1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2009-2010 5 * Swinburne University of Technology, Melbourne, Australia 6 * Copyright (c) 2010-2011 The FreeBSD Foundation 7 * All rights reserved. 8 * 9 * This software was developed at the Centre for Advanced Internet 10 * Architectures, Swinburne University of Technology, by David Hayes and 11 * Lawrence Stewart, made possible in part by a grant from the Cisco University 12 * Research Program Fund at Community Foundation Silicon Valley. 13 * 14 * Portions of this software were developed at the Centre for Advanced Internet 15 * Architectures, Swinburne University of Technology, Melbourne, Australia by 16 * David Hayes under sponsorship from the FreeBSD Foundation. 17 * 18 * Redistribution and use in source and binary forms, with or without 19 * modification, are permitted provided that the following conditions 20 * are met: 21 * 1. Redistributions of source code must retain the above copyright 22 * notice, this list of conditions and the following disclaimer. 23 * 2. Redistributions in binary form must reproduce the above copyright 24 * notice, this list of conditions and the following disclaimer in the 25 * documentation and/or other materials provided with the distribution. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 */ 39 40 /* 41 * An implementation of the CAIA-Hamilton delay based congestion control 42 * algorithm, based on "Improved coexistence and loss tolerance for delay based 43 * TCP congestion control" by D. A. Hayes and G. Armitage., in 35th Annual IEEE 44 * Conference on Local Computer Networks (LCN 2010), Denver, Colorado, USA, 45 * 11-14 October 2010. 46 * 47 * Originally released as part of the NewTCP research project at Swinburne 48 * University of Technology's Centre for Advanced Internet Architectures, 49 * Melbourne, Australia, which was made possible in part by a grant from the 50 * Cisco University Research Program Fund at Community Foundation Silicon 51 * Valley. More details are available at: 52 * http://caia.swin.edu.au/urp/newtcp/ 53 */ 54 55 #include <sys/param.h> 56 #include <sys/kernel.h> 57 #include <sys/khelp.h> 58 #include <sys/limits.h> 59 #include <sys/malloc.h> 60 #include <sys/module.h> 61 #include <sys/queue.h> 62 #include <sys/socket.h> 63 #include <sys/socketvar.h> 64 #include <sys/sysctl.h> 65 #include <sys/systm.h> 66 67 #include <net/vnet.h> 68 69 #include <net/route.h> 70 #include <net/route/nhop.h> 71 72 #include <netinet/in_pcb.h> 73 #include <netinet/tcp.h> 74 #include <netinet/tcp_seq.h> 75 #include <netinet/tcp_timer.h> 76 #include <netinet/tcp_var.h> 77 #include <netinet/cc/cc.h> 78 #include <netinet/cc/cc_module.h> 79 80 #include <netinet/khelp/h_ertt.h> 81 82 /* 83 * Private signal type for rate based congestion signal. 84 * See <netinet/cc.h> for appropriate bit-range to use for private signals. 85 */ 86 #define CC_CHD_DELAY 0x02000000 87 88 /* Largest possible number returned by random(). */ 89 #define RANDOM_MAX INT_MAX 90 91 static void chd_ack_received(struct cc_var *ccv, ccsignal_t ack_type); 92 static void chd_cb_destroy(struct cc_var *ccv); 93 static int chd_cb_init(struct cc_var *ccv, void *ptr); 94 static void chd_cong_signal(struct cc_var *ccv, ccsignal_t signal_type); 95 static void chd_conn_init(struct cc_var *ccv); 96 static int chd_mod_init(void); 97 static size_t chd_data_sz(void); 98 99 struct chd { 100 /* 101 * Shadow window - keeps track of what the NewReno congestion window 102 * would have been if delay-based cwnd backoffs had not been made. This 103 * functionality aids coexistence with loss-based TCP flows which may be 104 * sharing links along the path. 105 */ 106 unsigned long shadow_w; 107 /* 108 * Loss-based TCP compatibility flag - When set, it turns on the shadow 109 * window functionality. 