 451 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 452 /* Make sure the flags are all off we had a loss */ 453 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 454 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 455 cubic_log_hystart_event(ccv, cubic_data, 9, CCV(ccv, snd_ssthresh)); 456 } 457 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 458 cubic_ssthresh_update(ccv, mss); 459 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 460 cubic_data->t_last_cong = ticks; 461 cubic_data->K = cubic_k(cubic_data->max_cwnd / mss); 462 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 463 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 464 } 465 break; 466 467 case CC_RTO: 468 /* RFC8312 Section 4.7 */ 469 if (CCV(ccv, t_rxtshift) == 1) { 470 cubic_data->t_last_cong_prev = cubic_data->t_last_cong; 471 cubic_data->prev_max_cwnd_cp = cubic_data->prev_max_cwnd; 472 } 473 cubic_data->flags |= CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT; 474 cubic_data->prev_max_cwnd = cubic_data->max_cwnd; 475 CCV(ccv, snd_ssthresh) = ((uint64_t)CCV(ccv, snd_cwnd) * 476 CUBIC_BETA) >> CUBIC_SHIFT; 477 CCV(ccv, snd_cwnd) = mss; 478 break; 479 480 case CC_RTO_ERR: 481 cubic_data->flags &= ~(CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT); 482 cubic_data->max_cwnd = cubic_data->prev_max_cwnd; 483 cubic_data->prev_max_cwnd = cubic_data->prev_max_cwnd_cp; 484 cubic_data->t_last_cong = cubic_data->t_last_cong_prev; 485 cubic_data->K = cubic_k(cubic_data->max_cwnd / mss); 486 break; 487 } 488 } 489 490 static void 491 cubic_conn_init(struct cc_var *ccv) 492 { 493 struct cubic *cubic_data; 494 495 cubic_data = ccv->cc_data; 496 497 /* 498 * Ensure we have a sane initial value for max_cwnd recorded. Without 499 * this here bad things happen when entries from the TCP hostcache 500 * get used. 501 */ 502 cubic_data->max_cwnd = CCV(ccv, snd_cwnd); 503 } 504 505 static int 506 cubic_mod_init(void) 507 { 508 return (0); 509 } 510 511 /* 512 * Perform any necessary tasks before we exit congestion recovery. 513 */ 514 static void 515 cubic_post_recovery(struct cc_var *ccv) 516 { 517 struct cubic *cubic_data; 518 int pipe; 519 520 cubic_data = ccv->cc_data; 521 pipe = 0; 522 523 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 524 /* 525 * If inflight data is less than ssthresh, set cwnd 526 * conservatively to avoid a burst of data, as suggested in 527 * the NewReno RFC. Otherwise, use the CUBIC method. 528 * 529 * XXXLAS: Find a way to do this without needing curack 530 */ 531 if (V_tcp_do_newsack) 532 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 533 else 534 pipe = CCV(ccv, snd_max) - ccv->curack; 535 536 if (pipe < CCV(ccv, snd_ssthresh)) 537 /* 538 * Ensure that cwnd does not collapse to 1 MSS under 539 * adverse conditions. Implements RFC6582 540 */ 541 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) + 542 CCV(ccv, t_maxseg); 543 else 544 /* Update cwnd based on beta and adjusted max_cwnd. */ 545 CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->max_cwnd * 546 CUBIC_BETA) >> CUBIC_SHIFT, 547 2 * CCV(ccv, t_maxseg)); 548 } 549 550 /* Calculate the average RTT between congestion epochs. */ 551 if (cubic_data->epoch_ack_count > 0 && 552 cubic_data->sum_rtt_ticks >= cubic_data->epoch_ack_count) { 553 cubic_data->mean_rtt_ticks = (int)(cubic_data->sum_rtt_ticks / 554 cubic_data->epoch_ack_count); 555 } 556 557 cubic_data->epoch_ack_count = 0; 558 cubic_data->sum_rtt_ticks = 0; 559 } 560 561 /* 562 * Record the min RTT and sum samples for the epoch average RTT calculation. 563 */ 564 static void 565 cubic_record_rtt(struct cc_var *ccv) 566 { 567 struct cubic *cubic_data; 568 int t_srtt_ticks; 569 570 /* Ignore srtt until a min number of samples have been taken. */ 571 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) { 572 cubic_data = ccv->cc_data; 573 t_srtt_ticks = CCV(ccv, t_srtt) / TCP_RTT_SCALE; 574 575 /* 576 * Record the current SRTT as our minrtt if it's the smallest 577 * we've seen or minrtt is currently equal to its initialised 578 * value. 579 * 580 * XXXLAS: Should there be some hysteresis for minrtt? 581 */ 582 if ((t_srtt_ticks < cubic_data->min_rtt_ticks || 583 cubic_data->min_rtt_ticks == TCPTV_SRTTBASE)) { 584 cubic_data->min_rtt_ticks = max(1, t_srtt_ticks); 585 586 /* 587 * If the connection is within its first congestion 588 * epoch, ensure we prime mean_rtt_ticks with a 589 * reasonable value until the epoch average RTT is 590 * calculated in cubic_post_recovery(). 