1 /* $NetBSD: tcp_congctl.c,v 1.23 2017/01/02 09:29:38 skrll Exp $ */ 2 3 /*- 4 * Copyright (c) 1997, 1998, 1999, 2001, 2005, 2006 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe and Kevin M. Lahey of the Numerical Aerospace Simulation 9 * Facility, NASA Ames Research Center. 10 * This code is derived from software contributed to The NetBSD Foundation 11 * by Charles M. Hannum. 12 * This code is derived from software contributed to The NetBSD Foundation 13 * by Rui Paulo. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 26 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 27 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 28 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 29 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 30 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 31 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 32 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 33 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 34 * POSSIBILITY OF SUCH DAMAGE. 35 */ 36 37 /* 38 * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. 39 * All rights reserved. 40 * 41 * Redistribution and use in source and binary forms, with or without 42 * modification, are permitted provided that the following conditions 43 * are met: 44 * 1. Redistributions of source code must retain the above copyright 45 * notice, this list of conditions and the following disclaimer. 46 * 2. Redistributions in binary form must reproduce the above copyright 47 * notice, this list of conditions and the following disclaimer in the 48 * documentation and/or other materials provided with the distribution. 49 * 3. Neither the name of the project nor the names of its contributors 50 * may be used to endorse or promote products derived from this software 51 * without specific prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND 54 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 55 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 56 * ARE DISCLAIMED. IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE 57 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 58 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 59 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 63 * SUCH DAMAGE. 64 */ 65 66 /* 67 * @(#)COPYRIGHT 1.1 (NRL) 17 January 1995 68 * 69 * NRL grants permission for redistribution and use in source and binary 70 * forms, with or without modification, of the software and documentation 71 * created at NRL provided that the following conditions are met: 72 * 73 * 1. Redistributions of source code must retain the above copyright 74 * notice, this list of conditions and the following disclaimer. 75 * 2. Redistributions in binary form must reproduce the above copyright 76 * notice, this list of conditions and the following disclaimer in the 77 * documentation and/or other materials provided with the distribution. 78 * 3. All advertising materials mentioning features or use of this software 79 * must display the following acknowledgements: 80 * This product includes software developed by the University of 81 * California, Berkeley and its contributors. 82 * This product includes software developed at the Information 83 * Technology Division, US Naval Research Laboratory. 84 * 4. Neither the name of the NRL nor the names of its contributors 85 * may be used to endorse or promote products derived from this software 86 * without specific prior written permission. 