110 */ 111 int loss_compete; 112 /* The maximum round trip time seen within a measured rtt period. */ 113 int maxrtt_in_rtt; 114 /* The previous qdly that caused cwnd to backoff. */ 115 int prev_backoff_qdly; 116 }; 117 118 static int ertt_id; 119 120 VNET_DEFINE_STATIC(uint32_t, chd_qmin) = 5; 121 VNET_DEFINE_STATIC(uint32_t, chd_pmax) = 50; 122 VNET_DEFINE_STATIC(uint32_t, chd_loss_fair) = 1; 123 VNET_DEFINE_STATIC(uint32_t, chd_use_max) = 1; 124 VNET_DEFINE_STATIC(uint32_t, chd_qthresh) = 20; 125 #define V_chd_qthresh VNET(chd_qthresh) 126 #define V_chd_qmin VNET(chd_qmin) 127 #define V_chd_pmax VNET(chd_pmax) 128 #define V_chd_loss_fair VNET(chd_loss_fair) 129 #define V_chd_use_max VNET(chd_use_max) 130 131 132 struct cc_algo chd_cc_algo = { 133 .name = "chd", 134 .ack_received = chd_ack_received, 135 .cb_destroy = chd_cb_destroy, 136 .cb_init = chd_cb_init, 137 .cong_signal = chd_cong_signal, 138 .conn_init = chd_conn_init, 139 .mod_init = chd_mod_init, 140 .cc_data_sz = chd_data_sz, 141 .after_idle = newreno_cc_after_idle, 142 .post_recovery = newreno_cc_post_recovery, 143 }; 144 145 static __inline void 146 chd_window_decrease(struct cc_var *ccv) 147 { 148 unsigned long win; 149 150 win = min(CCV(ccv, snd_wnd), CCV(ccv, snd_cwnd)) / CCV(ccv, t_maxseg); 151 win -= max((win / 2), 1); 152 CCV(ccv, snd_ssthresh) = max(win, 2) * CCV(ccv, t_maxseg); 153 } 154 155 /* 156 * Probabilistic backoff function. Returns 1 if we should backoff or 0 157 * otherwise. The calculation of p is similar to the calculation of p in cc_hd. 158 */ 159 static __inline int 160 should_backoff(int qdly, int maxqdly, struct chd *chd_data) 161 { 162 unsigned long p, rand; 163 164 rand = random(); 165 166 if (qdly < V_chd_qthresh) { 167 chd_data->loss_compete = 0; 168 p = (((RANDOM_MAX / 100) * V_chd_pmax) / 169 (V_chd_qthresh - V_chd_qmin)) * 170 (qdly - V_chd_qmin); 171 } else { 172 if (qdly > V_chd_qthresh) { 173 p = (((RANDOM_MAX / 100) * V_chd_pmax) / 174 (maxqdly - V_chd_qthresh)) * 175 (maxqdly - qdly); 176 if (V_chd_loss_fair && rand < p) 177 chd_data->loss_compete = 1; 178 } else { 179 p = (RANDOM_MAX / 100) * V_chd_pmax; 180 chd_data->loss_compete = 0; 181 } 182 } 183 184 return (rand < p); 185 } 186 187 static __inline void 188 chd_window_increase(struct cc_var *ccv, int new_measurement) 189 { 190 struct chd *chd_data; 191 int incr; 192 193 chd_data = ccv->cc_data; 194 incr = 0; 195 196 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { 197 /* Adapted from NewReno slow start. */ 198 if (V_tcp_do_rfc3465) { 199 /* In slow-start with ABC enabled. */ 200 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) { 201 /* Not due to RTO. */ 202 incr = min(ccv->bytes_this_ack, 203 V_tcp_abc_l_var * CCV(ccv, t_maxseg)); 204 } else { 205 /* Due to RTO. */ 206 incr = min(ccv->bytes_this_ack, 207 CCV(ccv, t_maxseg)); 208 } 209 } else 210 incr = CCV(ccv, t_maxseg); 211 212 } else { /* Congestion avoidance. */ 213 if (V_tcp_do_rfc3465) { 214 if (ccv->flags & CCF_ABC_SENTAWND) { 215 ccv->flags &= ~CCF_ABC_SENTAWND; 216 incr = CCV(ccv, t_maxseg); 217 } 218 } else if (new_measurement) 219 incr = CCV(ccv, t_maxseg); 220 } 221 222 if (chd_data->shadow_w > 0) { 223 /* Track NewReno window. */ 224 chd_data->shadow_w = min(chd_data->shadow_w + incr, 225 TCP_MAXWIN << CCV(ccv, snd_scale)); 226 } 227 228 CCV(ccv,snd_cwnd) = min(CCV(ccv, snd_cwnd) + incr, 229 TCP_MAXWIN << CCV(ccv, snd_scale)); 230 } 231 232 /* 233 * All ACK signals are used for timing measurements to determine delay-based 234 * congestion. However, window increases are only performed when 235 * ack_type == CC_ACK. 236 */ 237 static void 238 chd_ack_received(struct cc_var *ccv, ccsignal_t ack_type) 239 { 240 struct chd *chd_data; 241 struct ertt *e_t; 242 int backoff, new_measurement, qdly, rtt; 243 244 e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); 245 chd_data = ccv->cc_data; 246 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT; 247 backoff = qdly = 0; 248 249 chd_data->maxrtt_in_rtt = imax(e_t->rtt, chd_data->maxrtt_in_rtt); 250 251 if (new_measurement) { 252 /* 253 * There is a new per RTT measurement, so check to see if there 254 * is delay based congestion. 