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, uint16_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, uint32_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->ccvc.tcp; 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 u_int incr = CCV(ccv, t_maxseg); 172 uint16_t abc_val; 173 174 cubicd->flags |= CUBICFLAG_IN_SLOWSTART; 175 if (ccv->flags & CCF_USE_LOCAL_ABC) 176 abc_val = ccv->labc; 177 else 178 abc_val = V_tcp_abc_l_var; 179 if ((ccv->flags & CCF_HYSTART_ALLOWED) && 180 (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) && 181 ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) == 0)) { 182 /* 183 * Hystart is allowed and still enabled and we are not yet 184 * in CSS. Lets check to see if we can make a decision on 185 * if we need to go into CSS. 186 */ 187 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 188 (cubicd->css_current_round_minrtt != 0xffffffff) && 189 (cubicd->css_lastround_minrtt != 0xffffffff)) { 190 uint32_t rtt_thresh; 191 192 /* Clamp (minrtt_thresh, lastround/8, maxrtt_thresh) */ 193 rtt_thresh = (cubicd->css_lastround_minrtt >> 3); 194 if (rtt_thresh < hystart_minrtt_thresh) 195 rtt_thresh = hystart_minrtt_thresh; 196 if (rtt_thresh > hystart_maxrtt_thresh) 197 rtt_thresh = hystart_maxrtt_thresh; 198 cubic_log_hystart_event(ccv, cubicd, 1, rtt_thresh); 199 200 if (cubicd->css_current_round_minrtt >= (cubicd->css_lastround_minrtt + rtt_thresh)) { 201 /* Enter CSS */ 202 cubicd->flags |= CUBICFLAG_HYSTART_IN_CSS; 203 cubicd->css_fas_at_css_entry = cubicd->css_lowrtt_fas; 204 /* 205 * The draft (v4) calls for us to set baseline to css_current_round_min 206 * but that can cause an oscillation. We probably shoudl be using 207 * css_lastround_minrtt, but the authors insist that will cause 208 * issues on exiting early. We will leave the draft version for now 209 * but I suspect this is incorrect. 210 */ 211 cubicd->css_baseline_minrtt = cubicd->css_current_round_minrtt; 212 cubicd->css_entered_at_round = cubicd->css_current_round; 213 cubic_log_hystart_event(ccv, cubicd, 2, rtt_thresh); 214 } 215 } 216 } 217 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) 218 incr = min(ccv->bytes_this_ack, 219 ccv->nsegs * abc_val * 220 CCV(ccv, t_maxseg)); 221 else 222 incr = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg)); 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, uint16_t type) 237 { 238 struct cubic *cubic_data; 239 unsigned long W_est, W_cubic; 240 int usecs_since_epoch; 241 242 cubic_data = ccv->cc_data; 243 cubic_record_rtt(ccv); 244 245 /* 246 * For a regular ACK and we're not in cong/fast recovery and 247 * we're cwnd limited, always recalculate cwnd. 248 */ 249 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 250 (ccv->flags & CCF_CWND_LIMITED)) { 251 /* Use the logic in NewReno ack_received() for slow start. */ 252 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 253 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE) { 254 cubic_does_slow_start(ccv, cubic_data); 255 } else { 256 if (cubic_data->flags & CUBICFLAG_HYSTART_IN_CSS) { 257 /* 258 * We have slipped into CA with 259 * CSS active. Deactivate all. 260 */ 261 /* Turn off the CSS flag */ 262 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 263 /* Disable use of CSS in the future except long idle */ 264 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 265 cubic_log_hystart_event(ccv, cubic_data, 11, CCV(ccv, snd_ssthresh)); 266 } 267 if ((cubic_data->flags & CUBICFLAG_RTO_EVENT) && 268 (cubic_data->flags & CUBICFLAG_IN_SLOWSTART)) { 269 /* RFC8312 Section 4.7 */ 270 cubic_data->flags &= ~(CUBICFLAG_RTO_EVENT | 271 CUBICFLAG_IN_SLOWSTART); 272 cubic_data->W_max = CCV(ccv, snd_cwnd); 273 cubic_data->K = 0; 274 } else if (cubic_data->flags & (CUBICFLAG_IN_SLOWSTART | 275 CUBICFLAG_IN_APPLIMIT)) { 276 cubic_data->flags &= ~(CUBICFLAG_IN_SLOWSTART | 277 CUBICFLAG_IN_APPLIMIT); 278 cubic_data->t_epoch = ticks; 279 cubic_data->K = cubic_k(cubic_data->W_max / 280 CCV(ccv, t_maxseg)); 281 } 282 usecs_since_epoch = (ticks - cubic_data->t_epoch) * tick; 283 if (usecs_since_epoch < 0) { 284 /* 285 * dragging t_epoch along 286 */ 287 usecs_since_epoch = INT_MAX; 288 cubic_data->t_epoch = ticks - INT_MAX; 289 } 290 /* 291 * The mean RTT is used to best reflect the equations in 292 * the I-D. Using min_rtt in the tf_cwnd calculation 293 * causes W_est to grow much faster than it should if the 294 * RTT is dominated by network buffering rather than 295 * propagation delay. 296 */ 297 W_est = tf_cwnd(usecs_since_epoch, cubic_data->mean_rtt_usecs, 298 cubic_data->W_max, CCV(ccv, t_maxseg)); 299 300 W_cubic = cubic_cwnd(usecs_since_epoch + 301 cubic_data->mean_rtt_usecs, 302 cubic_data->W_max, 303 CCV(ccv, t_maxseg), 304 cubic_data->K); 305 306 ccv->flags &= ~CCF_ABC_SENTAWND; 307 308 if (W_cubic < W_est) { 309 /* 310 * TCP-friendly region, follow tf 311 * cwnd growth. 312 */ 313 if (CCV(ccv, snd_cwnd) < W_est) 314 CCV(ccv, snd_cwnd) = ulmin(W_est, INT_MAX); 315 } else if (CCV(ccv, snd_cwnd) < W_cubic) { 316 /* 317 * Concave or convex region, follow CUBIC 318 * cwnd growth. 319 * Only update snd_cwnd, if it doesn't shrink. 320 */ 321 CCV(ccv, snd_cwnd) = ulmin(W_cubic, INT_MAX); 322 } 323 324 /* 325 * If we're not in slow start and we're probing for a 326 * new cwnd limit at the start of a connection 327 * (happens when hostcache has a relevant entry), 328 * keep updating our current estimate of the 329 * W_max. 330 */ 331 if (((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) && 332 cubic_data->W_max < CCV(ccv, snd_cwnd)) { 333 cubic_data->W_max = CCV(ccv, snd_cwnd); 334 cubic_data->K = cubic_k(cubic_data->W_max / 335 CCV(ccv, t_maxseg)); 336 } 337 } 338 } else if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 339 !(ccv->flags & CCF_CWND_LIMITED)) { 340 cubic_data->flags |= CUBICFLAG_IN_APPLIMIT; 341 } 342 } 343 344 /* 345 * This is a CUBIC specific implementation of after_idle. 346 * - Reset cwnd by calling New Reno implementation of after_idle. 347 * - Reset t_epoch. 