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 /* 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/param.h> 51 #include <sys/kernel.h> 52 #include <sys/limits.h> 53 #include <sys/malloc.h> 54 #include <sys/module.h> 55 #include <sys/socket.h> 56 #include <sys/socketvar.h> 57 #include <sys/sysctl.h> 58 #include <sys/systm.h> 59 60 #include <net/vnet.h> 61 62 #include <net/route.h> 63 #include <net/route/nhop.h> 64 65 #include <netinet/in_pcb.h> 66 #include <netinet/tcp.h> 67 #include <netinet/tcp_seq.h> 68 #include <netinet/tcp_timer.h> 69 #include <netinet/tcp_var.h> 70 #include <netinet/tcp_log_buf.h> 71 #include <netinet/tcp_hpts.h> 72 #include <netinet/cc/cc.h> 73 #include <netinet/cc/cc_cubic.h> 74 #include <netinet/cc/cc_module.h> 75 76 static void cubic_ack_received(struct cc_var *ccv, ccsignal_t type); 77 static void cubic_cb_destroy(struct cc_var *ccv); 78 static int cubic_cb_init(struct cc_var *ccv, void *ptr); 79 static void cubic_cong_signal(struct cc_var *ccv, ccsignal_t type); 80 static void cubic_conn_init(struct cc_var *ccv); 81 static int cubic_mod_init(void); 82 static void cubic_post_recovery(struct cc_var *ccv); 83 static void cubic_record_rtt(struct cc_var *ccv); 84 static void cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg); 85 static void cubic_after_idle(struct cc_var *ccv); 86 static size_t cubic_data_sz(void); 87 static void cubic_newround(struct cc_var *ccv, uint32_t round_cnt); 88 static void cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, 89 uint32_t rxtcnt, uint32_t fas); 90 91 struct cc_algo cubic_cc_algo = { 92 .name = "cubic", 93 .ack_received = cubic_ack_received, 94 .cb_destroy = cubic_cb_destroy, 95 .cb_init = cubic_cb_init, 96 .cong_signal = cubic_cong_signal, 97 .conn_init = cubic_conn_init, 98 .mod_init = cubic_mod_init, 99 .post_recovery = cubic_post_recovery, 100 .after_idle = cubic_after_idle, 101 .cc_data_sz = cubic_data_sz, 102 .rttsample = cubic_rttsample, 103 .newround = cubic_newround 104 }; 105 106 static void 107 cubic_log_hystart_event(struct cc_var *ccv, struct cubic *cubicd, uint8_t mod, uint32_t flex1) 108 { 109 /* 110 * Types of logs (mod value) 111 * 1 - rtt_thresh in flex1, checking to see if RTT is to great. 112 * 2 - rtt is too great, rtt_thresh in flex1. 113 * 3 - CSS is active incr in flex1 114 * 4 - A new round is beginning flex1 is round count 115 * 5 - A new RTT measurement flex1 is the new measurement. 116 * 6 - We enter CA ssthresh is also in flex1. 117 * 7 - Socket option to change hystart executed opt.val in flex1. 118 * 8 - Back out of CSS into SS, flex1 is the css_baseline_minrtt 119 * 9 - We enter CA, via an ECN mark. 120 * 10 - We enter CA, via a loss. 121 * 11 - We have slipped out of SS into CA via cwnd growth. 