1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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/param.h> 53 #include <sys/kernel.h> 54 #include <sys/limits.h> 55 #include <sys/malloc.h> 56 #include <sys/module.h> 57 #include <sys/socket.h> 58 #include <sys/socketvar.h> 59 #include <sys/sysctl.h> 60 #include <sys/systm.h> 61 62 #include <net/vnet.h> 63 64 #include <net/route.h> 65 #include <net/route/nhop.h> 66 67 #include <netinet/in_pcb.h> 68 #include <netinet/tcp.h> 69 #include <netinet/tcp_seq.h> 70 #include <netinet/tcp_timer.h> 71 #include <netinet/tcp_var.h> 72 #include <netinet/cc/cc.h> 73 #include <netinet/cc/cc_module.h> 74 75 /* Fixed point math shifts. */ 76 #define HTCP_SHIFT 8 77 #define HTCP_ALPHA_INC_SHIFT 4 78 79 #define HTCP_INIT_ALPHA 1 80 #define HTCP_DELTA_L hz /* 1 sec in ticks. */ 81 #define HTCP_MINBETA 128 /* 0.5 << HTCP_SHIFT. */ 82 #define HTCP_MAXBETA 204 /* ~0.8 << HTCP_SHIFT. */ 83 #define HTCP_MINROWE 26 /* ~0.1 << HTCP_SHIFT. */ 84 #define HTCP_MAXROWE 512 /* 2 << HTCP_SHIFT. */ 85 86 /* RTT_ref (ms) used in the calculation of alpha if RTT scaling is enabled. */ 87 #define HTCP_RTT_REF 100 88 89 /* Don't trust SRTT until this many samples have been taken. */ 90 #define HTCP_MIN_RTT_SAMPLES 8 91 92 /* 93 * HTCP_CALC_ALPHA performs a fixed point math calculation to determine the 94 * value of alpha, based on the function defined in the HTCP spec. 95 * 96 * i.e. 1 + 10(delta - delta_l) + ((delta - delta_l) / 2) ^ 2 97 * 98 * "diff" is passed in to the macro as "delta - delta_l" and is expected to be 99 * in units of ticks. 100 * 101 * The joyousnous of fixed point maths means our function implementation looks a 102 * little funky... 103 * 104 * In order to maintain some precision in the calculations, a fixed point shift 105 * HTCP_ALPHA_INC_SHIFT is used to ensure the integer divisions don't 106 * truncate the results too badly. 107 * 108 * The "16" value is the "1" term in the alpha function shifted up by 109 * HTCP_ALPHA_INC_SHIFT 110 * 111 * The "160" value is the "10" multiplier in the alpha function multiplied by 112 * 2^HTCP_ALPHA_INC_SHIFT 113 * 114 * Specifying these as constants reduces the computations required. After 115 * up-shifting all the terms in the function and performing the required 116 * calculations, we down-shift the final result by HTCP_ALPHA_INC_SHIFT to 117 * ensure it is back in the correct range. 118 * 119 * The "hz" terms are required as kernels can be configured to run with 120 * different tick timers, which we have to adjust for in the alpha calculation 121 * (which originally was defined in terms of seconds). 122 * 123 * We also have to be careful to constrain the value of diff such that it won't 124 * overflow whilst performing the calculation. The middle term i.e. (160 * diff) 125 * / hz is the limiting factor in the calculation. We must constrain diff to be 126 * less than the max size of an int divided by the constant 160 figure 127 * i.e. diff < INT_MAX / 160 128 * 129 * NB: Changing HTCP_ALPHA_INC_SHIFT will require you to MANUALLY update the 130 * constants used in this function! 