348 */ 349 static void 350 cubic_after_idle(struct cc_var *ccv) 351 { 352 struct cubic *cubic_data; 353 354 cubic_data = ccv->cc_data; 355 356 cubic_data->W_max = ulmax(cubic_data->W_max, CCV(ccv, snd_cwnd)); 357 cubic_data->K = cubic_k(cubic_data->W_max / CCV(ccv, t_maxseg)); 358 if ((cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) == 0) { 359 /* 360 * Re-enable hystart if we have been idle. 361 */ 362 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 363 cubic_data->flags |= CUBICFLAG_HYSTART_ENABLED; 364 cubic_log_hystart_event(ccv, cubic_data, 12, CCV(ccv, snd_ssthresh)); 365 } 366 newreno_cc_after_idle(ccv); 367 cubic_data->t_epoch = ticks; 368 } 369 370 static void 371 cubic_cb_destroy(struct cc_var *ccv) 372 { 373 free(ccv->cc_data, M_CC_MEM); 374 } 375 376 static size_t 377 cubic_data_sz(void) 378 { 379 return (sizeof(struct cubic)); 380 } 381 382 static int 383 cubic_cb_init(struct cc_var *ccv, void *ptr) 384 { 385 struct cubic *cubic_data; 386 387 INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); 388 if (ptr == NULL) { 389 cubic_data = malloc(sizeof(struct cubic), M_CC_MEM, M_NOWAIT|M_ZERO); 390 if (cubic_data == NULL) 391 return (ENOMEM); 392 } else 393 cubic_data = ptr; 394 395 /* Init some key variables with sensible defaults. */ 396 cubic_data->t_epoch = ticks; 397 cubic_data->min_rtt_usecs = TCPTV_SRTTBASE; 398 cubic_data->mean_rtt_usecs = 1; 399 400 ccv->cc_data = cubic_data; 401 cubic_data->flags = CUBICFLAG_HYSTART_ENABLED; 402 /* At init set both to infinity */ 403 cubic_data->css_lastround_minrtt = 0xffffffff; 404 cubic_data->css_current_round_minrtt = 0xffffffff; 405 cubic_data->css_current_round = 0; 406 cubic_data->css_baseline_minrtt = 0xffffffff; 407 cubic_data->css_rttsample_count = 0; 408 cubic_data->css_entered_at_round = 0; 409 cubic_data->css_fas_at_css_entry = 0; 410 cubic_data->css_lowrtt_fas = 0; 411 cubic_data->css_last_fas = 0; 412 413 return (0); 414 } 415 416 /* 417 * Perform any necessary tasks before we enter congestion recovery. 418 */ 419 static void 420 cubic_cong_signal(struct cc_var *ccv, uint32_t type) 421 { 422 struct cubic *cubic_data; 423 uint32_t mss, pipe; 424 425 cubic_data = ccv->cc_data; 426 mss = tcp_fixed_maxseg(ccv->ccvc.tcp); 427 428 switch (type) { 429 case CC_NDUPACK: 430 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 431 /* Make sure the flags are all off we had a loss */ 432 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 433 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 434 cubic_log_hystart_event(ccv, cubic_data, 10, CCV(ccv, snd_ssthresh)); 435 } 436 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 437 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 438 cubic_ssthresh_update(ccv, mss); 439 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 440 cubic_data->t_epoch = ticks; 441 cubic_data->K = cubic_k(cubic_data->W_max / mss); 442 } 443 ENTER_RECOVERY(CCV(ccv, t_flags)); 444 } 445 break; 446 447 case CC_ECN: 448 if (cubic_data->flags & CUBICFLAG_HYSTART_ENABLED) { 449 /* Make sure the flags are all off we had a loss */ 450 cubic_data->flags &= ~CUBICFLAG_HYSTART_ENABLED; 451 cubic_data->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 452 