122 * 12 - After idle has re-enabled hystart++ 123 */ 124 struct tcpcb *tp; 125 126 if (hystart_bblogs == 0) 127 return; 128 tp = ccv->tp; 129 if (tcp_bblogging_on(tp)) { 130 union tcp_log_stackspecific log; 131 struct timeval tv; 132 133 memset(&log, 0, sizeof(log)); 134 log.u_bbr.flex1 = flex1; 135 log.u_bbr.flex2 = cubicd->css_current_round_minrtt; 136 log.u_bbr.flex3 = cubicd->css_lastround_minrtt; 137 log.u_bbr.flex4 = cubicd->css_rttsample_count; 138 log.u_bbr.flex5 = cubicd->css_entered_at_round; 139 log.u_bbr.flex6 = cubicd->css_baseline_minrtt; 140 /* We only need bottom 16 bits of flags */ 141 log.u_bbr.flex7 = cubicd->flags & 0x0000ffff; 142 log.u_bbr.flex8 = mod; 143 log.u_bbr.epoch = cubicd->css_current_round; 144 log.u_bbr.timeStamp = tcp_get_usecs(&tv); 145 log.u_bbr.lt_epoch = cubicd->css_fas_at_css_entry; 146 log.u_bbr.pkts_out = cubicd->css_last_fas; 147 log.u_bbr.delivered = cubicd->css_lowrtt_fas; 148 log.u_bbr.pkt_epoch = ccv->flags; 149 TCP_LOG_EVENTP(tp, NULL, 150 &tptosocket(tp)->so_rcv, 151 &tptosocket(tp)->so_snd, 152 TCP_HYSTART, 0, 153 0, &log, false, &tv); 154 } 155 } 156 157 static void 158 cubic_does_slow_start(struct cc_var *ccv, struct cubic *cubicd) 159 { 160 /* 161 * In slow-start with ABC enabled and no RTO in sight? 162 * (Must not use abc_l_var > 1 if slow starting after 163 * an RTO. On RTO, snd_nxt = snd_una, so the 164 * snd_nxt == snd_max check is sufficient to 165 * handle this). 166 * 167 * XXXLAS: Find a way to signal SS after RTO that 168 * doesn't rely on tcpcb vars. 169 */ 170 u_int cw = CCV(ccv, snd_cwnd); 171 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 172 u_int incr = mss; 173 uint16_t abc_val; 174 175 cubicd->flags |= CUBICFLAG_IN_SLOWSTART; 176 if (ccv->flags & CCF_USE_LOCAL_ABC) 177 abc_val = ccv->labc; 178 else 179 abc_val = V_tcp_abc_l_var; 180 if ((ccv->flags & CCF_HYSTART_ALLOWED) && 181 (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) && 182 ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) == 0)) { 183 /* 184 * Hystart is allowed and still enabled and we are not yet 185 * in CSS. Lets check to see if we can make a decision on 186 * if we need to go into CSS. 187 */ 188 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 189 (cubicd->css_current_round_minrtt != 0xffffffff) && 190 (cubicd->css_lastround_minrtt != 0xffffffff)) { 191 uint32_t rtt_thresh; 192 193 /* Clamp (minrtt_thresh, lastround/8, maxrtt_thresh) */ 194 rtt_thresh = (cubicd->css_lastround_minrtt >> 3); 195 if (rtt_thresh < hystart_minrtt_thresh) 196 rtt_thresh = hystart_minrtt_thresh; 197 if (rtt_thresh > hystart_maxrtt_thresh) 198 rtt_thresh = hystart_maxrtt_thresh; 199 cubic_log_hystart_event(ccv, cubicd, 1, rtt_thresh); 200 201 if (cubicd->css_current_round_minrtt >= (cubicd->css_lastround_minrtt + rtt_thresh)) { 202 /* Enter CSS */ 203 cubicd->flags |= CUBICFLAG_HYSTART_IN_CSS; 204 cubicd->css_fas_at_css_entry = cubicd->css_lowrtt_fas; 205 /* 206 * The draft (v4) calls for us to set baseline to css_current_round_min 207 * but that can cause an oscillation. We probably shoudl be using 208 * css_lastround_minrtt, but the authors insist that will cause 209 * issues on exiting early. We will leave the draft version for now 210 * but I suspect this is incorrect. 