1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2007-2008 5 * Swinburne University of Technology, Melbourne, Australia 6 * Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org> 7 * Copyright (c) 2010 The FreeBSD Foundation 8 * All rights reserved. 9 * 10 * This software was developed at the Centre for Advanced Internet 11 * Architectures, Swinburne University of Technology, by Lawrence Stewart and 12 * James Healy, made possible in part by a grant from the Cisco University 13 * Research Program Fund at Community Foundation Silicon Valley. 14 * 15 * Portions of this software were developed at the Centre for Advanced 16 * Internet Architectures, Swinburne University of Technology, Melbourne, 17 * Australia by David Hayes under sponsorship from the FreeBSD Foundation. 18 * 19 * Redistribution and use in source and binary forms, with or without 20 * modification, are permitted provided that the following conditions 21 * are met: 22 * 1. Redistributions of source code must retain the above copyright 23 * notice, this list of conditions and the following disclaimer. 24 * 2. Redistributions in binary form must reproduce the above copyright 25 * notice, this list of conditions and the following disclaimer in the 26 * documentation and/or other materials provided with the distribution. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 29 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 32 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38 * SUCH DAMAGE. 39 */ 40 41 /* 42 * An implementation of the H-TCP congestion control algorithm for FreeBSD, 43 * based on the Internet Draft "draft-leith-tcp-htcp-06.txt" by Leith and 44 * Shorten. Originally released as part of the NewTCP research project at 45 * Swinburne University of Technology's Centre for Advanced Internet 46 * Architectures, Melbourne, Australia, which was made possible in part by a 47 * grant from the Cisco University Research Program Fund at Community Foundation 48 * Silicon Valley. More details are available at: 49 * http://caia.swin.edu.au/urp/newtcp/ 50 */ 51 52 #include <sys/cdefs.h> 53 __FBSDID("$FreeBSD$"); 54 55 #include <sys/param.h> 56 #include <sys/kernel.h> 57 #include <sys/limits.h> 58 #include <sys/malloc.h> 59 #include <sys/module.h> 60 #include <sys/socket.h> 61 #include <sys/socketvar.h> 62 #include <sys/sysctl.h> 63 #include <sys/systm.h> 64 65 #include <net/vnet.h> 66 67 #include <net/route.h> 68 #include <net/route/nhop.h> 69 70 #include <netinet/in_pcb.h> 71 #include <netinet/tcp.h> 72 #include <netinet/tcp_seq.h> 73 #include <netinet/tcp_timer.h> 74 #include <netinet/tcp_var.h> 75 #include <netinet/cc/cc.h> 76 #include <netinet/cc/cc_module.h> 77 78 /* Fixed point math shifts. */ 79 #define HTCP_SHIFT 8 80 #define HTCP_ALPHA_INC_SHIFT 4 81 82 #define HTCP_INIT_ALPHA 1 83 #define HTCP_DELTA_L hz /* 1 sec in ticks. */ 84 #define HTCP_MINBETA 128 /* 0.5 << HTCP_SHIFT. */ 85 #define HTCP_MAXBETA 204 /* ~0.8 << HTCP_SHIFT. */ 86 #define HTCP_MINROWE 26 /* ~0.1 << HTCP_SHIFT. */ 87 #define HTCP_MAXROWE 512 /* 2 << HTCP_SHIFT. */ 88 89 /* RTT_ref (ms) used in the calculation of alpha if RTT scaling is enabled. */ 90 #define HTCP_RTT_REF 100 91 92 /* Don't trust SRTT until this many samples have been taken. */ 93 #define HTCP_MIN_RTT_SAMPLES 8 94 95 /* 96 * HTCP_CALC_ALPHA performs a fixed point math calculation to determine the 97 * value of alpha, based on the function defined in the HTCP spec. 98 * 99 * i.e. 1 + 10(delta - delta_l) + ((delta - delta_l) / 2) ^ 2 100 * 101 * "diff" is passed in to the macro as "delta - delta_l" and is expected to be 102 * in units of ticks. 