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 <netinet/tcp.h> 68 #include <netinet/tcp_seq.h> 69 #include <netinet/tcp_timer.h> 70 #include <netinet/tcp_var.h> 71 #include <netinet/cc/cc.h> 72 #include <netinet/cc/cc_module.h> 73 74 /* Fixed point math shifts. */ 75 #define HTCP_SHIFT 8 76 #define HTCP_ALPHA_INC_SHIFT 4 77 78 #define HTCP_INIT_ALPHA 1 79 #define HTCP_DELTA_L hz /* 1 sec in ticks. */ 80 #define HTCP_MINBETA 128 /* 0.5 << HTCP_SHIFT. */ 81 #define HTCP_MAXBETA 204 /* ~0.8 << HTCP_SHIFT. */ 82 #define HTCP_MINROWE 26 /* ~0.1 << HTCP_SHIFT. */ 83 #define HTCP_MAXROWE 512 /* 2 << HTCP_SHIFT. */ 84 85 /* RTT_ref (ms) used in the calculation of alpha if RTT scaling is enabled. */ 86 #define HTCP_RTT_REF 100 87 88 /* Don't trust SRTT until this many samples have been taken. */ 89 #define HTCP_MIN_RTT_SAMPLES 8 90 91 /* 92 * HTCP_CALC_ALPHA performs a fixed point math calculation to determine the 93 * value of alpha, based on the function defined in the HTCP spec. 94 * 95 * i.e. 1 + 10(delta - delta_l) + ((delta - delta_l) / 2) ^ 2 96 * 97 * "diff" is passed in to the macro as "delta - delta_l" and is expected to be 98 * in units of ticks. 99 * 100 * The joyousnous of fixed point maths means our function implementation looks a 101 * little funky... 102 * 103 * In order to maintain some precision in the calculations, a fixed point shift 104 * HTCP_ALPHA_INC_SHIFT is used to ensure the integer divisions don't 105 * truncate the results too badly. 106 * 107 * The "16" value is the "1" term in the alpha function shifted up by 108 * HTCP_ALPHA_INC_SHIFT 109 * 110 * The "160" value is the "10" multiplier in the alpha function multiplied by 111 * 2^HTCP_ALPHA_INC_SHIFT 112 * 113 * Specifying these as constants reduces the computations required. After 114 * up-shifting all the terms in the function and performing the required 115 * calculations, we down-shift the final result by HTCP_ALPHA_INC_SHIFT to 116 * ensure it is back in the correct range. 117 * 118 * The "hz" terms are required as kernels can be configured to run with 119 * different tick timers, which we have to adjust for in the alpha calculation 120 * (which originally was defined in terms of seconds). 121 * 122 * We also have to be careful to constrain the value of diff such that it won't 123 * overflow whilst performing the calculation. The middle term i.e. (160 * diff) 124 * / hz is the limiting factor in the calculation. We must constrain diff to be 125 * less than the max size of an int divided by the constant 160 figure 126 * i.e. diff < INT_MAX / 160 127 * 128 * NB: Changing HTCP_ALPHA_INC_SHIFT will require you to MANUALLY update the 129 * constants used in this function! 130 */ 131 #define HTCP_CALC_ALPHA(diff) \ 132 ((\ 133 (16) + \ 134 ((160 * (diff)) / hz) + \ 135 (((diff) / hz) * (((diff) << HTCP_ALPHA_INC_SHIFT) / (4 * hz))) \ 136 ) >> HTCP_ALPHA_INC_SHIFT) 137 138 static void htcp_ack_received(struct cc_var *ccv, uint16_t type); 139 static void htcp_cb_destroy(struct cc_var *ccv); 140 static int htcp_cb_init(struct cc_var *ccv); 141 static void htcp_cong_signal(struct cc_var *ccv, uint32_t type); 142 static int htcp_mod_init(void); 143 static void htcp_post_recovery(struct cc_var *ccv); 144 static void htcp_recalc_alpha(struct cc_var *ccv); 145 static void htcp_recalc_beta(struct cc_var *ccv); 146 static void htcp_record_rtt(struct cc_var *ccv); 147 static void htcp_ssthresh_update(struct cc_var *ccv); 148 149 struct htcp { 150 /* cwnd before entering cong recovery. */ 151 unsigned long prev_cwnd; 152 /* cwnd additive increase parameter. */ 153 int alpha; 154 /* cwnd multiplicative decrease parameter. */ 155 int beta; 156 /* Largest rtt seen for the flow. */ 157 int maxrtt; 158 /* Shortest rtt seen for the flow. */ 159 int minrtt; 160 /* Time of last congestion event in ticks. */ 161 int t_last_cong; 162 }; 163 164 static int htcp_rtt_ref; 165 /* 166 * The maximum number of ticks the value of diff can reach in 167 * htcp_recalc_alpha() before alpha will stop increasing due to overflow. 