211 */ 212 cubicd->css_baseline_minrtt = cubicd->css_current_round_minrtt; 213 cubicd->css_entered_at_round = cubicd->css_current_round; 214 cubic_log_hystart_event(ccv, cubicd, 2, rtt_thresh); 215 } 216 } 217 } 218 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) 219 incr = min(ccv->bytes_this_ack, 220 ccv->nsegs * abc_val * mss); 221 else 222 incr = min(ccv->bytes_this_ack, mss); 223 224 /* Only if Hystart is enabled will the flag get set */ 225 if (cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) { 226 incr /= hystart_css_growth_div; 227 cubic_log_hystart_event(ccv, cubicd, 3, incr); 228 } 229 /* ABC is on by default, so incr equals 0 frequently. */ 230 if (incr > 0) 231 CCV(ccv, snd_cwnd) = min((cw + incr), 232 TCP_MAXWIN << CCV(ccv, snd_scale)); 233 } 234 235 static void 236 cubic_ack_received(struct cc_var *ccv, ccsignal_t type) 237 { 238 struct cubic *cubic_data; 239 unsigned long W_est, W_cubic; 240 int usecs_since_epoch; 241 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 242 243 cubic_data = ccv->cc_data; 244 cubic_record_rtt(ccv); 245 246 /* 247 * For a regular ACK and we're not in cong/fast recovery and 248 * we're cwnd limited, always recalculate cwnd. 249 */ 250 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 251 (ccv->flags & CCF_CWND_LIMITED)) { 252 /* Use the logic in NewReno ack_received() for slow start. */ 253 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 254 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE) { 255 cubic_does_slow_start(ccv, cubic_data); 256 } else { 257 if (cubic_data->flags & CUBICFLAG_HYSTART_IN_CSS) { 258 /* 259 * We have slipped into CA with 260 * CSS active. Deactivate all. 261 */ 262 /* Turn off the CSS flag */ 263 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 264 /* Disable use of CSS in the future except long idle */ 265 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 266 cubic_log_hystart_event(ccv, cubic_data, 11, CCV(ccv, snd_ssthresh)); 267 } 268 if ((cubic_data->flags & CUBICFLAG_RTO_EVENT) && 269 (cubic_data->flags & CUBICFLAG_IN_SLOWSTART)) { 270 /* RFC8312 Section 4.7 */ 271 cubic_data->flags &= ~(CUBICFLAG_RTO_EVENT | 272 CUBICFLAG_IN_SLOWSTART); 273 cubic_data->W_max = CCV(ccv, snd_cwnd); 274 cubic_data->t_epoch = ticks; 275 cubic_data->K = 0; 276 } else if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART | 277 CUBICFLAG_IN_APPLIMIT)) { 278 cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART | 279 CUBICFLAG_IN_APPLIMIT); 280 cubic_data->t_epoch = ticks; 281 cubic_data->K = cubic_k(cubic_data->W_max / mss); 282 } 283 usecs_since_epoch = (ticks - cubic_data->t_epoch) * tick; 284 if (usecs_since_epoch < 0) { 285 /* 286 * dragging t_epoch along 287 */ 288 usecs_since_epoch = INT_MAX; 289 cubic_data->t_epoch = ticks - INT_MAX; 290 } 291 292 W_est = tf_cwnd(ccv); 293 294 /* 295 * The mean RTT is used to best reflect the equations in 296 * the I-D. 297 */ 298 W_cubic = cubic_cwnd(usecs_since_epoch + 299 cubic_data->mean_rtt_usecs, 300 cubic_data->W_max, 301 mss, 302 cubic_data->K); 303 304 if (W_cubic < W_est) { 305 /* 306 * TCP-friendly region, follow tf 307 * cwnd growth. 308 */ 309 CCV(ccv, snd_cwnd) = ulmin(W_est, INT_MAX); 310 cubic_data->flags |= CUBICFLAG_IN_TF; 311 } else if (CCV(ccv, snd_cwnd) < W_cubic) { 312 /* 313 * Concave or convex region, follow CUBIC 314 * cwnd growth. 315 * Only update snd_cwnd, if it doesn't shrink. 