131 */ 132 #define HTCP_CALC_ALPHA(diff) \ 133 ((\ 134 (16) + \ 135 ((160 * (diff)) / hz) + \ 136 (((diff) / hz) * (((diff) << HTCP_ALPHA_INC_SHIFT) / (4 * hz))) \ 137 ) >> HTCP_ALPHA_INC_SHIFT) 138 139 static void htcp_ack_received(struct cc_var *ccv, ccsignal_t type); 140 static void htcp_cb_destroy(struct cc_var *ccv); 141 static int htcp_cb_init(struct cc_var *ccv, void *ptr); 142 static void htcp_cong_signal(struct cc_var *ccv, ccsignal_t type); 143 static int htcp_mod_init(void); 144 static void htcp_post_recovery(struct cc_var *ccv); 145 static void htcp_recalc_alpha(struct cc_var *ccv); 146 static void htcp_recalc_beta(struct cc_var *ccv); 147 static void htcp_record_rtt(struct cc_var *ccv); 148 static void htcp_ssthresh_update(struct cc_var *ccv); 149 static size_t htcp_data_sz(void); 150 151 struct htcp { 152 /* cwnd before entering cong recovery. */ 153 unsigned long prev_cwnd; 154 /* cwnd additive increase parameter. */ 155 int alpha; 156 /* cwnd multiplicative decrease parameter. */ 157 int beta; 158 /* Largest rtt seen for the flow. */ 159 int maxrtt; 160 /* Shortest rtt seen for the flow. */ 161 int minrtt; 162 /* Time of last congestion event in ticks. */ 163 int t_last_cong; 164 }; 165 166 static int htcp_rtt_ref; 167 /* 168 * The maximum number of ticks the value of diff can reach in 169 * htcp_recalc_alpha() before alpha will stop increasing due to overflow. 170 * See comment above HTCP_CALC_ALPHA for more info. 171 */ 172 static int htcp_max_diff = INT_MAX / ((1 << HTCP_ALPHA_INC_SHIFT) * 10); 173 174 /* Per-netstack vars. */ 175 VNET_DEFINE_STATIC(u_int, htcp_adaptive_backoff) = 0; 176 VNET_DEFINE_STATIC(u_int, htcp_rtt_scaling) = 0; 177 #define V_htcp_adaptive_backoff VNET(htcp_adaptive_backoff) 178 #define V_htcp_rtt_scaling VNET(htcp_rtt_scaling) 179 180 struct cc_algo htcp_cc_algo = { 181 .name = "htcp", 182 .ack_received = htcp_ack_received, 183 .cb_destroy = htcp_cb_destroy, 184 .cb_init = htcp_cb_init, 185 .cong_signal = htcp_cong_signal, 186 .mod_init = htcp_mod_init, 187 .post_recovery = htcp_post_recovery, 188 .cc_data_sz = htcp_data_sz, 189 .after_idle = newreno_cc_after_idle, 190 }; 191 192 static void 193 htcp_ack_received(struct cc_var *ccv, ccsignal_t type) 194 { 195 struct htcp *htcp_data; 196 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 197 198 htcp_data = ccv->cc_data; 199 htcp_record_rtt(ccv); 200 201 /* 202 * Regular ACK and we're not in cong/fast recovery and we're cwnd 203 * limited and we're either not doing ABC or are slow starting or are 204 * doing ABC and we've sent a cwnd's worth of bytes. 205 */ 206 if (type == CC_ACK && !IN_RECOVERY(CCV(ccv, t_flags)) && 207 (ccv->flags & CCF_CWND_LIMITED) && (!V_tcp_do_rfc3465 || 208 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh) || 209 (V_tcp_do_rfc3465 && ccv->flags & CCF_ABC_SENTAWND))) { 210 htcp_recalc_beta(ccv); 211 htcp_recalc_alpha(ccv); 212 /* 213 * Use the logic in NewReno ack_received() for slow start and 214 * for the first HTCP_DELTA_L ticks after either the flow starts 215 * or a congestion event (when alpha equals 1). 