103 * 104 * The joyousnous of fixed point maths means our function implementation looks a 105 * little funky... 106 * 107 * In order to maintain some precision in the calculations, a fixed point shift 108 * HTCP_ALPHA_INC_SHIFT is used to ensure the integer divisions don't 109 * truncate the results too badly. 110 * 111 * The "16" value is the "1" term in the alpha function shifted up by 112 * HTCP_ALPHA_INC_SHIFT 113 * 114 * The "160" value is the "10" multiplier in the alpha function multiplied by 115 * 2^HTCP_ALPHA_INC_SHIFT 116 * 117 * Specifying these as constants reduces the computations required. After 118 * up-shifting all the terms in the function and performing the required 119 * calculations, we down-shift the final result by HTCP_ALPHA_INC_SHIFT to 120 * ensure it is back in the correct range. 121 * 122 * The "hz" terms are required as kernels can be configured to run with 123 * different tick timers, which we have to adjust for in the alpha calculation 124 * (which originally was defined in terms of seconds). 125 * 126 * We also have to be careful to constrain the value of diff such that it won't 127 * overflow whilst performing the calculation. The middle term i.e. (160 * diff) 128 * / hz is the limiting factor in the calculation. We must constrain diff to be 129 * less than the max size of an int divided by the constant 160 figure 130 * i.e. diff < INT_MAX / 160 131 * 132 * NB: Changing HTCP_ALPHA_INC_SHIFT will require you to MANUALLY update the 133 * constants used in this function! 134 */ 135 #define HTCP_CALC_ALPHA(diff) \ 136 ((\ 137 (16) + \ 138 ((160 * (diff)) / hz) + \ 139 (((diff) / hz) * (((diff) << HTCP_ALPHA_INC_SHIFT) / (4 * hz))) \ 140 ) >> HTCP_ALPHA_INC_SHIFT) 141 142 static void htcp_ack_received(struct cc_var *ccv, uint16_t type); 143 static void htcp_cb_destroy(struct cc_var *ccv); 144 static int htcp_cb_init(struct cc_var *ccv, void *ptr); 145 static void htcp_cong_signal(struct cc_var *ccv, uint32_t type); 146 static int htcp_mod_init(void); 147 static void htcp_post_recovery(struct cc_var *ccv); 148 static void htcp_recalc_alpha(struct cc_var *ccv); 149 static void htcp_recalc_beta(struct cc_var *ccv); 150 static void htcp_record_rtt(struct cc_var *ccv); 151 static void htcp_ssthresh_update(struct cc_var *ccv); 152 static size_t htcp_data_sz(void); 153 154 struct htcp { 155 /* cwnd before entering cong recovery. */ 156 unsigned long prev_cwnd; 157 /* cwnd additive increase parameter. */ 158 int alpha; 159 /* cwnd multiplicative decrease parameter. */ 160 int beta; 161 /* Largest rtt seen for the flow. */ 162 int maxrtt; 163 /* Shortest rtt seen for the flow. */ 164 int minrtt; 165 /* Time of last congestion event in ticks. */ 166 int t_last_cong; 167 }; 168 169 static int htcp_rtt_ref; 170 /* 171 * The maximum number of ticks the value of diff can reach in 172 * htcp_recalc_alpha() before alpha will stop increasing due to overflow. 173 * See comment above HTCP_CALC_ALPHA for more info. 174 */ 175 static int htcp_max_diff = INT_MAX / ((1 << HTCP_ALPHA_INC_SHIFT) * 10); 176 177 /* Per-netstack vars. */ 178 VNET_DEFINE_STATIC(u_int, htcp_adaptive_backoff) = 0; 179 VNET_DEFINE_STATIC(u_int, htcp_rtt_scaling) = 0; 180 #define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff) 181 #define V_htcp_rtt_scaling VNET(htcp_rtt_scaling) 182 183 struct cc_algo htcp_cc_algo = { 184 .name = "htcp", 185 .ack_received = htcp_ack_received, 186 .cb_destroy = htcp_cb_destroy, 187 .cb_init = htcp_cb_init, 188 .cong_signal = htcp_cong_signal, 