316 */ 317 CCV(ccv, snd_cwnd) = ulmin(W_cubic, INT_MAX); 318 cubic_data->flags &= ~CUBICFLAG_IN_TF; 319 } 320 321 /* 322 * If we're not in slow start and we're probing for a 323 * new cwnd limit at the start of a connection 324 * (happens when hostcache has a relevant entry), 325 * keep updating our current estimate of the 326 * W_max. 327 */ 328 if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) && 329 cubic_data->W_max < CCV(ccv, snd_cwnd)) { 330 cubic_data->W_max = CCV(ccv, snd_cwnd); 331 cubic_data->K = cubic_k(cubic_data->W_max / mss); 332 } 333 } 334 } else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 335 !(ccv->flags & CCF_CWND_LIMITED)) { 336 cubic_data->flags |= CUBICFLAG_IN_APPLIMIT; 337 } 338 } 339 340 /* 341 * This is a CUBIC specific implementation of after_idle. 342 * - Reset cwnd by calling New Reno implementation of after_idle. 343 * - Reset t_epoch. 344 */ 345 static void 346 cubic_after_idle(struct cc_var *ccv) 347 { 348 struct cubic *cubic_data; 349 350 cubic_data = ccv->cc_data; 351 352 cubic_data->W_max = ulmax(cubic_data->W_max, CCV(ccv, snd_cwnd)); 353 cubic_data->K = cubic_k(cubic_data->W_max / tcp_fixed_maxseg(ccv->tp)); 354 if ((cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) == 0) { 355 /* 356 * Re-enable hystart if we have been idle. 357 */ 358 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 359 cubic_data->flags |= CUBICFLAG_HYSTART_ENABLED; 360 cubic_log_hystart_event(ccv, cubic_data, 12, CCV(ccv, snd_ssthresh)); 361 } 362 newreno_cc_after_idle(ccv); 363 cubic_data->t_epoch = ticks; 364 } 365 366 static void 367 cubic_cb_destroy(struct cc_var *ccv) 368 { 369 free(ccv->cc_data, M_CC_MEM); 370 } 371 372 static size_t 373 cubic_data_sz(void) 374 { 375 return (sizeof(struct cubic)); 376 } 377 378 static int 379 cubic_cb_init(struct cc_var *ccv, void *ptr) 380 { 381 struct cubic *cubic_data; 382 383 INP_WLOCK_ASSERT(tptoinpcb(ccv->tp)); 384 if (ptr == NULL) { 385 cubic_data = malloc(sizeof(struct cubic), M_CC_MEM, M_NOWAIT|M_ZERO); 386 if (cubic_data == NULL) 387 return (ENOMEM); 388 } else 389 cubic_data = ptr; 390 391 /* Init some key variables with sensible defaults. */ 392 cubic_data->t_epoch = ticks; 393 cubic_data->min_rtt_usecs = TCPTV_SRTTBASE; 394 cubic_data->mean_rtt_usecs = 1; 395 396 ccv->cc_data = cubic_data; 397 cubic_data->flags = CUBICFLAG_HYSTART_ENABLED; 398 /* At init set both to infinity */ 399 cubic_data->css_lastround_minrtt = 0xffffffff; 400 cubic_data->css_current_round_minrtt = 0xffffffff; 401 cubic_data->css_current_round = 0; 402 cubic_data->css_baseline_minrtt = 0xffffffff; 403 cubic_data->css_rttsample_count = 0; 404 cubic_data->css_entered_at_round = 0; 405 cubic_data->css_fas_at_css_entry = 0; 406 cubic_data->css_lowrtt_fas = 0; 407 cubic_data->css_last_fas = 0; 408 409 return (0); 410 } 411 412 /* 413 * Perform any necessary tasks before we enter congestion recovery. 