216 */ 217 if (htcp_data->alpha == 1 || 218 CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) 219 newreno_cc_ack_received(ccv, type); 220 else { 221 if (V_tcp_do_rfc3465) { 222 /* Increment cwnd by alpha segments. */ 223 CCV(ccv, snd_cwnd) += htcp_data->alpha * 224 mss; 225 ccv->flags &= ~CCF_ABC_SENTAWND; 226 } else 227 /* 228 * Increment cwnd by alpha/cwnd segments to 229 * approximate an increase of alpha segments 230 * per RTT. 231 */ 232 CCV(ccv, snd_cwnd) += (((htcp_data->alpha << 233 HTCP_SHIFT) / (max(1, 234 CCV(ccv, snd_cwnd) / mss))) * 235 mss) >> HTCP_SHIFT; 236 } 237 } 238 } 239 240 static void 241 htcp_cb_destroy(struct cc_var *ccv) 242 { 243 free(ccv->cc_data, M_CC_MEM); 244 } 245 246 static size_t 247 htcp_data_sz(void) 248 { 249 return(sizeof(struct htcp)); 250 } 251 252 static int 253 htcp_cb_init(struct cc_var *ccv, void *ptr) 254 { 255 struct htcp *htcp_data; 256 257 INP_WLOCK_ASSERT(tptoinpcb(ccv->tp)); 258 if (ptr == NULL) { 259 htcp_data = malloc(sizeof(struct htcp), M_CC_MEM, M_NOWAIT); 260 if (htcp_data == NULL) 261 return (ENOMEM); 262 } else 263 htcp_data = ptr; 264 265 /* Init some key variables with sensible defaults. */ 266 htcp_data->alpha = HTCP_INIT_ALPHA; 267 htcp_data->beta = HTCP_MINBETA; 268 htcp_data->maxrtt = TCPTV_SRTTBASE; 269 htcp_data->minrtt = TCPTV_SRTTBASE; 270 htcp_data->prev_cwnd = 0; 271 htcp_data->t_last_cong = ticks; 272 273 ccv->cc_data = htcp_data; 274 275 return (0); 276 } 277 278 /* 279 * Perform any necessary tasks before we enter congestion recovery. 280 */ 281 static void 282 htcp_cong_signal(struct cc_var *ccv, ccsignal_t type) 283 { 284 struct htcp *htcp_data; 285 uint32_t mss, pipe; 286 287 htcp_data = ccv->cc_data; 288 mss = tcp_fixed_maxseg(ccv->tp); 289 290 switch (type) { 291 case CC_NDUPACK: 292 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 293 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 294 /* 295 * Apply hysteresis to maxrtt to ensure 296 * reductions in the RTT are reflected in our 297 * measurements. 298 */ 299 htcp_data->maxrtt = (htcp_data->minrtt + 300 (htcp_data->maxrtt - htcp_data->minrtt) * 301 95) / 100; 302 htcp_ssthresh_update(ccv); 303 htcp_data->t_last_cong = ticks; 304 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 305 } 306 ENTER_RECOVERY(CCV(ccv, t_flags)); 307 } 308 break; 309 310 case CC_ECN: 311 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 312 /* 313 * Apply hysteresis to maxrtt to ensure reductions in 314 * the RTT are reflected in our measurements. 315 */ 316 htcp_data->maxrtt = (htcp_data->minrtt + (htcp_data->maxrtt - 317 htcp_data->minrtt) * 95) / 100; 318 htcp_ssthresh_update(ccv); 319 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 320 htcp_data->t_last_cong = ticks; 321 htcp_data->prev_cwnd = CCV(ccv, snd_cwnd); 322 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 323 } 324 break; 325 326 case CC_RTO: 327 if (CCV(ccv, t_rxtshift) == 1) { 328 pipe = tcp_compute_pipe(ccv->tp); 329 CCV(ccv, snd_ssthresh) = max(2, 330 min(CCV(ccv, snd_wnd), pipe) / 2 / mss) * mss; 331 } 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 default: 344 break; 345 } 346 } 347 348 static int 349 htcp_mod_init(void) 350 { 351 /* 352 * HTCP_RTT_REF is defined in ms, and t_srtt in the tcpcb is stored in 353 * units of TCP_RTT_SCALE*hz. Scale HTCP_RTT_REF to be in the same units 354 * as t_srtt. 355 */ 356 htcp_rtt_ref = (HTCP_RTT_REF * TCP_RTT_SCALE * hz) / 1000; 357 return (0); 358 } 359 360 /* 361 * Perform any necessary tasks before we exit congestion recovery. 