189 .mod_init = htcp_mod_init, 190 .post_recovery = htcp_post_recovery, 191 .cc_data_sz = htcp_data_sz, 192 .after_idle = newreno_cc_after_idle, 193 }; 194 195 static void 196 htcp_ack_received(struct cc_var *ccv, uint16_t type) 197 { 198 struct htcp *htcp_data; 199 200 htcp_data = ccv->cc_data; 201 htcp_record_rtt(ccv); 202 203 /* 204 * Regular ACK and we're not in cong/fast recovery and we're cwnd 205 * limited and we're either not doing ABC or are slow starting or are 206 * doing ABC and we've sent a cwnd's worth of bytes. 207 */ 208 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 209 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 || 210 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 211 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) { 212 htcp_recalc_beta(ccv); 213 htcp_recalc_alpha(ccv); 214 /* 215 * Use the logic in NewReno ack_received() for slow start and 216 * for the first HTCP_DELTA_L ticks after either the flow starts 217 * or a congestion event (when alpha equals 1). 218 */ 219 if (htcp_data->alpha == 1 || 220 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) 221 newreno_cc_ack_received(ccv, type); 222 else { 223 if (V_tcp_do_rfc3465) { 224 /* Increment cwnd by alpha segments. */ 225 CCV(ccv, snd_cwnd) += htcp_data->alpha * 226 CCV(ccv, t_maxseg); 227 ccv->flags &= ~CCF_ABC_SENTAWND; 228 } else 229 /* 230 * Increment cwnd by alpha/cwnd segments to 231 * approximate an increase of alpha segments 232 * per RTT. 233 */ 234 CCV(ccv, snd_cwnd) += (((htcp_data->alpha << 235 HTCP_SHIFT) / (CCV(ccv, snd_cwnd) / 236 CCV(ccv, t_maxseg))) * CCV(ccv, t_maxseg)) 237 >> HTCP_SHIFT; 238 } 239 } 240 } 241 242 static void 243 htcp_cb_destroy(struct cc_var *ccv) 244 { 245 free(ccv->cc_data, M_CC_MEM); 246 } 247 248 static size_t 249 htcp_data_sz(void) 250 { 251 return(sizeof(struct htcp)); 252 } 253 254 static int 255 htcp_cb_init(struct cc_var *ccv, void *ptr) 256 { 257 struct htcp *htcp_data; 258 259 INP_WLOCK_ASSERT(ccv->ccvc.tcp->t_inpcb); 260 if (ptr == NULL) { 261 htcp_data = malloc(sizeof(struct htcp), M_CC_MEM, M_NOWAIT); 262 if (htcp_data == NULL) 263 return (ENOMEM); 264 } else 265 htcp_data = ptr; 266 267 /* Init some key variables with sensible defaults. */ 268 htcp_data->alpha = HTCP_INIT_ALPHA; 269 htcp_data->beta = HTCP_MINBETA; 270 htcp_data->maxrtt = TCPTV_SRTTBASE; 271 htcp_data->minrtt = TCPTV_SRTTBASE; 272 htcp_data->prev_cwnd = 0; 273 htcp_data->t_last_cong = ticks; 274 275 ccv->cc_data = htcp_data; 276 277 return (0); 278 } 279 280 /* 281 * Perform any necessary tasks before we enter congestion recovery. 282 */ 283 static void 284 htcp_cong_signal(struct cc_var *ccv, uint32_t type) 285 { 286 struct htcp *htcp_data; 287 u_int mss; 288 289 htcp_data = ccv->cc_data; 290 mss = tcp_maxseg(ccv->ccvc.tcp); 291 292 switch (type) { 293 case CC_NDUPACK: 294 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 295 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 296 /* 297 * Apply hysteresis to maxrtt to ensure 298 * reductions in the RTT are reflected in our 299 * measurements. 300 */ 301 htcp_data->maxrtt = (htcp_data->minrtt + 302 (htcp_data->maxrtt - htcp_data->minrtt) * 303 95) / 100; 304 htcp_ssthresh_update(ccv); 305 htcp_data->t_last_cong = ticks; 306 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 307 } 308 ENTER_RECOVERY(CCV(ccv, t_flags)); 309 } 310 break; 311 312 case CC_ECN: 313 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 314 /* 315 * Apply hysteresis to maxrtt to ensure reductions in 316 * the RTT are reflected in our measurements. 