168 * See comment above HTCP_CALC_ALPHA for more info. 169 */ 170 static int htcp_max_diff = INT_MAX / ((1 << HTCP_ALPHA_INC_SHIFT) * 10); 171 172 /* Per-netstack vars. */ 173 VNET_DEFINE_STATIC(u_int, htcp_adaptive_backoff) = 0; 174 VNET_DEFINE_STATIC(u_int, htcp_rtt_scaling) = 0; 175 #define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff) 176 #define V_htcp_rtt_scaling VNET(htcp_rtt_scaling) 177 178 static MALLOC_DEFINE(M_HTCP, "htcp data", 179 "Per connection data required for the HTCP congestion control algorithm"); 180 181 struct cc_algo htcp_cc_algo = { 182 .name = "htcp", 183 .ack_received = htcp_ack_received, 184 .cb_destroy = htcp_cb_destroy, 185 .cb_init = htcp_cb_init, 186 .cong_signal = htcp_cong_signal, 187 .mod_init = htcp_mod_init, 188 .post_recovery = htcp_post_recovery, 189 }; 190 191 static void 192 htcp_ack_received(struct cc_var *ccv, uint16_t type) 193 { 194 struct htcp *htcp_data; 195 196 htcp_data = ccv->cc_data; 197 htcp_record_rtt(ccv); 198 199 /* 200 * Regular ACK and we're not in cong/fast recovery and we're cwnd 201 * limited and we're either not doing ABC or are slow starting or are 202 * doing ABC and we've sent a cwnd's worth of bytes. 203 */ 204 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 205 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 || 206 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 207 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) { 208 htcp_recalc_beta(ccv); 209 htcp_recalc_alpha(ccv); 210 /* 211 * Use the logic in NewReno ack_received() for slow start and 212 * for the first HTCP_DELTA_L ticks after either the flow starts 213 * or a congestion event (when alpha equals 1). 214 */ 215 if (htcp_data->alpha == 1 || 216 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) 217 newreno_cc_algo.ack_received(ccv, type); 218 else { 219 if (V_tcp_do_rfc3465) { 220 /* Increment cwnd by alpha segments. */ 221 CCV(ccv, snd_cwnd) += htcp_data->alpha * 222 CCV(ccv, t_maxseg); 223 ccv->flags &= ~CCF_ABC_SENTAWND; 224 } else 225 /* 226 * Increment cwnd by alpha/cwnd segments to 227 * approximate an increase of alpha segments 228 * per RTT. 229 */ 230 CCV(ccv, snd_cwnd) += (((htcp_data->alpha << 231 HTCP_SHIFT) / (CCV(ccv, snd_cwnd) / 232 CCV(ccv, t_maxseg))) * CCV(ccv, t_maxseg)) 233 >> HTCP_SHIFT; 234 } 235 } 236 } 237 238 static void 239 htcp_cb_destroy(struct cc_var *ccv) 240 { 241 free(ccv->cc_data, M_HTCP); 242 } 243 244 static int 245 htcp_cb_init(struct cc_var *ccv) 246 { 247 struct htcp *htcp_data; 248 249 htcp_data = malloc(sizeof(struct htcp), M_HTCP, M_NOWAIT); 250 251 if (htcp_data == NULL) 252 return (ENOMEM); 253 254 /* Init some key variables with sensible defaults. */ 255 htcp_data->alpha = HTCP_INIT_ALPHA; 256 htcp_data->beta = HTCP_MINBETA; 257 htcp_data->maxrtt = TCPTV_SRTTBASE; 258 htcp_data->minrtt = TCPTV_SRTTBASE; 259 htcp_data->prev_cwnd = 0; 260 htcp_data->t_last_cong = ticks; 261 262 ccv->cc_data = htcp_data; 263 264 return (0); 265 } 266 267 /* 268 * Perform any necessary tasks before we enter congestion recovery. 269 */ 270 static void 271 htcp_cong_signal(struct cc_var *ccv, uint32_t type) 272 { 273 struct htcp *htcp_data; 274 u_int mss; 275 276 htcp_data = ccv->cc_data; 277 mss = tcp_maxseg(ccv->ccvc.tcp); 278 279 switch (type) { 280 case CC_NDUPACK: 281 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 282 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 283 /* 284 * Apply hysteresis to maxrtt to ensure 285 * reductions in the RTT are reflected in our 286 * measurements. 