255 */ 256 rtt = V_chd_use_max ? chd_data->maxrtt_in_rtt : e_t->rtt; 257 chd_data->maxrtt_in_rtt = 0; 258 259 if (rtt && e_t->minrtt && !IN_RECOVERY(CCV(ccv, t_flags))) { 260 qdly = rtt - e_t->minrtt; 261 if (qdly > V_chd_qmin) { 262 /* 263 * Probabilistic delay based congestion 264 * indication. 265 */ 266 backoff = should_backoff(qdly, 267 e_t->maxrtt - e_t->minrtt, chd_data); 268 } else 269 chd_data->loss_compete = 0; 270 } 271 /* Reset per RTT measurement flag to start a new measurement. */ 272 e_t->flags &= ~ERTT_NEW_MEASUREMENT; 273 } 274 275 if (backoff) { 276 /* 277 * Update shadow_w before delay based backoff. 278 */ 279 if (chd_data->loss_compete || 280 qdly > chd_data->prev_backoff_qdly) { 281 /* 282 * Delay is higher than when we backed off previously, 283 * so it is possible that this flow is competing with 284 * loss based flows. 285 */ 286 chd_data->shadow_w = max(CCV(ccv, snd_cwnd), 287 chd_data->shadow_w); 288 } else { 289 /* 290 * Reset shadow_w, as it is probable that this flow is 291 * not competing with loss based flows at the moment. 292 */ 293 chd_data->shadow_w = 0; 294 } 295 296 chd_data->prev_backoff_qdly = qdly; 297 /* 298 * Send delay-based congestion signal to the congestion signal 299 * handler. 300 */ 301 chd_cong_signal(ccv, CC_CHD_DELAY); 302 303 } else if (ack_type == CC_ACK) 304 chd_window_increase(ccv, new_measurement); 305 } 306 307 static void 308 chd_cb_destroy(struct cc_var *ccv) 309 { 310 free(ccv->cc_data, M_CC_MEM); 311 } 312 313 size_t 314 chd_data_sz(void) 315 { 316 return (sizeof(struct chd)); 317 } 318 319 static int 320 chd_cb_init(struct cc_var *ccv, void *ptr) 321 { 322 struct chd *chd_data; 323 324 INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); 325 if (ptr == NULL) { 326 chd_data = malloc(sizeof(struct chd), M_CC_MEM, M_NOWAIT); 327 if (chd_data == NULL) 328 return (ENOMEM); 329 } else 330 chd_data = ptr; 331 332 chd_data->shadow_w = 0; 333 ccv->cc_data = chd_data; 334 335 return (0); 336 } 337 338 static void 339 chd_cong_signal(struct cc_var *ccv, ccsignal_t signal_type) 340 { 341 struct ertt *e_t; 342 struct chd *chd_data; 343 int qdly; 344 345 e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); 346 chd_data = ccv->cc_data; 347 qdly = imax(e_t->rtt, chd_data->maxrtt_in_rtt) - e_t->minrtt; 348 349 switch((int)signal_type) { 350 case CC_CHD_DELAY: 351 chd_window_decrease(ccv); /* Set new ssthresh. */ 352 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 353 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 354 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 355 break; 356 357 case CC_NDUPACK: /* Packet loss. */ 358 /* 359 * Only react to loss as a congestion signal if qdly > 360 * V_chd_qthresh. If qdly is less than qthresh, presume that 361 * this is a non congestion related loss. If qdly is greater 362 * than qthresh, assume that we are competing with loss based 363 * tcp flows and restore window from any unnecessary backoffs, 364 * before the decrease. 365 */ 366 if (!IN_RECOVERY(CCV(ccv, t_flags)) && qdly > V_chd_qthresh) { 367 if (chd_data->loss_compete) { 368 CCV(ccv, snd_cwnd) = max(CCV(ccv, snd_cwnd), 369 chd_data->shadow_w); 370 } 371 chd_window_decrease(ccv); 372 } else { 373 /* 374 * This loss isn't congestion related, or already 375 * recovering from congestion. 