317 */ 318 htcp_data->maxrtt = (htcp_data->minrtt + (htcp_data->maxrtt - 319 htcp_data->minrtt) * 95) / 100; 320 htcp_ssthresh_update(ccv); 321 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 322 htcp_data->t_last_cong = ticks; 323 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 324 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 325 } 326 break; 327 328 case CC_RTO: 329 CCV(ccv, snd_ssthresh) = max(min(CCV(ccv, snd_wnd), 330 CCV(ccv, snd_cwnd)) / 2 / mss, 331 2) * mss; 332 CCV(ccv, snd_cwnd) = mss; 333 /* 334 * Grab the current time and record it so we know when the 335 * most recent congestion event was. Only record it when the 336 * timeout has fired more than once, as there is a reasonable 337 * chance the first one is a false alarm and may not indicate 338 * congestion. 339 */ 340 if (CCV(ccv, t_rxtshift) >= 2) 341 htcp_data->t_last_cong = ticks; 342 break; 343 } 344 } 345 346 static int 347 htcp_mod_init(void) 348 { 349 /* 350 * HTCP_RTT_REF is defined in ms, and t_srtt in the tcpcb is stored in 351 * units of TCP_RTT_SCALE*hz. Scale HTCP_RTT_REF to be in the same units 352 * as t_srtt. 353 */ 354 htcp_rtt_ref = (HTCP_RTT_REF * TCP_RTT_SCALE * hz) / 1000; 355 return (0); 356 } 357 358 /* 359 * Perform any necessary tasks before we exit congestion recovery. 360 */ 361 static void 362 htcp_post_recovery(struct cc_var *ccv) 363 { 364 int pipe; 365 struct htcp *htcp_data; 366 367 pipe = 0; 368 htcp_data = ccv->cc_data; 369 370 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 371 /* 372 * If inflight data is less than ssthresh, set cwnd 373 * conservatively to avoid a burst of data, as suggested in the 374 * NewReno RFC. Otherwise, use the HTCP method. 375 * 376 * XXXLAS: Find a way to do this without needing curack 377 */ 378 if (V_tcp_do_newsack) 379 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 380 else 381 pipe = CCV(ccv, snd_max) - ccv->curack; 382 383 if (pipe < CCV(ccv, snd_ssthresh)) 384 /* 385 * Ensure that cwnd down not collape to 1 MSS under 386 * adverse conditions. Implements RFC6582 387 */ 388 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) + 389 CCV(ccv, t_maxseg); 390 else 391 CCV(ccv, snd_cwnd) = max(1, ((htcp_data->beta * 392 htcp_data->prev_cwnd / CCV(ccv, t_maxseg)) 393 >> HTCP_SHIFT)) * CCV(ccv, t_maxseg); 394 } 395 } 396 397 static void 398 htcp_recalc_alpha(struct cc_var *ccv) 399 { 400 struct htcp *htcp_data; 401 int alpha, diff, now; 402 403 htcp_data = ccv->cc_data; 404 now = ticks; 405 406 /* 407 * If ticks has wrapped around (will happen approximately once every 49 408 * days on a machine with the default kern.hz=1000) and a flow straddles 409 * the wrap point, our alpha calcs will be completely wrong. We cut our 410 * losses and restart alpha from scratch by setting t_last_cong = now - 411 * HTCP_DELTA_L. 412 * 413 * This does not deflate our cwnd at all. It simply slows the rate cwnd 414 * is growing by until alpha regains the value it held prior to taking 415 * this drastic measure. 416 */ 417 if (now < htcp_data->t_last_cong) 418 htcp_data->t_last_cong = now - HTCP_DELTA_L; 419 420 diff = now - htcp_data->t_last_cong - HTCP_DELTA_L; 421 422 /* Cap alpha if the value of diff would overflow HTCP_CALC_ALPHA(). */ 423 if (diff < htcp_max_diff) { 424 /* 425 * If it has been more than HTCP_DELTA_L ticks since congestion, 426 * increase alpha according to the function defined in the spec. 427 */ 428 if (diff > 0) { 429 alpha = HTCP_CALC_ALPHA(diff); 430 431 /* 432 * Adaptive backoff fairness adjustment: 433 * 2 * (1 - beta) * alpha_raw 434 */ 435 if (V_htcp_adaptive_backoff) 436 alpha = max(1, (2 * ((1 << HTCP_SHIFT) - 437 htcp_data->beta) * alpha) >> HTCP_SHIFT); 438 439 /* 440 * RTT scaling: (RTT / RTT_ref) * alpha 441 * alpha will be the raw value from HTCP_CALC_ALPHA() if 442 * adaptive backoff is off, or the adjusted value if 443 * adaptive backoff is on. 444 */ 445 if (V_htcp_rtt_scaling) 446 alpha = max(1, (min(max(HTCP_MINROWE, 447 (CCV(ccv, t_srtt) << HTCP_SHIFT) / 448 htcp_rtt_ref), HTCP_MAXROWE) * alpha) 449 >> HTCP_SHIFT); 450 451 } else 452 alpha = 1; 453 454 htcp_data->alpha = alpha; 455 } 456 } 457 458 static void 459 htcp_recalc_beta(struct cc_var *ccv) 460 { 461 struct htcp *htcp_data; 462 463 htcp_data = ccv->cc_data; 464 465 /* 466 * TCPTV_SRTTBASE is the initialised value of each connection's SRTT, so 467 * we only calc beta if the connection's SRTT has been changed from its 468 * initial value. beta is bounded to ensure it is always between 469 * HTCP_MINBETA and HTCP_MAXBETA. 470 */ 471 if (V_htcp_adaptive_backoff && htcp_data->minrtt != TCPTV_SRTTBASE && 472 htcp_data->maxrtt != TCPTV_SRTTBASE) 473 htcp_data->beta = min(max(HTCP_MINBETA, 474 (htcp_data->minrtt << HTCP_SHIFT) / htcp_data->maxrtt), 475 HTCP_MAXBETA); 476 else 477 htcp_data->beta = HTCP_MINBETA; 478 } 479 480 /* 481 * Record the minimum and maximum RTT seen for the connection. These are used in 482 * the calculation of beta if adaptive backoff is enabled. 483 */ 484 static void 485 htcp_record_rtt(struct cc_var *ccv) 486 { 487 struct htcp *htcp_data; 488 489 htcp_data = ccv->cc_data; 490 491 /* XXXLAS: Should there be some hysteresis for minrtt? */ 492 493 /* 494 * Record the current SRTT as our minrtt if it's the smallest we've seen 495 * or minrtt is currently equal to its initialised value. Ignore SRTT 496 * until a min number of samples have been taken. 497 */ 498 if ((CCV(ccv, t_srtt) < htcp_data->minrtt || 499 htcp_data->minrtt == TCPTV_SRTTBASE) && 500 (CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES)) 501 htcp_data->minrtt = CCV(ccv, t_srtt); 502 503 /* 504 * Record the current SRTT as our maxrtt if it's the largest we've 505 * seen. Ignore SRTT until a min number of samples have been taken. 506 */ 507 if (CCV(ccv, t_srtt) > htcp_data->maxrtt 508 && CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES) 509 htcp_data->maxrtt = CCV(ccv, t_srtt); 510 } 511 512 /* 513 * Update the ssthresh in the event of congestion. 514 */ 515 static void 516 htcp_ssthresh_update(struct cc_var *ccv) 517 { 518 struct htcp *htcp_data; 519 520 htcp_data = ccv->cc_data; 521 522 /* 523 * On the first congestion event, set ssthresh to cwnd * 0.5, on 524 * subsequent congestion events, set it to cwnd * beta. 525 */ 526 if (CCV(ccv, snd_ssthresh) == TCP_MAXWIN << TCP_MAX_WINSHIFT) 527 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 528 HTCP_MINBETA) >> HTCP_SHIFT; 529 else { 530 htcp_recalc_beta(ccv); 531 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 532 htcp_data->beta) >> HTCP_SHIFT; 533 } 534 } 535 536 SYSCTL_DECL(_net_inet_tcp_cc_htcp); 537 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 538 "H-TCP related settings"); 539 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff, 540 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_adaptive_backoff), 0, 541 "enable H-TCP adaptive backoff"); 542 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling, 543 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_rtt_scaling), 0, 544 "enable H-TCP RTT scaling"); 545 546 DECLARE_CC_MODULE(htcp, &htcp_cc_algo); 547 MODULE_VERSION(htcp, 2); 548