287 */ 288 htcp_data->maxrtt = (htcp_data->minrtt + 289 (htcp_data->maxrtt - htcp_data->minrtt) * 290 95) / 100; 291 htcp_ssthresh_update(ccv); 292 htcp_data->t_last_cong = ticks; 293 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 294 } 295 ENTER_RECOVERY(CCV(ccv, t_flags)); 296 } 297 break; 298 299 case CC_ECN: 300 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 301 /* 302 * Apply hysteresis to maxrtt to ensure reductions in 303 * the RTT are reflected in our measurements. 304 */ 305 htcp_data->maxrtt = (htcp_data->minrtt + (htcp_data->maxrtt - 306 htcp_data->minrtt) * 95) / 100; 307 htcp_ssthresh_update(ccv); 308 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 309 htcp_data->t_last_cong = ticks; 310 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 311 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 312 } 313 break; 314 315 case CC_RTO: 316 CCV(ccv, snd_ssthresh) = max(min(CCV(ccv, snd_wnd), 317 CCV(ccv, snd_cwnd)) / 2 / mss, 318 2) * mss; 319 CCV(ccv, snd_cwnd) = mss; 320 /* 321 * Grab the current time and record it so we know when the 322 * most recent congestion event was. Only record it when the 323 * timeout has fired more than once, as there is a reasonable 324 * chance the first one is a false alarm and may not indicate 325 * congestion. 326 */ 327 if (CCV(ccv, t_rxtshift) >= 2) 328 htcp_data->t_last_cong = ticks; 329 break; 330 } 331 } 332 333 static int 334 htcp_mod_init(void) 335 { 336 337 htcp_cc_algo.after_idle = newreno_cc_algo.after_idle; 338 339 /* 340 * HTCP_RTT_REF is defined in ms, and t_srtt in the tcpcb is stored in 341 * units of TCP_RTT_SCALE*hz. Scale HTCP_RTT_REF to be in the same units 342 * as t_srtt. 343 */ 344 htcp_rtt_ref = (HTCP_RTT_REF * TCP_RTT_SCALE * hz) / 1000; 345 346 return (0); 347 } 348 349 /* 350 * Perform any necessary tasks before we exit congestion recovery. 351 */ 352 static void 353 htcp_post_recovery(struct cc_var *ccv) 354 { 355 int pipe; 356 struct htcp *htcp_data; 357 358 pipe = 0; 359 htcp_data = ccv->cc_data; 360 361 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 362 /* 363 * If inflight data is less than ssthresh, set cwnd 364 * conservatively to avoid a burst of data, as suggested in the 365 * NewReno RFC. Otherwise, use the HTCP method. 366 * 367 * XXXLAS: Find a way to do this without needing curack 368 */ 369 if (V_tcp_do_newsack) 370 pipe = tcp_compute_pipe(ccv->ccvc.tcp); 371 else 372 pipe = CCV(ccv, snd_max) - ccv->curack; 373 374 if (pipe < CCV(ccv, snd_ssthresh)) 375 /* 376 * Ensure that cwnd down not collape to 1 MSS under 377 * adverse conditions. Implements RFC6582 378 */ 379 CCV(ccv, snd_cwnd) = max(pipe, CCV(ccv, t_maxseg)) + 380 CCV(ccv, t_maxseg); 381 else 382 CCV(ccv, snd_cwnd) = max(1, ((htcp_data->beta * 383 htcp_data->prev_cwnd / CCV(ccv, t_maxseg)) 384 >> HTCP_SHIFT)) * CCV(ccv, t_maxseg); 385 } 386 } 387 388 static void 389 htcp_recalc_alpha(struct cc_var *ccv) 390 { 391 struct htcp *htcp_data; 392 int alpha, diff, now; 393 394 htcp_data = ccv->cc_data; 395 now = ticks; 396 397 /* 398 * If ticks has wrapped around (will happen approximately once every 49 399 * days on a machine with the default kern.hz=1000) and a flow straddles 400 * the wrap point, our alpha calcs will be completely wrong. We cut our 401 * losses and restart alpha from scratch by setting t_last_cong = now - 402 * HTCP_DELTA_L. 403 * 404 * This does not deflate our cwnd at all. It simply slows the rate cwnd 405 * is growing by until alpha regains the value it held prior to taking 406 * this drastic measure. 