cubic_log_hystart_event(ccv, cubic_data, 9, CCV(ccv, snd_ssthresh)); 453 } 454 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 455 cubic_ssthresh_update(ccv, mss); 456 cubic_data->flags |= CUBICFLAG_CONG_EVENT; 457 cubic_data->t_epoch = ticks; 458 cubic_data->K = cubic_k(cubic_data->W_max / mss); 459 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 460 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 461 } 462 break; 463 464 case CC_RTO: 465 /* RFC8312 Section 4.7 */ 466 if (CCV(ccv, t_rxtshift) == 1) { 467 /* 468 * Remember the state only for the first RTO event. This 469 * will help us restore the state to the values seen 470 * at the most recent congestion avoidance stage before 471 * the current RTO event. 472 */ 473 cubic_data->undo_t_epoch = cubic_data->t_epoch; 474 cubic_data->undo_cwnd_epoch = cubic_data->cwnd_epoch; 475 cubic_data->undo_W_est = cubic_data->W_est; 476 cubic_data->undo_cwnd_prior = cubic_data->cwnd_prior; 477 cubic_data->undo_W_max = cubic_data->W_max; 478 cubic_data->undo_K = cubic_data->K; 479 if (V_tcp_do_newsack) { 480 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 481 } else { 482 pipe = CCV(ccv, snd_max) - 483 CCV(ccv, snd_fack) + 484 CCV(ccv, sackhint.sack_bytes_rexmit); 485 } 486 CCV(ccv, snd_ssthresh) = max(2, 487 (((uint64_t)min(CCV(ccv, snd_wnd), pipe) * 488 CUBIC_BETA) >> CUBIC_SHIFT) / mss) * mss; 489 } 490 cubic_data->flags |= CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT; 491 cubic_data->undo_W_max = cubic_data->W_max; 492 cubic_data->num_cong_events++; 493 CCV(ccv, snd_cwnd) = mss; 494 break; 495 496 case CC_RTO_ERR: 497 cubic_data->flags &= ~(CUBICFLAG_CONG_EVENT | CUBICFLAG_RTO_EVENT); 498 cubic_data->num_cong_events--; 499 cubic_data->K = cubic_data->undo_K; 500 cubic_data->cwnd_prior = cubic_data->undo_cwnd_prior; 501 cubic_data->W_max = cubic_data->undo_W_max; 502 cubic_data->W_est = cubic_data->undo_W_est; 503 cubic_data->cwnd_epoch = cubic_data->undo_cwnd_epoch; 504 cubic_data->t_epoch = cubic_data->undo_t_epoch; 505 break; 506 } 507 } 508 509 static void 510 cubic_conn_init(struct cc_var *ccv) 511 { 512 struct cubic *cubic_data; 513 514 cubic_data = ccv->cc_data; 515 516 /* 517 * Ensure we have a sane initial value for W_max recorded. Without 518 * this here bad things happen when entries from the TCP hostcache 519 * get used. 520 */ 521 cubic_data->W_max = CCV(ccv, snd_cwnd); 522 } 523 524 static int 525 cubic_mod_init(void) 526 { 527 return (0); 528 } 529 530 /* 531 * Perform any necessary tasks before we exit congestion recovery. 532 */ 533 static void 534 cubic_post_recovery(struct cc_var *ccv) 535 { 536 struct cubic *cubic_data; 537 int pipe; 538 539 cubic_data = ccv->cc_data; 540 pipe = 0; 541 542 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 543 /* 544 * If inflight data is less than ssthresh, set cwnd 545 * conservatively to avoid a burst of data, as suggested in 546 * the NewReno RFC. Otherwise, use the CUBIC method. 