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/prng.h> 62 #include <sys/queue.h> 63 #include <sys/socket.h> 64 #include <sys/socketvar.h> 65 #include <sys/sysctl.h> 66 #include <sys/systm.h> 67 68 #include <net/vnet.h> 69 70 #include <net/route.h> 71 #include <net/route/nhop.h> 72 73 #include <netinet/in_pcb.h> 74 #include <netinet/tcp.h> 75 #include <netinet/tcp_seq.h> 76 #include <netinet/tcp_timer.h> 77 #include <netinet/tcp_var.h> 78 #include <netinet/cc/cc.h> 79 #include <netinet/cc/cc_module.h> 80 81 #include <netinet/khelp/h_ertt.h> 82 83 /* 84 * Private signal type for rate based congestion signal. 85 * See <netinet/cc.h> for appropriate bit-range to use for private signals. 86 */ 87 #define CC_CHD_DELAY 0x02000000 88 89 /* Largest possible number returned by prng32(). */ 90 #define RANDOM_MAX UINT32_MAX 91 92 static void chd_ack_received(struct cc_var *ccv, ccsignal_t ack_type); 93 static void chd_cb_destroy(struct cc_var *ccv); 94 static int chd_cb_init(struct cc_var *ccv, void *ptr); 95 static void chd_cong_signal(struct cc_var *ccv, ccsignal_t signal_type); 96 static void chd_conn_init(struct cc_var *ccv); 97 static int chd_mod_init(void); 98 static size_t chd_data_sz(void); 99 100 struct chd { 101 /* 102 * Shadow window - keeps track of what the NewReno congestion window 103 * would have been if delay-based cwnd backoffs had not been made. This 104 * functionality aids coexistence with loss-based TCP flows which may be 105 * sharing links along the path. 106 */ 107 unsigned long shadow_w; 108 /* 109 * Loss-based TCP compatibility flag - When set, it turns on the shadow 110 * window functionality. 111 */ 112 int loss_compete; 113 /* The maximum round trip time seen within a measured rtt period. */ 114 int maxrtt_in_rtt; 115 /* The previous qdly that caused cwnd to backoff. */ 116 int prev_backoff_qdly; 117 }; 118 119 static int ertt_id; 120 121 VNET_DEFINE_STATIC(uint32_t, chd_qmin) = 5; 122 VNET_DEFINE_STATIC(uint32_t, chd_pmax) = 50; 123 VNET_DEFINE_STATIC(uint32_t, chd_loss_fair) = 1; 124 VNET_DEFINE_STATIC(uint32_t, chd_use_max) = 1; 125 VNET_DEFINE_STATIC(uint32_t, chd_qthresh) = 20; 126 #define V_chd_qthresh VNET(chd_qthresh) 127 #define V_chd_qmin VNET(chd_qmin) 128 #define V_chd_pmax VNET(chd_pmax) 129 #define V_chd_loss_fair VNET(chd_loss_fair) 130 #define V_chd_use_max VNET(chd_use_max) 131 132 133 struct cc_algo chd_cc_algo = { 134 .name = "chd", 135 .ack_received = chd_ack_received, 136 .cb_destroy = chd_cb_destroy, 137 .cb_init = chd_cb_init, 138 .cong_signal = chd_cong_signal, 139 .conn_init = chd_conn_init, 140 .mod_init = chd_mod_init, 141 .cc_data_sz = chd_data_sz, 142 .after_idle = newreno_cc_after_idle, 143 .post_recovery = newreno_cc_post_recovery, 144 }; 145 146 static __inline void 147 chd_window_decrease(struct cc_var *ccv) 148 { 149 unsigned long win; 150 151 win = min(CCV(ccv, snd_wnd), CCV(ccv, snd_cwnd)) / CCV(ccv, t_maxseg); 152 win -= max((win / 2), 1); 153 CCV(ccv, snd_ssthresh) = max(win, 2) * CCV(ccv, t_maxseg); 154 } 155 156 /* 157 * Probabilistic backoff function. Returns 1 if we should backoff or 0 158 * otherwise. The calculation of p is similar to the calculation