407 */ 408 if (now < htcp_data->t_last_cong) 409 htcp_data->t_last_cong = now - HTCP_DELTA_L; 410 411 diff = now - htcp_data->t_last_cong - HTCP_DELTA_L; 412 413 /* Cap alpha if the value of diff would overflow HTCP_CALC_ALPHA(). */ 414 if (diff < htcp_max_diff) { 415 /* 416 * If it has been more than HTCP_DELTA_L ticks since congestion, 417 * increase alpha according to the function defined in the spec. 418 */ 419 if (diff > 0) { 420 alpha = HTCP_CALC_ALPHA(diff); 421 422 /* 423 * Adaptive backoff fairness adjustment: 424 * 2 * (1 - beta) * alpha_raw 425 */ 426 if (V_htcp_adaptive_backoff) 427 alpha = max(1, (2 * ((1 << HTCP_SHIFT) - 428 htcp_data->beta) * alpha) >> HTCP_SHIFT); 429 430 /* 431 * RTT scaling: (RTT / RTT_ref) * alpha 432 * alpha will be the raw value from HTCP_CALC_ALPHA() if 433 * adaptive backoff is off, or the adjusted value if 434 * adaptive backoff is on. 435 */ 436 if (V_htcp_rtt_scaling) 437 alpha = max(1, (min(max(HTCP_MINROWE, 438 (CCV(ccv, t_srtt) << HTCP_SHIFT) / 439 htcp_rtt_ref), HTCP_MAXROWE) * alpha) 440 >> HTCP_SHIFT); 441 442 } else 443 alpha = 1; 444 445 htcp_data->alpha = alpha; 446 } 447 } 448 449 static void 450 htcp_recalc_beta(struct cc_var *ccv) 451 { 452 struct htcp *htcp_data; 453 454 htcp_data = ccv->cc_data; 455 456 /* 457 * TCPTV_SRTTBASE is the initialised value of each connection's SRTT, so 458 * we only calc beta if the connection's SRTT has been changed from its 459 * initial value. beta is bounded to ensure it is always between 460 * HTCP_MINBETA and HTCP_MAXBETA. 461 */ 462 if (V_htcp_adaptive_backoff && htcp_data->minrtt != TCPTV_SRTTBASE && 463 htcp_data->maxrtt != TCPTV_SRTTBASE) 464 htcp_data->beta = min(max(HTCP_MINBETA, 465 (htcp_data->minrtt << HTCP_SHIFT) / htcp_data->maxrtt), 466 HTCP_MAXBETA); 467 else 468 htcp_data->beta = HTCP_MINBETA; 469 } 470 471 /* 472 * Record the minimum and maximum RTT seen for the connection. These are used in 473 * the calculation of beta if adaptive backoff is enabled. 474 */ 475 static void 476 htcp_record_rtt(struct cc_var *ccv) 477 { 478 struct htcp *htcp_data; 479 480 htcp_data = ccv->cc_data; 481 482 /* XXXLAS: Should there be some hysteresis for minrtt? */ 483 484 /* 485 * Record the current SRTT as our minrtt if it's the smallest we've seen 486 * or minrtt is currently equal to its initialised value. Ignore SRTT 487 * until a min number of samples have been taken. 488 */ 489 if ((CCV(ccv, t_srtt) < htcp_data->minrtt || 490 htcp_data->minrtt == TCPTV_SRTTBASE) && 491 (CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES)) 492 htcp_data->minrtt = CCV(ccv, t_srtt); 493 494 /* 495 * Record the current SRTT as our maxrtt if it's the largest we've 496 * seen. Ignore SRTT until a min number of samples have been taken. 497 */ 498 if (CCV(ccv, t_srtt) > htcp_data->maxrtt 499 && CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES) 500 htcp_data->maxrtt = CCV(ccv, t_srtt); 501 } 502 503 /* 504 * Update the ssthresh in the event of congestion. 505 */ 506 static void 507 htcp_ssthresh_update(struct cc_var *ccv) 508 { 509 struct htcp *htcp_data; 510 511 htcp_data = ccv->cc_data; 512 513 /* 514 * On the first congestion event, set ssthresh to cwnd * 0.5, on 515 * subsequent congestion events, set it to cwnd * beta. 516 */ 517 if (CCV(ccv, snd_ssthresh) == TCP_MAXWIN << TCP_MAX_WINSHIFT) 518 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 519 HTCP_MINBETA) >> HTCP_SHIFT; 520 else { 521 htcp_recalc_beta(ccv); 522 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 523 htcp_data->beta) >> HTCP_SHIFT; 524 } 525 } 526 527 SYSCTL_DECL(_net_inet_tcp_cc_htcp); 528 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 529 "H-TCP related settings"); 530 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff, 531 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_adaptive_backoff), 0, 532 "enable H-TCP adaptive backoff"); 533 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling, 534 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_rtt_scaling), 0, 535 "enable H-TCP RTT scaling"); 536 537 DECLARE_CC_MODULE(htcp, &htcp_cc_algo); 538 MODULE_VERSION(htcp, 1); 539