87 * 88 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 89 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 90 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 91 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 92 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 93 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 94 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 95 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 96 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 97 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 98 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 99 * 100 * The views and conclusions contained in the software and documentation 101 * are those of the authors and should not be interpreted as representing 102 * official policies, either expressed or implied, of the US Naval 103 * Research Laboratory (NRL). 104 */ 105 106 /* 107 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 108 * The Regents of the University of California. All rights reserved. 109 * 110 * Redistribution and use in source and binary forms, with or without 111 * modification, are permitted provided that the following conditions 112 * are met: 113 * 1. Redistributions of source code must retain the above copyright 114 * notice, this list of conditions and the following disclaimer. 115 * 2. Redistributions in binary form must reproduce the above copyright 116 * notice, this list of conditions and the following disclaimer in the 117 * documentation and/or other materials provided with the distribution. 118 * 3. Neither the name of the University nor the names of its contributors 119 * may be used to endorse or promote products derived from this software 120 * without specific prior written permission. 121 * 122 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 123 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 124 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 125 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 126 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 127 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 128 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 129 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 130 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 131 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 132 * SUCH DAMAGE. 133 * 134 * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95 135 */ 136 137 #include <sys/cdefs.h> 138 __KERNEL_RCSID(0, "$NetBSD: tcp_congctl.c,v 1.23 2017/01/02 09:29:38 skrll Exp $"); 139 140 #ifdef _KERNEL_OPT 141 #include "opt_inet.h" 142 #include "opt_tcp_debug.h" 143 #include "opt_tcp_congctl.h" 144 #endif 145 146 #include <sys/param.h> 147 #include <sys/systm.h> 148 #include <sys/malloc.h> 149 #include <sys/mbuf.h> 150 #include <sys/protosw.h> 151 #include <sys/socket.h> 152 #include <sys/socketvar.h> 153 #include <sys/errno.h> 154 #include <sys/syslog.h> 155 #include <sys/pool.h> 156 #include <sys/domain.h> 157 #include <sys/kernel.h> 158 #include <sys/mutex.h> 159 160 #include <net/if.h> 161 162 #include <netinet/in.h> 163 #include <netinet/in_systm.h> 164 #include <netinet/ip.h> 165 #include <netinet/in_pcb.h> 166 #include <netinet/in_var.h> 167 #include <netinet/ip_var.h> 168 169 #ifdef INET6 170 #ifndef INET 171 #include <netinet/in.h> 172 #endif 173 #include <netinet/ip6.h> 174 #include <netinet6/ip6_var.h> 175 #include <netinet6/in6_pcb.h> 176 #include <netinet6/ip6_var.h> 177 #include <netinet6/in6_var.h> 178 #include <netinet/icmp6.h> 179 #endif 180 181 #include <netinet/tcp.h> 182 #include <netinet/tcp_fsm.h> 183 #include <netinet/tcp_seq.h> 184 #include <netinet/tcp_timer.h> 185 #include <netinet/tcp_var.h> 186 #include <netinet/tcpip.h> 187 #include <netinet/tcp_congctl.h> 