591 */ 592 if (cubic_data->min_rtt_ticks > 593 cubic_data->mean_rtt_ticks) 594 cubic_data->mean_rtt_ticks = 595 cubic_data->min_rtt_ticks; 596 } 597 598 /* Sum samples for epoch average RTT calculation. */ 599 cubic_data->sum_rtt_ticks += t_srtt_ticks; 600 cubic_data->epoch_ack_count++; 601 } 602 } 603 604 /* 605 * Update the ssthresh in the event of congestion. 606 */ 607 static void 608 cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg) 609 { 610 struct cubic *cubic_data; 611 uint32_t ssthresh; 612 uint32_t cwnd; 613 614 cubic_data = ccv->cc_data; 615 cwnd = CCV(ccv, snd_cwnd); 616 617 /* Fast convergence heuristic. */ 618 if (cwnd < cubic_data->max_cwnd) { 619 cwnd = ((uint64_t)cwnd * CUBIC_FC_FACTOR) >> CUBIC_SHIFT; 620 } 621 cubic_data->prev_max_cwnd = cubic_data->max_cwnd; 622 cubic_data->max_cwnd = cwnd; 623 624 /* 625 * On the first congestion event, set ssthresh to cwnd * 0.5 626 * and reduce max_cwnd to cwnd * beta. This aligns the cubic concave 627 * region appropriately. On subsequent congestion events, set 628 * ssthresh to cwnd * beta. 629 */ 630 if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) { 631 ssthresh = cwnd >> 1; 632 cubic_data->max_cwnd = ((uint64_t)cwnd * 633 CUBIC_BETA) >> CUBIC_SHIFT; 634 } else { 635 ssthresh = ((uint64_t)cwnd * 636 CUBIC_BETA) >> CUBIC_SHIFT; 637 } 638 CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * maxseg); 639 } 640 641 static void 642 cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas) 643 { 644 struct cubic *cubicd; 645 646 cubicd = ccv->cc_data; 647 if (rxtcnt > 1) { 648 /* 649 * Only look at RTT's that are non-ambiguous. 650 */ 651 return; 652 } 653 cubicd->css_rttsample_count++; 654 cubicd->css_last_fas = fas; 655 if (cubicd->css_current_round_minrtt > usec_rtt) { 656 cubicd->css_current_round_minrtt = usec_rtt; 657 cubicd->css_lowrtt_fas = cubicd->css_last_fas; 658 } 659 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 660 (cubicd->css_current_round_minrtt != 0xffffffff) && 661 (cubicd->css_current_round_minrtt < cubicd->css_baseline_minrtt) && 662 (cubicd->css_lastround_minrtt != 0xffffffff)) { 663 /* 664 * We were in CSS and the RTT is now less, we 665 * entered CSS erroneously. 666 */ 667 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 668 cubic_log_hystart_event(ccv, cubicd, 8, cubicd->css_baseline_minrtt); 669 cubicd->css_baseline_minrtt = 0xffffffff; 670 } 671 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 672 cubic_log_hystart_event(ccv, cubicd, 5, usec_rtt); 673 } 674 675 static void 676 cubic_newround(struct cc_var *ccv, uint32_t round_cnt) 677 { 678 struct cubic *cubicd; 679 680 cubicd = ccv->cc_data; 681 /* We have entered a new round */ 682 cubicd->css_lastround_minrtt = cubicd->css_current_round_minrtt; 683 cubicd->css_current_round_minrtt = 0xffffffff; 684 cubicd->css_rttsample_count = 0; 685 cubicd->css_current_round = round_cnt; 686 if ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) && 687 ((round_cnt - cubicd->css_entered_at_round) >= hystart_css_rounds)) { 688 /* Enter CA */ 689 if (ccv->flags & CCF_HYSTART_CAN_SH_CWND) { 690 /* 691 * We engage more than snd_ssthresh, engage 692 * the brakes!! Though we will stay in SS to 693 * creep back up again, so lets leave CSS active 694 * and give us hystart_css_rounds more rounds. 695 */ 696 if (ccv->flags & CCF_HYSTART_CONS_SSTH) { 697 CCV(ccv, snd_ssthresh) = ((cubicd->css_lowrtt_fas + cubicd->css_fas_at_css_entry) / 2); 698 } else { 699 CCV(ccv, snd_ssthresh) = cubicd->css_lowrtt_fas; 700 } 701 CCV(ccv, snd_cwnd) = cubicd->css_fas_at_css_entry; 702 cubicd->css_entered_at_round = round_cnt; 703 } else { 704 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 705 /* Turn off the CSS flag */ 706 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 707 /* Disable use of CSS in the future except long idle */ 708 cubicd->flags &= ~CUBICFLAG_HYSTART_ENABLED; 709 } 710 cubic_log_hystart_event(ccv, cubicd, 6, CCV(ccv, snd_ssthresh)); 711 } 712 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 713 cubic_log_hystart_event(ccv, cubicd, 4, round_cnt); 714 } 715 716 DECLARE_CC_MODULE(cubic, &cubic_cc_algo); 717 MODULE_VERSION(cubic, 2); 718