414 */ 415 static void 416 cubic_cong_signal(struct cc_var *ccv, ccsignal_t type) 417 { 418 struct cubic *cubic_data; 419 uint32_t mss, pipe; 420 421 cubic_data = ccv->cc_data; 422 mss = tcp_fixed_maxseg(ccv->tp); 423 424 switch (type) { 425 case CC_NDUPACK: 426 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 427 /* Make sure the flags are all off we had a loss */ 428 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 429 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 430 cubic_log_hystart_event(ccv, cubic_data, 10, CCV(ccv, snd_ssthresh)); 431 } 432 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 433 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 434 cubic_ssthresh_update(ccv, mss); 435 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 436 cubic_data->t_epoch = ticks; 437 cubic_data->K = cubic_k(cubic_data->W_max / mss); 438 } 439 ENTER_RECOVERY(CCV(ccv, t_flags)); 440 } 441 break; 442 443 case CC_ECN: 444 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 445 /* Make sure the flags are all off we had a loss */ 446 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 447 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 448 cubic_log_hystart_event(ccv, cubic_data, 9, CCV(ccv, snd_ssthresh)); 449 } 450 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 451 cubic_ssthresh_update(ccv, mss); 452 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 453 cubic_data->t_epoch = ticks; 454 cubic_data->K = cubic_k(cubic_data->W_max / mss); 455 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 456 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 457 } 458 break; 459 460 case CC_RTO: 461 /* RFC8312 Section 4.7 */ 462 if (CCV(ccv, t_rxtshift) == 1) { 463 /* 464 * Remember the state only for the first RTO event. This 465 * will help us restore the state to the values seen 466 * at the most recent congestion avoidance stage before 467 * the current RTO event. 468 */ 469 cubic_data->undo_t_epoch = cubic_data->t_epoch; 470 cubic_data->undo_cwnd_epoch = cubic_data->cwnd_epoch; 471 cubic_data->undo_W_max = cubic_data->W_max; 472 cubic_data->undo_K = cubic_data->K; 473 pipe = tcp_compute_pipe(ccv->tp); 474 CCV(ccv, snd_ssthresh) = max(2, 475 (((uint64_t)min(CCV(ccv, snd_wnd), pipe) * 476 CUBIC_BETA) >> CUBIC_SHIFT) / mss) * mss; 477 } 478 cubic_data->flags |= CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT; 479 CCV(ccv, snd_cwnd) = mss; 480 break; 481 482 case CC_RTO_ERR: 483 cubic_data->flags &= ~(CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT); 484 cubic_data->K = cubic_data->undo_K; 485 cubic_data->W_max = cubic_data->undo_W_max; 486 cubic_data->cwnd_epoch = cubic_data->undo_cwnd_epoch; 487 cubic_data->t_epoch = cubic_data->undo_t_epoch; 488 break; 489 default: 490 break; 491 } 492 } 493 494 static void 495 cubic_conn_init(struct cc_var *ccv) 496 { 497 struct cubic *cubic_data; 498 499 cubic_data = ccv->cc_data; 500 501 /* 502 * Ensure we have a sane initial value for W_max recorded. Without 503 * this here bad things happen when entries from the TCP hostcache 504 * get used. 505 */ 506 cubic_data->W_max = CCV(ccv, snd_cwnd); 507 } 508 509 static int 510 cubic_mod_init(void) 511 { 512 return (0); 513 } 514 515 /* 516 * Perform any necessary tasks before we exit congestion recovery. 517 */ 518 static void 519 cubic_post_recovery(struct cc_var *ccv) 520 { 521 struct cubic *cubic_data; 522 int pipe; 523 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 524 525 cubic_data = ccv->cc_data; 526 pipe = 0; 527 528 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 529 /* 530 * If inflight data is less than ssthresh, set cwnd 531 * conservatively to avoid a burst of data, as suggested in 532 * the NewReno RFC. Otherwise, use the CUBIC method. 