188 #ifdef TCP_DEBUG 189 #include <netinet/tcp_debug.h> 190 #endif 191 192 /* 193 * TODO: 194 * consider separating the actual implementations in another file. 195 */ 196 197 static void tcp_common_congestion_exp(struct tcpcb *, int, int); 198 199 static int tcp_reno_do_fast_retransmit(struct tcpcb *, const struct tcphdr *); 200 static int tcp_reno_fast_retransmit(struct tcpcb *, const struct tcphdr *); 201 static void tcp_reno_slow_retransmit(struct tcpcb *); 202 static void tcp_reno_fast_retransmit_newack(struct tcpcb *, 203 const struct tcphdr *); 204 static void tcp_reno_newack(struct tcpcb *, const struct tcphdr *); 205 static void tcp_reno_congestion_exp(struct tcpcb *tp); 206 207 static int tcp_newreno_fast_retransmit(struct tcpcb *, const struct tcphdr *); 208 static void tcp_newreno_fast_retransmit_newack(struct tcpcb *, 209 const struct tcphdr *); 210 static void tcp_newreno_newack(struct tcpcb *, const struct tcphdr *); 211 212 static int tcp_cubic_fast_retransmit(struct tcpcb *, const struct tcphdr *); 213 static void tcp_cubic_slow_retransmit(struct tcpcb *tp); 214 static void tcp_cubic_newack(struct tcpcb *, const struct tcphdr *); 215 static void tcp_cubic_congestion_exp(struct tcpcb *); 216 217 static void tcp_congctl_fillnames(void); 218 219 extern int tcprexmtthresh; 220 221 MALLOC_DEFINE(M_TCPCONGCTL, "tcpcongctl", "TCP congestion control structures"); 222 223 /* currently selected global congestion control */ 224 char tcp_congctl_global_name[TCPCC_MAXLEN]; 225 226 /* available global congestion control algorithms */ 227 char tcp_congctl_avail[10 * TCPCC_MAXLEN]; 228 229 /* 230 * Used to list the available congestion control algorithms. 231 */ 232 TAILQ_HEAD(, tcp_congctlent) tcp_congctlhd = 233 TAILQ_HEAD_INITIALIZER(tcp_congctlhd); 234 235 static struct tcp_congctlent * tcp_congctl_global; 236 237 static kmutex_t tcp_congctl_mtx; 238 239 void 240 tcp_congctl_init(void) 241 { 242 int r __diagused; 243 244 mutex_init(&tcp_congctl_mtx, MUTEX_DEFAULT, IPL_NONE); 245 246 /* Base algorithms. */ 247 r = tcp_congctl_register("reno", &tcp_reno_ctl); 248 KASSERT(r == 0); 249 r = tcp_congctl_register("newreno", &tcp_newreno_ctl); 250 KASSERT(r == 0); 251 r = tcp_congctl_register("cubic", &tcp_cubic_ctl); 252 KASSERT(r == 0); 253 254 /* NewReno is the default. */ 255 #ifndef TCP_CONGCTL_DEFAULT 256 #define TCP_CONGCTL_DEFAULT "newreno" 257 #endif 258 259 r = tcp_congctl_select(NULL, TCP_CONGCTL_DEFAULT); 260 KASSERT(r == 0); 261 } 262 263 /* 264 * Register a congestion algorithm and select it if we have none. 265 */ 266 int 267 tcp_congctl_register(const char *name, const struct tcp_congctl *tcc) 268 { 269 struct tcp_congctlent *ntcc, *tccp; 270 271 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) 272 if (!strcmp(name, tccp->congctl_name)) { 273 /* name already registered */ 274 return EEXIST; 275 } 276 277 ntcc = malloc(sizeof(*ntcc), M_TCPCONGCTL, M_WAITOK|M_ZERO); 278 279 strlcpy(ntcc->congctl_name, name, sizeof(ntcc->congctl_name) - 1); 280 ntcc->congctl_ctl = tcc; 281 282 TAILQ_INSERT_TAIL(&tcp_congctlhd, ntcc, congctl_ent); 283 tcp_congctl_fillnames(); 284 285 if (TAILQ_FIRST(&tcp_congctlhd) == ntcc) 286 tcp_congctl_select(NULL, name); 287 288 return 0; 289 } 290 291 int 292 tcp_congctl_unregister(const char *name) 293 { 294 struct tcp_congctlent *tccp, *rtccp; 295 unsigned int size; 296 297 rtccp = NULL; 298 size = 0; 299 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) { 300 if (!strcmp(name, tccp->congctl_name)) 301 rtccp = tccp; 302 size++; 303 } 304 305 if (!rtccp) 306 return ENOENT; 307 308 if (size <= 1 || tcp_congctl_global == rtccp || rtccp->congctl_refcnt) 309 return EBUSY; 310 311 TAILQ_REMOVE(&tcp_congctlhd, rtccp, congctl_ent); 312 free(rtccp, M_TCPCONGCTL); 313 tcp_congctl_fillnames(); 314 315 return 0; 316 } 317 318 /* 319 * Select a congestion algorithm by name. 320 */ 321 int 322 tcp_congctl_select(struct tcpcb *tp, const char *name) 323 { 324 struct tcp_congctlent *tccp, *old_tccp, *new_tccp; 325 bool old_found, new_found; 326 327 KASSERT(name); 328 329 old_found = (tp == NULL || tp->t_congctl == NULL); 330 old_tccp = NULL; 331 new_found = false; 332 new_tccp = NULL; 333 334 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) { 335 if (!old_found && tccp->congctl_ctl == tp->t_congctl) { 336 old_tccp = tccp; 337 old_found = true; 338 } 339 340 if (!new_found && !strcmp(name, tccp->congctl_name)) { 341 new_tccp = tccp; 342 new_found = true; 343 } 344 345 if (new_found && old_found) { 346 if (tp) { 347 mutex_enter(&tcp_congctl_mtx); 348 if (old_tccp) 349 old_tccp->congctl_refcnt--; 350 tp->t_congctl = new_tccp->congctl_ctl; 351 new_tccp->congctl_refcnt++; 352 mutex_exit(&tcp_congctl_mtx); 353 } else { 354 tcp_congctl_global = new_tccp; 355 strlcpy(tcp_congctl_global_name, 356 new_tccp->congctl_name, 357 sizeof(tcp_congctl_global_name) - 1); 358 } 359 return 0; 360 } 361 } 362 363 return EINVAL; 364 } 365 366 void 367 tcp_congctl_release(struct tcpcb *tp) 368 { 369 struct tcp_congctlent *tccp; 370 371 KASSERT(tp->t_congctl); 372 373 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) { 374 if (tccp->congctl_ctl == tp->t_congctl) { 375 tccp->congctl_refcnt--; 376 return; 377 } 378 } 379 } 380 381 /* 382 * Returns the name of a congestion algorithm. 383 */ 384 const char * 385 tcp_congctl_bystruct(const struct tcp_congctl *tcc) 386 { 387 struct tcp_congctlent *tccp; 388 389 KASSERT(tcc); 390 391 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) 392 if (tccp->congctl_ctl == tcc) 393 return tccp->congctl_name; 394 395 return NULL; 396 } 397 398 static void 399 tcp_congctl_fillnames(void) 400 { 401 struct tcp_congctlent *tccp; 402 const char *delim = " "; 403 404 tcp_congctl_avail[0] = '\0'; 405 TAILQ_FOREACH(tccp, &tcp_congctlhd, congctl_ent) { 406 strlcat(tcp_congctl_avail, tccp->congctl_name, 407 sizeof(tcp_congctl_avail) - 1); 408 if (TAILQ_NEXT(tccp, congctl_ent)) 409 strlcat(tcp_congctl_avail, delim, 410 sizeof(tcp_congctl_avail) - 1); 411 } 412 413 } 414 415 /* ------------------------------------------------------------------------ */ 416 417 /* 418 * Common stuff 419 */ 420 421 /* Window reduction (1-beta) for [New]Reno: 0.5 */ 422 #define RENO_BETAA 1 423 #define RENO_BETAB 2 424 /* Window reduction (1-beta) for Cubic: 0.8 */ 425 #define CUBIC_BETAA 4 426 #define CUBIC_BETAB 5 427 /* Draft Rhee Section 4.1 */ 428 #define CUBIC_CA 4 429 #define CUBIC_CB 10 430 431 static void 432 tcp_common_congestion_exp(struct tcpcb *tp, int betaa, int betab) 433 { 434 u_int win; 435 436 /* 437 * Reduce the congestion window and the slow start threshold. 438 */ 439 win = min(tp->snd_wnd, tp->snd_cwnd) * betaa / betab / tp->t_segsz; 440 if (win < 2) 441 win = 2; 442 443 tp->snd_ssthresh = win * tp->t_segsz; 444 tp->snd_recover = tp->snd_max; 445 tp->snd_cwnd = tp->snd_ssthresh; 446 447 /* 448 * When using TCP ECN, notify the peer that 449 * we reduced the cwnd. 450 */ 451 if (TCP_ECN_ALLOWED(tp)) 452 tp->t_flags |= TF_ECN_SND_CWR; 453 } 454 455 456 /* ------------------------------------------------------------------------ */ 457 458 /* 459 * TCP/Reno congestion control. 