547 * 548 * XXXLAS: Find a way to do this without needing curack 549 */ 550 if (V_tcp_do_newsack) 551 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 552 else 553 pipe = CCV(ccv, snd_max) - ccv->curack; 554 555 if (pipe < CCV(ccv, snd_ssthresh)) 556 /* 557 * Ensure that cwnd does not collapse to 1 MSS under 558 * adverse conditions. Implements RFC6582 559 */ 560 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) + 561 CCV(ccv, t_maxseg); 562 else 563 /* Update cwnd based on beta and adjusted W_max. */ 564 CCV(ccv, snd_cwnd) = max(((uint64_t)cubic_data->W_max * 565 CUBIC_BETA) >> CUBIC_SHIFT, 566 2 * CCV(ccv, t_maxseg)); 567 } 568 569 /* Calculate the average RTT between congestion epochs. */ 570 if (cubic_data->epoch_ack_count > 0 && 571 cubic_data->sum_rtt_usecs >= cubic_data->epoch_ack_count) { 572 cubic_data->mean_rtt_usecs = (int)(cubic_data->sum_rtt_usecs / 573 cubic_data->epoch_ack_count); 574 } 575 576 cubic_data->epoch_ack_count = 0; 577 cubic_data->sum_rtt_usecs = 0; 578 } 579 580 /* 581 * Record the min RTT and sum samples for the epoch average RTT calculation. 582 */ 583 static void 584 cubic_record_rtt(struct cc_var *ccv) 585 { 586 struct cubic *cubic_data; 587 uint32_t t_srtt_usecs; 588 589 /* Ignore srtt until a min number of samples have been taken. */ 590 if (CCV(ccv, t_rttupdated) >= CUBIC_MIN_RTT_SAMPLES) { 591 cubic_data = ccv->cc_data; 592 t_srtt_usecs = tcp_get_srtt(ccv->ccvc.tcp, 593 TCP_TMR_GRANULARITY_USEC); 594 /* 595 * Record the current SRTT as our minrtt if it's the smallest 596 * we've seen or minrtt is currently equal to its initialised 597 * value. 598 * 599 * XXXLAS: Should there be some hysteresis for minrtt? 600 */ 601 if ((t_srtt_usecs < cubic_data->min_rtt_usecs || 602 cubic_data->min_rtt_usecs == TCPTV_SRTTBASE)) { 603 /* A minimal rtt is a single unshifted tick of a ticks 604 * timer. */ 605 cubic_data->min_rtt_usecs = max(tick >> TCP_RTT_SHIFT, 606 t_srtt_usecs); 607 608 /* 609 * If the connection is within its first congestion 610 * epoch, ensure we prime mean_rtt_usecs with a 611 * reasonable value until the epoch average RTT is 612 * calculated in cubic_post_recovery(). 613 */ 614 if (cubic_data->min_rtt_usecs > 615 cubic_data->mean_rtt_usecs) 616 cubic_data->mean_rtt_usecs = 617 cubic_data->min_rtt_usecs; 618 } 619 620 /* Sum samples for epoch average RTT calculation. */ 621 cubic_data->sum_rtt_usecs += t_srtt_usecs; 622 cubic_data->epoch_ack_count++; 623 } 624 } 625 626 /* 627 * Update the ssthresh in the event of congestion. 628 */ 629 static void 630 cubic_ssthresh_update(struct cc_var *ccv, uint32_t maxseg) 631 { 632 struct cubic *cubic_data; 633 uint32_t ssthresh; 634 uint32_t cwnd; 635 636 cubic_data = ccv->cc_data; 637 cwnd = CCV(ccv, snd_cwnd); 638 639 /* Fast convergence heuristic. */ 640 if (cwnd < cubic_data->W_max) { 641 cwnd = ((uint64_t)cwnd * CUBIC_FC_FACTOR) >> CUBIC_SHIFT; 642 } 643 cubic_data->undo_W_max = cubic_data->W_max; 644 cubic_data->W_max = cwnd; 645 646 /* 647 * On the first congestion event, set ssthresh to cwnd * 0.5 648 * and reduce W_max to cwnd * beta. This aligns the cubic concave 649 * region appropriately. On subsequent congestion events, set 650 * ssthresh to cwnd * beta. 651 */ 652 if ((cubic_data->flags & CUBICFLAG_CONG_EVENT) == 0) { 653 ssthresh = cwnd >> 1; 654 cubic_data->W_max = ((uint64_t)cwnd * 655 CUBIC_BETA) >> CUBIC_SHIFT; 656 } else { 657 ssthresh = ((uint64_t)cwnd * 658 CUBIC_BETA) >> CUBIC_SHIFT; 659 } 660 CCV(ccv, snd_ssthresh) = max(ssthresh, 2 * maxseg); 661 } 662 663 static void 664 cubic_rttsample(struct cc_var *ccv, uint32_t usec_rtt, uint32_t rxtcnt, uint32_t fas) 665 { 666 struct cubic *cubicd; 667 668 cubicd = ccv->cc_data; 669 if (rxtcnt > 1) { 670 /* 671 * Only look at RTT's that are non-ambiguous. 672 */ 673 return; 674 } 675 cubicd->css_rttsample_count++; 676 cubicd->css_last_fas = fas; 677 if (cubicd->css_current_round_minrtt > usec_rtt) { 678 cubicd->css_current_round_minrtt = usec_rtt; 679 cubicd->css_lowrtt_fas = cubicd->css_last_fas; 680 } 681 if ((cubicd->css_rttsample_count >= hystart_n_rttsamples) && 682 (cubicd->css_current_round_minrtt != 0xffffffff) && 683 (cubicd->css_current_round_minrtt < cubicd->css_baseline_minrtt) && 684 (cubicd->css_lastround_minrtt != 0xffffffff)) { 685 /* 686 * We were in CSS and the RTT is now less, we 687 * entered CSS erroneously. 688 */ 689 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 690 cubic_log_hystart_event(ccv, cubicd, 8, cubicd->css_baseline_minrtt); 691 cubicd->css_baseline_minrtt = 0xffffffff; 692 } 693 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 694 cubic_log_hystart_event(ccv, cubicd, 5, usec_rtt); 695 } 696 697 static void 698 cubic_newround(struct cc_var *ccv, uint32_t round_cnt) 699 { 700 struct cubic *cubicd; 701 702 cubicd = ccv->cc_data; 703 /* We have entered a new round */ 704 cubicd->css_lastround_minrtt = cubicd->css_current_round_minrtt; 705 cubicd->css_current_round_minrtt = 0xffffffff; 706 cubicd->css_rttsample_count = 0; 707 cubicd->css_current_round = round_cnt; 708 if ((cubicd->flags & CUBICFLAG_HYSTART_IN_CSS) && 709 ((round_cnt - cubicd->css_entered_at_round) >= hystart_css_rounds)) { 710 /* Enter CA */ 711 if (ccv->flags & CCF_HYSTART_CAN_SH_CWND) { 712 /* 713 * We engage more than snd_ssthresh, engage 714 * the brakes!! Though we will stay in SS to 715 * creep back up again, so lets leave CSS active 716 * and give us hystart_css_rounds more rounds. 717 */ 718 if (ccv->flags & CCF_HYSTART_CONS_SSTH) { 719 CCV(ccv, snd_ssthresh) = ((cubicd->css_lowrtt_fas + cubicd->css_fas_at_css_entry) / 2); 720 } else { 721 CCV(ccv, snd_ssthresh) = cubicd->css_lowrtt_fas; 722 } 723 CCV(ccv, snd_cwnd) = cubicd->css_fas_at_css_entry; 724 cubicd->css_entered_at_round = round_cnt; 725 } else { 726 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 727 /* Turn off the CSS flag */ 728 cubicd->flags &= ~CUBICFLAG_HYSTART_IN_CSS; 729 /* Disable use of CSS in the future except long idle */ 730 cubicd->flags &= ~CUBICFLAG_HYSTART_ENABLED; 731 } 732 cubic_log_hystart_event(ccv, cubicd, 6, CCV(ccv, snd_ssthresh)); 733 } 734 if (cubicd->flags & CUBICFLAG_HYSTART_ENABLED) 735 cubic_log_hystart_event(ccv, cubicd, 4, round_cnt); 736 } 737 738 DECLARE_CC_MODULE(cubic, &cubic_cc_algo); 739 MODULE_VERSION(cubic, 2); 740