533 */ 534 pipe = tcp_compute_pipe(ccv->tp); 535 if (pipe < CCV(ccv, snd_ssthresh)) 536 /* 537 * Ensure that cwnd does not collapse to 1 MSS under 538 * adverse conditions. Implements RFC6582 539 */ 540 CCV(ccv, snd_cwnd) = max(pipe, mss) + mss; 541 else 542 /* Update cwnd based on beta and adjusted W_max. */ 543 CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->W_max * 544 CUBIC_BETA) >> CUBIC_SHIFT, 545 2 * mss); 546 } 547 548 /* Calculate the average RTT between congestion epochs. */ 549 if (cubic_data->epoch_ack_count > 0 && 550 cubic_data->sum_rtt_usecs >= cubic_data->epoch_ack_count) { 551 cubic_data->mean_rtt_usecs = (int)(cubic_data->sum_rtt_usecs / 552 cubic_data->epoch_ack_count); 553 } 554 555 cubic_data->epoch_ack_count = 0; 556 cubic_data->sum_rtt_usecs = 0; 557 } 558 559 /* 560 * Record the min RTT and sum samples for the epoch average RTT calculation. 561 */ 562 static void 563 cubic_record_rtt(struct cc_var *ccv) 564 { 565 struct cubic *cubic_data; 566 uint32_t t_srtt_usecs; 567 568 /* Ignore srtt until a min number of samples have been taken. */ 569 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) { 570 cubic_data = ccv->cc_data; 571 t_srtt_usecs = tcp_get_srtt(ccv->tp, 572 TCP_TMR_GRANULARITY_USEC); 573 /* 574 * Record the current SRTT as our minrtt if it's the smallest 575 * we've seen or minrtt is currently equal to its initialised 576 * value. 577 * 578 * XXXLAS: Should there be some hysteresis for minrtt? 579 */ 580 if ((t_srtt_usecs < cubic_data->min_rtt_usecs || 581 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE)) { 582 /* A minimal rtt is a single unshifted tick of a ticks 583 * timer. */ 584 cubic_data->min_rtt_usecs = max(tick >> TCP_RTT_SHIFT, 585 t_srtt_usecs); 586 587 /* 588 * If the connection is within its first congestion 589 * epoch, ensure we prime mean_rtt_usecs with a 590 * reasonable value until the epoch average RTT is 591 * calculated in cubic_post_recovery(). 592 */ 593 if (cubic_data->min_rtt_usecs > 594 cubic_data->mean_rtt_usecs) 595 cubic_data->mean_rtt_usecs = 596 cubic_data->min_rtt_usecs; 597 } 598 599 /* Sum samples for epoch average RTT calculation. */ 600 cubic_data->sum_rtt_usecs += t_srtt_usecs; 601 cubic_data->epoch_ack_count++; 602 } 603 } 604 605 /* 606 * Update the ssthresh in the event of congestion. 607 */ 608 static void 609 cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg) 610 { 611 struct cubic *cubic_data; 612 uint32_t ssthresh; 613 uint32_t cwnd; 614 615 cubic_data = ccv->cc_data; 616 cwnd = CCV(ccv, snd_cwnd); 617 618 /* Fast convergence heuristic. */ 619 if (cwnd < cubic_data->W_max) { 620 cwnd = ((uint64_t)cwnd * CUBIC_FC_FACTOR) >> CUBIC_SHIFT; 621 } 622 cubic_data->undo_W_max = cubic_data->W_max; 623 cubic_data->W_max = cwnd; 624 625 if (cubic_data->flags & CUBICFLAG_IN_TF) { 626 /* If in the TCP friendly region, follow what newreno does */ 627 ssthresh = newreno_cc_cwnd_on_multiplicative_decrease(ccv, maxseg); 628 629 } else if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) { 630 /* 631 * On the first congestion event, set ssthresh to cwnd * 0.5 632 * and reduce W_max to cwnd * beta. This aligns the cubic 633 * concave region appropriately. 