460 */ 461 static void 462 tcp_reno_congestion_exp(struct tcpcb *tp) 463 { 464 465 tcp_common_congestion_exp(tp, RENO_BETAA, RENO_BETAB); 466 } 467 468 static int 469 tcp_reno_do_fast_retransmit(struct tcpcb *tp, const struct tcphdr *th) 470 { 471 /* 472 * Dup acks mean that packets have left the 473 * network (they're now cached at the receiver) 474 * so bump cwnd by the amount in the receiver 475 * to keep a constant cwnd packets in the 476 * network. 477 * 478 * If we are using TCP/SACK, then enter 479 * Fast Recovery if the receiver SACKs 480 * data that is tcprexmtthresh * MSS 481 * bytes past the last ACKed segment, 482 * irrespective of the number of DupAcks. 483 */ 484 485 tcp_seq onxt = tp->snd_nxt; 486 487 tp->t_partialacks = 0; 488 TCP_TIMER_DISARM(tp, TCPT_REXMT); 489 tp->t_rtttime = 0; 490 if (TCP_SACK_ENABLED(tp)) { 491 tp->t_dupacks = tcprexmtthresh; 492 tp->sack_newdata = tp->snd_nxt; 493 tp->snd_cwnd = tp->t_segsz; 494 (void) tcp_output(tp); 495 return 0; 496 } 497 tp->snd_nxt = th->th_ack; 498 tp->snd_cwnd = tp->t_segsz; 499 (void) tcp_output(tp); 500 tp->snd_cwnd = tp->snd_ssthresh + tp->t_segsz * tp->t_dupacks; 501 if (SEQ_GT(onxt, tp->snd_nxt)) 502 tp->snd_nxt = onxt; 503 504 return 0; 505 } 506 507 static int 508 tcp_reno_fast_retransmit(struct tcpcb *tp, const struct tcphdr *th) 509 { 510 511 /* 512 * We know we're losing at the current 513 * window size so do congestion avoidance 514 * (set ssthresh to half the current window 515 * and pull our congestion window back to 516 * the new ssthresh). 517 */ 518 519 tcp_reno_congestion_exp(tp); 520 return tcp_reno_do_fast_retransmit(tp, th); 521 } 522 523 static void 524 tcp_reno_slow_retransmit(struct tcpcb *tp) 525 { 526 u_int win; 527 528 /* 529 * Close the congestion window down to one segment 530 * (we'll open it by one segment for each ack we get). 531 * Since we probably have a window's worth of unacked 532 * data accumulated, this "slow start" keeps us from 533 * dumping all that data as back-to-back packets (which 534 * might overwhelm an intermediate gateway). 535 * 536 * There are two phases to the opening: Initially we 537 * open by one mss on each ack. This makes the window 538 * size increase exponentially with time. If the 539 * window is larger than the path can handle, this 540 * exponential growth results in dropped packet(s) 541 * almost immediately. To get more time between 542 * drops but still "push" the network to take advantage 543 * of improving conditions, we switch from exponential 544 * to linear window opening at some threshhold size. 545 * For a threshhold, we use half the current window 546 * size, truncated to a multiple of the mss. 547 * 548 * (the minimum cwnd that will give us exponential 549 * growth is 2 mss. We don't allow the threshhold 550 * to go below this.) 551 */ 552 553 win = min(tp->snd_wnd, tp->snd_cwnd) / 2 / tp->t_segsz; 554 if (win < 2) 555 win = 2; 556 /* Loss Window MUST be one segment. */ 557 tp->snd_cwnd = tp->t_segsz; 558 tp->snd_ssthresh = win * tp->t_segsz; 559 tp->t_partialacks = -1; 560 tp->t_dupacks = 0; 561 tp->t_bytes_acked = 0; 562 563 if (TCP_ECN_ALLOWED(tp)) 564 tp->t_flags |= TF_ECN_SND_CWR; 565 } 566 567 static void 568 tcp_reno_fast_retransmit_newack(struct tcpcb *tp, 569 const struct tcphdr *th) 570 { 571 if (tp->t_partialacks < 0) { 572 /* 573 * We were not in fast recovery. Reset the duplicate ack 574 * counter. 575 */ 576 tp->t_dupacks = 0; 577 } else { 578 /* 579 * Clamp the congestion window to the crossover point and 580 * exit fast recovery. 