362 */ 363 static void 364 htcp_post_recovery(struct cc_var *ccv) 365 { 366 int pipe; 367 struct htcp *htcp_data; 368 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 369 370 pipe = 0; 371 htcp_data = ccv->cc_data; 372 373 if (IN_FASTRECOVERY(CCV(ccv, t_flags))) { 374 /* 375 * If inflight data is less than ssthresh, set cwnd 376 * conservatively to avoid a burst of data, as suggested in the 377 * NewReno RFC. Otherwise, use the HTCP method. 378 */ 379 pipe = tcp_compute_pipe(ccv->tp); 380 if (pipe < CCV(ccv, snd_ssthresh)) 381 /* 382 * Ensure that cwnd down not collape to 1 MSS under 383 * adverse conditions. Implements RFC6582 384 */ 385 CCV(ccv, snd_cwnd) = max(pipe, mss) + mss; 386 else 387 CCV(ccv, snd_cwnd) = max(1, ((htcp_data->beta * 388 htcp_data->prev_cwnd / mss) 389 >> HTCP_SHIFT)) * mss; 390 } 391 } 392 393 static void 394 htcp_recalc_alpha(struct cc_var *ccv) 395 { 396 struct htcp *htcp_data; 397 int alpha, diff, now; 398 399 htcp_data = ccv->cc_data; 400 now = ticks; 401 402 /* 403 * If ticks has wrapped around (will happen approximately once every 49 404 * days on a machine with the default kern.hz=1000) and a flow straddles 405 * the wrap point, our alpha calcs will be completely wrong. We cut our 406 * losses and restart alpha from scratch by setting t_last_cong = now - 407 * HTCP_DELTA_L. 408 * 409 * This does not deflate our cwnd at all. It simply slows the rate cwnd 410 * is growing by until alpha regains the value it held prior to taking 411 * this drastic measure. 412 */ 413 if (now < htcp_data->t_last_cong) 414 htcp_data->t_last_cong = now - HTCP_DELTA_L; 415 416 diff = now - htcp_data->t_last_cong - HTCP_DELTA_L; 417 418 /* Cap alpha if the value of diff would overflow HTCP_CALC_ALPHA(). */ 419 if (diff < htcp_max_diff) { 420 /* 421 * If it has been more than HTCP_DELTA_L ticks since congestion, 422 * increase alpha according to the function defined in the spec. 423 */ 424 if (diff > 0) { 425 alpha = HTCP_CALC_ALPHA(diff); 426 427 /* 428 * Adaptive backoff fairness adjustment: 429 * 2 * (1 - beta) * alpha_raw 430 */ 431 if (V_htcp_adaptive_backoff) 432 alpha = max(1, (2 * ((1 << HTCP_SHIFT) - 433 htcp_data->beta) * alpha) >> HTCP_SHIFT); 434 435 /* 436 * RTT scaling: (RTT / RTT_ref) * alpha 437 * alpha will be the raw value from HTCP_CALC_ALPHA() if 438 * adaptive backoff is off, or the adjusted value if 439 * adaptive backoff is on. 440 */ 441 if (V_htcp_rtt_scaling) 442 alpha = max(1, (min(max(HTCP_MINROWE, 443 (tcp_get_srtt(ccv->tp, TCP_TMR_GRANULARITY_TICKS) << HTCP_SHIFT) / 444 htcp_rtt_ref), HTCP_MAXROWE) * alpha) 445 >> HTCP_SHIFT); 446 447 } else 448 alpha = 1; 449 450 htcp_data->alpha = alpha; 451 } 452 } 453 454 static void 455 htcp_recalc_beta(struct cc_var *ccv) 456 { 457 struct htcp *htcp_data; 458 459 htcp_data = ccv->cc_data; 460 461 /* 462 * TCPTV_SRTTBASE is the initialised value of each connection's SRTT, so 463 * we only calc beta if the connection's SRTT has been changed from its 464 * initial value. beta is bounded to ensure it is always between 465 * HTCP_MINBETA and HTCP_MAXBETA. 