634 */ 635 ssthresh = cwnd >> 1; 636 cubic_data->W_max = ((uint64_t)cwnd * CUBIC_BETA) >> CUBIC_SHIFT; 637 } else { 638 /* 639 * On subsequent congestion events, set ssthresh to cwnd * beta. 640 */ 641 ssthresh = ((uint64_t)cwnd * CUBIC_BETA) >> CUBIC_SHIFT; 642 } 643 CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * maxseg); 644 } 645 646 static void 647 cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas) 648 { 649 struct cubic *cubicd; 650 651 cubicd = ccv->cc_data; 652 if (rxtcnt > 1) { 653 /* 654 * Only look at RTT's that are non-ambiguous. 655 */ 656 return; 657 } 658 cubicd->css_rttsample_count++; 659 cubicd->css_last_fas = fas; 660 if (cubicd->css_current_round_minrtt > usec_rtt) { 661 cubicd->css_current_round_minrtt = usec_rtt; 662 cubicd->css_lowrtt_fas = cubicd->css_last_fas; 663 } 664 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 665 (cubicd->css_current_round_minrtt != 0xffffffff) && 666 (cubicd->css_current_round_minrtt < cubicd->css_baseline_minrtt) && 667 (cubicd->css_lastround_minrtt != 0xffffffff)) { 668 /* 669 * We were in CSS and the RTT is now less, we 670 * entered CSS erroneously. 671 */ 672 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 673 cubic_log_hystart_event(ccv, cubicd, 8, cubicd->css_baseline_minrtt); 674 cubicd->css_baseline_minrtt = 0xffffffff; 675 } 676 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 677 cubic_log_hystart_event(ccv, cubicd, 5, usec_rtt); 678 } 679 680 static void 681 cubic_newround(struct cc_var *ccv, uint32_t round_cnt) 682 { 683 struct cubic *cubicd; 684 685 cubicd = ccv->cc_data; 686 /* We have entered a new round */ 687 cubicd->css_lastround_minrtt = cubicd->css_current_round_minrtt; 688 cubicd->css_current_round_minrtt = 0xffffffff; 689 cubicd->css_rttsample_count = 0; 690 cubicd->css_current_round = round_cnt; 691 if ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) && 692 ((round_cnt - cubicd->css_entered_at_round) >= hystart_css_rounds)) { 693 /* Enter CA */ 694 if (ccv->flags & CCF_HYSTART_CAN_SH_CWND) { 695 /* 696 * We engage more than snd_ssthresh, engage 697 * the brakes!! Though we will stay in SS to 698 * creep back up again, so lets leave CSS active 699 * and give us hystart_css_rounds more rounds. 700 */ 701 if (ccv->flags & CCF_HYSTART_CONS_SSTH) { 702 CCV(ccv, snd_ssthresh) = ((cubicd->css_lowrtt_fas + cubicd->css_fas_at_css_entry) / 2); 703 } else { 704 CCV(ccv, snd_ssthresh) = cubicd->css_lowrtt_fas; 705 } 706 CCV(ccv, snd_cwnd) = cubicd->css_fas_at_css_entry; 707 cubicd->css_entered_at_round = round_cnt; 708 } else { 709 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 710 /* Turn off the CSS flag */ 711 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 712 /* Disable use of CSS in the future except long idle */ 713 cubicd->flags &= ~CUBICFLAG_HYSTART_ENABLED; 714 } 715 cubic_log_hystart_event(ccv, cubicd, 6, CCV(ccv, snd_ssthresh)); 716 } 717 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 718 cubic_log_hystart_event(ccv, cubicd, 4, round_cnt); 719 } 720 721 DECLARE_CC_MODULE(cubic, &cubic_cc_algo); 722 MODULE_VERSION(cubic, 2); 723