581 */ 582 if (tp->snd_cwnd > tp->snd_ssthresh) 583 tp->snd_cwnd = tp->snd_ssthresh; 584 tp->t_partialacks = -1; 585 tp->t_dupacks = 0; 586 tp->t_bytes_acked = 0; 587 if (TCP_SACK_ENABLED(tp) && SEQ_GT(th->th_ack, tp->snd_fack)) 588 tp->snd_fack = th->th_ack; 589 } 590 } 591 592 static void 593 tcp_reno_newack(struct tcpcb *tp, const struct tcphdr *th) 594 { 595 /* 596 * When new data is acked, open the congestion window. 597 */ 598 599 u_int cw = tp->snd_cwnd; 600 u_int incr = tp->t_segsz; 601 602 if (tcp_do_abc) { 603 604 /* 605 * RFC 3465 Appropriate Byte Counting (ABC) 606 */ 607 608 int acked = th->th_ack - tp->snd_una; 609 610 if (cw >= tp->snd_ssthresh) { 611 tp->t_bytes_acked += acked; 612 if (tp->t_bytes_acked >= cw) { 613 /* Time to increase the window. */ 614 tp->t_bytes_acked -= cw; 615 } else { 616 /* No need to increase yet. */ 617 incr = 0; 618 } 619 } else { 620 /* 621 * use 2*SMSS or 1*SMSS for the "L" param, 622 * depending on sysctl setting. 623 * 624 * (See RFC 3465 2.3 Choosing the Limit) 625 */ 626 u_int abc_lim; 627 628 abc_lim = (tcp_abc_aggressive == 0 || 629 tp->snd_nxt != tp->snd_max) ? incr : incr * 2; 630 incr = min(acked, abc_lim); 631 } 632 } else { 633 634 /* 635 * If the window gives us less than ssthresh packets 636 * in flight, open exponentially (segsz per packet). 637 * Otherwise open linearly: segsz per window 638 * (segsz^2 / cwnd per packet). 639 */ 640 641 if (cw >= tp->snd_ssthresh) { 642 incr = incr * incr / cw; 643 } 644 } 645 646 tp->snd_cwnd = min(cw + incr, TCP_MAXWIN << tp->snd_scale); 647 } 648 649 const struct tcp_congctl tcp_reno_ctl = { 650 .fast_retransmit = tcp_reno_fast_retransmit, 651 .slow_retransmit = tcp_reno_slow_retransmit, 652 .fast_retransmit_newack = tcp_reno_fast_retransmit_newack, 653 .newack = tcp_reno_newack, 654 .cong_exp = tcp_reno_congestion_exp, 655 }; 656 657 /* 658 * TCP/NewReno Congestion control. 659 */ 660 static int 661 tcp_newreno_fast_retransmit(struct tcpcb *tp, const struct tcphdr *th) 662 { 663 664 if (SEQ_LT(th->th_ack, tp->snd_high)) { 665 /* 666 * False fast retransmit after timeout. 667 * Do not enter fast recovery 668 */ 669 tp->t_dupacks = 0; 670 return 1; 671 } 672 /* 673 * Fast retransmit is same as reno. 674 */ 675 return tcp_reno_fast_retransmit(tp, th); 676 } 677 678 /* 679 * Implement the NewReno response to a new ack, checking for partial acks in 680 * fast recovery. 681 */ 682 static void 683 tcp_newreno_fast_retransmit_newack(struct tcpcb *tp, const struct tcphdr *th) 684 { 685 if (tp->t_partialacks < 0) { 686 /* 687 * We were not in fast recovery. Reset the duplicate ack 688 * counter. 689 */ 690 tp->t_dupacks = 0; 691 } else if (SEQ_LT(th->th_ack, tp->snd_recover)) { 692 /* 693 * This is a partial ack. Retransmit the first unacknowledged 694 * segment and deflate the congestion window by the amount of 695 * acknowledged data. Do not exit fast recovery. 696 */ 697 tcp_seq onxt = tp->snd_nxt; 698 u_long ocwnd = tp->snd_cwnd; 699 int sack_num_segs = 1, sack_bytes_rxmt = 0; 700 701 /* 702 * snd_una has not yet been updated and the socket's send 703 * buffer has not yet drained off the ACK'd data, so we 704 * have to leave snd_una as it was to get the correct data 705 * offset in tcp_output(). 706 */ 707 tp->t_partialacks++; 708 TCP_TIMER_DISARM(tp, TCPT_REXMT); 709 tp->t_rtttime = 0; 710 711 if (TCP_SACK_ENABLED(tp)) { 712 /* 713 * Partial ack handling within a sack recovery episode. 714 * Keeping this very simple for now. When a partial ack 715 * is received, force snd_cwnd to a value that will 716 * allow the sender to transmit no more than 2 segments. 717 * If necessary, a fancier scheme can be adopted at a 718 * later point, but for now, the goal is to prevent the 719 * sender from bursting a large amount of data in the 720 * midst of sack recovery. 