466 */ 467 if (V_htcp_adaptive_backoff && htcp_data->minrtt != TCPTV_SRTTBASE && 468 htcp_data->maxrtt != TCPTV_SRTTBASE) 469 htcp_data->beta = min(max(HTCP_MINBETA, 470 (htcp_data->minrtt << HTCP_SHIFT) / htcp_data->maxrtt), 471 HTCP_MAXBETA); 472 else 473 htcp_data->beta = HTCP_MINBETA; 474 } 475 476 /* 477 * Record the minimum and maximum RTT seen for the connection. These are used in 478 * the calculation of beta if adaptive backoff is enabled. 479 */ 480 static void 481 htcp_record_rtt(struct cc_var *ccv) 482 { 483 struct htcp *htcp_data; 484 485 htcp_data = ccv->cc_data; 486 487 /* XXXLAS: Should there be some hysteresis for minrtt? */ 488 489 /* 490 * Record the current SRTT as our minrtt if it's the smallest we've seen 491 * or minrtt is currently equal to its initialised value. Ignore SRTT 492 * until a min number of samples have been taken. 493 */ 494 if ((tcp_get_srtt(ccv->tp, TCP_TMR_GRANULARITY_TICKS) < htcp_data->minrtt || 495 htcp_data->minrtt == TCPTV_SRTTBASE) && 496 (CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES)) 497 htcp_data->minrtt = tcp_get_srtt(ccv->tp, TCP_TMR_GRANULARITY_TICKS); 498 499 /* 500 * Record the current SRTT as our maxrtt if it's the largest we've 501 * seen. Ignore SRTT until a min number of samples have been taken. 502 */ 503 if (tcp_get_srtt(ccv->tp, TCP_TMR_GRANULARITY_TICKS) > htcp_data->maxrtt 504 && CCV(ccv, t_rttupdated) >= HTCP_MIN_RTT_SAMPLES) 505 htcp_data->maxrtt = tcp_get_srtt(ccv->tp, TCP_TMR_GRANULARITY_TICKS); 506 } 507 508 /* 509 * Update the ssthresh in the event of congestion. 510 */ 511 static void 512 htcp_ssthresh_update(struct cc_var *ccv) 513 { 514 struct htcp *htcp_data; 515 516 htcp_data = ccv->cc_data; 517 518 /* 519 * On the first congestion event, set ssthresh to cwnd * 0.5, on 520 * subsequent congestion events, set it to cwnd * beta. 521 */ 522 if (CCV(ccv, snd_ssthresh) == TCP_MAXWIN << TCP_MAX_WINSHIFT) 523 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 524 HTCP_MINBETA) >> HTCP_SHIFT; 525 else { 526 htcp_recalc_beta(ccv); 527 CCV(ccv, snd_ssthresh) = ((u_long)CCV(ccv, snd_cwnd) * 528 htcp_data->beta) >> HTCP_SHIFT; 529 } 530 } 531 532 SYSCTL_DECL(_net_inet_tcp_cc_htcp); 533 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, htcp, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 534 "H-TCP related settings"); 535 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, adaptive_backoff, 536 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_adaptive_backoff), 0, 537 "enable H-TCP adaptive backoff"); 538 SYSCTL_UINT(_net_inet_tcp_cc_htcp, OID_AUTO, rtt_scaling, 539 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(htcp_rtt_scaling), 0, 540 "enable H-TCP RTT scaling"); 541 542 DECLARE_CC_MODULE(htcp, &htcp_cc_algo); 543 MODULE_VERSION(htcp, 2); 544