376 */ 377 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 378 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 379 } 380 381 if (chd_data->shadow_w > 0) { 382 chd_data->shadow_w = max(chd_data->shadow_w / 383 CCV(ccv, t_maxseg) / 2, 2) * CCV(ccv, t_maxseg); 384 } 385 ENTER_FASTRECOVERY(CCV(ccv, t_flags)); 386 break; 387 388 default: 389 newreno_cc_cong_signal(ccv, signal_type); 390 break; 391 } 392 } 393 394 static void 395 chd_conn_init(struct cc_var *ccv) 396 { 397 struct chd *chd_data; 398 399 chd_data = ccv->cc_data; 400 chd_data->prev_backoff_qdly = 0; 401 chd_data->maxrtt_in_rtt = 0; 402 chd_data->loss_compete = 0; 403 /* 404 * Initialise the shadow_cwnd to be equal to snd_cwnd in case we are 405 * competing with loss based flows from the start. 406 */ 407 chd_data->shadow_w = CCV(ccv, snd_cwnd); 408 } 409 410 static int 411 chd_mod_init(void) 412 { 413 414 ertt_id = khelp_get_id("ertt"); 415 if (ertt_id <= 0) { 416 printf("%s: h_ertt module not found\n", __func__); 417 return (ENOENT); 418 } 419 return (0); 420 } 421 422 static int 423 chd_loss_fair_handler(SYSCTL_HANDLER_ARGS) 424 { 425 int error; 426 uint32_t new; 427 428 new = V_chd_loss_fair; 429 error = sysctl_handle_int(oidp, &new, 0, req); 430 if (error == 0 && req->newptr != NULL) { 431 if (new > 1) 432 error = EINVAL; 433 else 434 V_chd_loss_fair = new; 435 } 436 437 return (error); 438 } 439 440 static int 441 chd_pmax_handler(SYSCTL_HANDLER_ARGS) 442 { 443 int error; 444 uint32_t new; 445 446 new = V_chd_pmax; 447 error = sysctl_handle_int(oidp, &new, 0, req); 448 if (error == 0 && req->newptr != NULL) { 449 if (new == 0 || new > 100) 450 error = EINVAL; 451 else 452 V_chd_pmax = new; 453 } 454 455 return (error); 456 } 457 458 static int 459 chd_qthresh_handler(SYSCTL_HANDLER_ARGS) 460 { 461 int error; 462 uint32_t new; 463 464 new = V_chd_qthresh; 465 error = sysctl_handle_int(oidp, &new, 0, req); 466 if (error == 0 && req->newptr != NULL) { 467 if (new <= V_chd_qmin) 468 error = EINVAL; 469 else 470 V_chd_qthresh = new; 471 } 472 473 return (error); 474 } 475 476 SYSCTL_DECL(_net_inet_tcp_cc_chd); 477 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, chd, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 478 "CAIA Hamilton delay-based congestion control related settings"); 479 480 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, loss_fair, 481 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 482 &VNET_NAME(chd_loss_fair), 1, &chd_loss_fair_handler, 483 "IU", "Flag to enable shadow window functionality."); 484 485 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, pmax, 486 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 487 &VNET_NAME(chd_pmax), 5, &chd_pmax_handler, 488 "IU", "Per RTT maximum backoff probability as a percentage"); 489 490 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, queue_threshold, 491 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 492 &VNET_NAME(chd_qthresh), 20, &chd_qthresh_handler, 493 "IU", "Queueing congestion threshold in ticks"); 494 495 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, queue_min, 496 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_qmin), 5, 497 "Minimum queueing delay threshold in ticks"); 498 499 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, use_max, 500 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_use_max), 1, 501 "Use the maximum RTT seen within the measurement period (RTT) " 502 "as the basic delay measurement for the algorithm."); 503 504 DECLARE_CC_MODULE(chd, &chd_cc_algo); 505 MODULE_VERSION(chd, 2); 506 MODULE_DEPEND(chd, ertt, 1, 1, 1); 507