721 */ 722 723 /* 724 * send one or 2 segments based on how much 725 * new data was acked 726 */ 727 if (((th->th_ack - tp->snd_una) / tp->t_segsz) > 2) 728 sack_num_segs = 2; 729 (void)tcp_sack_output(tp, &sack_bytes_rxmt); 730 tp->snd_cwnd = sack_bytes_rxmt + 731 (tp->snd_nxt - tp->sack_newdata) + 732 sack_num_segs * tp->t_segsz; 733 tp->t_flags |= TF_ACKNOW; 734 (void) tcp_output(tp); 735 } else { 736 tp->snd_nxt = th->th_ack; 737 /* 738 * Set snd_cwnd to one segment beyond ACK'd offset 739 * snd_una is not yet updated when we're called 740 */ 741 tp->snd_cwnd = tp->t_segsz + (th->th_ack - tp->snd_una); 742 (void) tcp_output(tp); 743 tp->snd_cwnd = ocwnd; 744 if (SEQ_GT(onxt, tp->snd_nxt)) 745 tp->snd_nxt = onxt; 746 /* 747 * Partial window deflation. Relies on fact that 748 * tp->snd_una not updated yet. 749 */ 750 tp->snd_cwnd -= (th->th_ack - tp->snd_una - 751 tp->t_segsz); 752 } 753 } else { 754 /* 755 * Complete ack. Inflate the congestion window to ssthresh 756 * and exit fast recovery. 757 * 758 * Window inflation should have left us with approx. 759 * snd_ssthresh outstanding data. But in case we 760 * would be inclined to send a burst, better to do 761 * it via the slow start mechanism. 762 */ 763 if (SEQ_SUB(tp->snd_max, th->th_ack) < tp->snd_ssthresh) 764 tp->snd_cwnd = SEQ_SUB(tp->snd_max, th->th_ack) 765 + tp->t_segsz; 766 else 767 tp->snd_cwnd = tp->snd_ssthresh; 768 tp->t_partialacks = -1; 769 tp->t_dupacks = 0; 770 tp->t_bytes_acked = 0; 771 if (TCP_SACK_ENABLED(tp) && SEQ_GT(th->th_ack, tp->snd_fack)) 772 tp->snd_fack = th->th_ack; 773 } 774 } 775 776 static void 777 tcp_newreno_newack(struct tcpcb *tp, const struct tcphdr *th) 778 { 779 /* 780 * If we are still in fast recovery (meaning we are using 781 * NewReno and we have only received partial acks), do not 782 * inflate the window yet. 783 */ 784 if (tp->t_partialacks < 0) 785 tcp_reno_newack(tp, th); 786 } 787 788 789 const struct tcp_congctl tcp_newreno_ctl = { 790 .fast_retransmit = tcp_newreno_fast_retransmit, 791 .slow_retransmit = tcp_reno_slow_retransmit, 792 .fast_retransmit_newack = tcp_newreno_fast_retransmit_newack, 793 .newack = tcp_newreno_newack, 794 .cong_exp = tcp_reno_congestion_exp, 795 }; 796 797 /* 798 * CUBIC - http://tools.ietf.org/html/draft-rhee-tcpm-cubic-02 799 */ 800 801 /* Cubic prototypes */ 802 static void tcp_cubic_update_ctime(struct tcpcb *tp); 803 static uint32_t tcp_cubic_diff_ctime(struct tcpcb *); 804 static uint32_t tcp_cubic_cbrt(uint32_t); 805 static ulong tcp_cubic_getW(struct tcpcb *, uint32_t, uint32_t); 806 807 /* Cubic TIME functions - XXX I don't like using timevals and microuptime */ 808 /* 809 * Set congestion timer to now 810 */ 811 static void 812 tcp_cubic_update_ctime(struct tcpcb *tp) 813 { 814 struct timeval now_timeval; 815 816 getmicrouptime(&now_timeval); 817 tp->snd_cubic_ctime = now_timeval.tv_sec * 1000 + 818 now_timeval.tv_usec / 1000; 819 } 820 821 /* 822 * miliseconds from last congestion 823 */ 824 static uint32_t 825 tcp_cubic_diff_ctime(struct tcpcb *tp) 826 { 827 struct timeval now_timeval; 828 829 getmicrouptime(&now_timeval); 830 return now_timeval.tv_sec * 1000 + now_timeval.tv_usec / 1000 - 831 tp->snd_cubic_ctime; 832 } 833 834 /* 835 * Approximate cubic root 836 */ 837 #define CBRT_ROUNDS 30 838 static uint32_t 839 tcp_cubic_cbrt(uint32_t v) 840 { 841 int i, rounds = CBRT_ROUNDS; 842 uint64_t x = v / 3; 843 844 /* We fail to calculate correct for small numbers */ 845 if (v == 0) 846 return 0; 847 else if (v < 4) 848 return 1; 849 850 /* 851 * largest x that 2*x^3+3*x fits 64bit 852 * Avoid overflow for a time cost 853 */ 854 if (x > 2097151) 855 rounds += 10; 856 857 for (i = 0; i < rounds; i++) 858 if (rounds == CBRT_ROUNDS) 859 x = (v + 2 * x * x * x) / (3 * x * x); 860 else 861 /* Avoid overflow */ 862 x = v / (3 * x * x) + 2 * x / 3; 863 864 return (uint32_t)x; 865 } 866 867 /* Draft Rhee Section 3.1 - get W(t+rtt) - Eq. 1 */ 868 static ulong 869 tcp_cubic_getW(struct tcpcb *tp, uint32_t ms_elapsed, uint32_t rtt) 870 { 871 uint32_t K; 872 long tK3; 873 874 /* Section 3.1 Eq. 2 */ 875 K = tcp_cubic_cbrt(tp->snd_cubic_wmax / CUBIC_BETAB * 876 CUBIC_CB / CUBIC_CA); 877 /* (t-K)^3 - not clear why is the measure unit mattering */ 878 tK3 = (long)(ms_elapsed + rtt) - (long)K; 879 tK3 = tK3 * tK3 * tK3; 880 881 return CUBIC_CA * tK3 / CUBIC_CB + tp->snd_cubic_wmax; 882 } 883 884 static void 885 tcp_cubic_congestion_exp(struct tcpcb *tp) 886 { 887 888 /* 889 * Congestion - Set WMax and shrink cwnd 890 */ 891 tcp_cubic_update_ctime(tp); 892 893 /* Section 3.6 - Fast Convergence */ 894 if (tp->snd_cubic_wmax < tp->snd_cubic_wmax_last) { 895 tp->snd_cubic_wmax_last = tp->snd_cubic_wmax; 896 tp->snd_cubic_wmax = tp->snd_cubic_wmax / 2 + 897 tp->snd_cubic_wmax * CUBIC_BETAA / CUBIC_BETAB / 2; 898 } else { 899 tp->snd_cubic_wmax_last = tp->snd_cubic_wmax; 900 tp->snd_cubic_wmax = tp->snd_cwnd; 901 } 902 903 tp->snd_cubic_wmax = max(tp->t_segsz, tp->snd_cubic_wmax); 904 905 /* Shrink CWND */ 906 tcp_common_congestion_exp(tp, CUBIC_BETAA, CUBIC_BETAB); 907 } 908 909 static int 910 tcp_cubic_fast_retransmit(struct tcpcb *tp, const struct tcphdr *th) 911 { 912 913 if (SEQ_LT(th->th_ack, tp->snd_high)) { 914 /* See newreno */ 915 tp->t_dupacks = 0; 916 return 1; 917 } 918 919 /* 920 * mark WMax 921 */ 922 tcp_cubic_congestion_exp(tp); 923 924 /* Do fast retransmit */ 925 return tcp_reno_do_fast_retransmit(tp, th); 926 } 927 928 static void 929 tcp_cubic_newack(struct tcpcb *tp, const struct tcphdr *th) 930 { 931 uint32_t ms_elapsed, rtt; 932 u_long w_tcp; 933 934 /* Congestion avoidance and not in fast recovery and usable rtt */ 935 if (tp->snd_cwnd > tp->snd_ssthresh && tp->t_partialacks < 0 && 936 /* 937 * t_srtt is 1/32 units of slow ticks 938 * converting it in ms would be equal to 939 * (t_srtt >> 5) * 1000 / PR_SLOWHZ ~= (t_srtt << 5) / PR_SLOWHZ 940 */ 941 (rtt = (tp->t_srtt << 5) / PR_SLOWHZ) > 0) { 942 ms_elapsed = tcp_cubic_diff_ctime(tp); 943 944 /* Compute W_tcp(t) */ 945 w_tcp = tp->snd_cubic_wmax * CUBIC_BETAA / CUBIC_BETAB + 946 ms_elapsed / rtt / 3; 947 948 if (tp->snd_cwnd > w_tcp) { 949 /* Not in TCP friendly mode */ 950 tp->snd_cwnd += (tcp_cubic_getW(tp, ms_elapsed, rtt) - 951 tp->snd_cwnd) / tp->snd_cwnd; 952 } else { 953 /* friendly TCP mode */ 954 tp->snd_cwnd = w_tcp; 955 } 956 957 /* Make sure we are within limits */ 958 tp->snd_cwnd = max(tp->snd_cwnd, tp->t_segsz); 959 tp->snd_cwnd = min(tp->snd_cwnd, TCP_MAXWIN << tp->snd_scale); 960 } else { 961 /* Use New Reno */ 962 tcp_newreno_newack(tp, th); 963 } 964 } 965 966 static void 967 tcp_cubic_slow_retransmit(struct tcpcb *tp) 968 { 969 970 /* Timeout - Mark new congestion */ 971 tcp_cubic_congestion_exp(tp); 972 973 /* Loss Window MUST be one segment. */ 974 tp->snd_cwnd = tp->t_segsz; 975 tp->t_partialacks = -1; 976 tp->t_dupacks = 0; 977 tp->t_bytes_acked = 0; 978 979 if (TCP_ECN_ALLOWED(tp)) 980 tp->t_flags |= TF_ECN_SND_CWR; 981 } 982 983 const struct tcp_congctl tcp_cubic_ctl = { 984 .fast_retransmit = tcp_cubic_fast_retransmit, 985 .slow_retransmit = tcp_cubic_slow_retransmit, 986 .fast_retransmit_newack = tcp_newreno_fast_retransmit_newack, 987 .newack = tcp_cubic_newack, 988 .cong_exp = tcp_cubic_congestion_exp, 989 }; 990