of p in cc_hd. 159 */ 160 static __inline int 161 should_backoff(int qdly, int maxqdly, struct chd *chd_data) 162 { 163 uint32_t rand, p; 164 165 rand = prng32(); 166 167 if (qdly < V_chd_qthresh) { 168 chd_data->loss_compete = 0; 169 p = (((RANDOM_MAX / 100) * V_chd_pmax) / 170 (V_chd_qthresh - V_chd_qmin)) * 171 (qdly - V_chd_qmin); 172 } else { 173 if (qdly > V_chd_qthresh) { 174 p = (((RANDOM_MAX / 100) * V_chd_pmax) / 175 (maxqdly - V_chd_qthresh)) * 176 (maxqdly - qdly); 177 if (V_chd_loss_fair && rand < p) 178 chd_data->loss_compete = 1; 179 } else { 180 p = (RANDOM_MAX / 100) * V_chd_pmax; 181 chd_data->loss_compete = 0; 182 } 183 } 184 185 return (rand < p); 186 } 187 188 static __inline void 189 chd_window_increase(struct cc_var *ccv, int new_measurement) 190 { 191 struct chd *chd_data; 192 int incr; 193 194 chd_data = ccv->cc_data; 195 incr = 0; 196 197 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { 198 /* Adapted from NewReno slow start. */ 199 if (V_tcp_do_rfc3465) { 200 /* In slow-start with ABC enabled. */ 201 if (CCV(ccv, snd_nxt) == CCV(ccv, snd_max)) { 202 /* Not due to RTO. */ 203 incr = min(ccv->bytes_this_ack, 204 V_tcp_abc_l_var * CCV(ccv, t_maxseg)); 205 } else { 206 /* Due to RTO. */ 207 incr = min(ccv->bytes_this_ack, 208 CCV(ccv, t_maxseg)); 209 } 210 } else 211 incr = CCV(ccv, t_maxseg); 212 213 } else { /* Congestion avoidance. */ 214 if (V_tcp_do_rfc3465) { 215 if (ccv->flags & CCF_ABC_SENTAWND) { 216 ccv->flags &= ~CCF_ABC_SENTAWND; 217 incr = CCV(ccv, t_maxseg); 218 } 219 } else if (new_measurement) 220 incr = CCV(ccv, t_maxseg); 221 } 222 223 if (chd_data->shadow_w > 0) { 224 /* Track NewReno window. */ 225 chd_data->shadow_w = min(chd_data->shadow_w + incr, 226 TCP_MAXWIN << CCV(ccv, snd_scale)); 227 } 228 229 CCV(ccv,snd_cwnd) = min(CCV(ccv, snd_cwnd) + incr, 230 TCP_MAXWIN << CCV(ccv, snd_scale)); 231 } 232 233 /* 234 * All ACK signals are used for timing measurements to determine delay-based 235 * congestion. However, window increases are only performed when 236 * ack_type == CC_ACK. 237 */ 238 static void 239 chd_ack_received(struct cc_var *ccv, ccsignal_t ack_type) 240 { 241 struct chd *chd_data; 242 struct ertt *e_t; 243 int backoff, new_measurement, qdly, rtt; 244 245 e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); 246 chd_data = ccv->cc_data; 247 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT; 248 backoff = qdly = 0; 249 250 chd_data->maxrtt_in_rtt = imax(e_t->rtt, chd_data->maxrtt_in_rtt); 251 252 if (new_measurement) { 253 /* 254 * There is a new per RTT measurement, so check to see if there 255 * is delay based congestion. 256 */ 257 rtt = V_chd_use_max ? chd_data->maxrtt_in_rtt : e_t->rtt; 258 chd_data->maxrtt_in_rtt = 0; 259 260 if (rtt && e_t->minrtt && !IN_RECOVERY(CCV(ccv, t_flags))) { 261 qdly = rtt - e_t->minrtt; 262 if (qdly > V_chd_qmin) { 263 /* 264 * Probabilistic delay based congestion 265 * indication. 266 */ 267 backoff = should_backoff(qdly, 268 e_t->maxrtt - e_t->minrtt, chd_data); 269 } else 270 chd_data->loss_compete = 0; 271 } 272 /* Reset per RTT measurement flag to start a new measurement. */ 273 e_t->flags &= ~ERTT_NEW_MEASUREMENT; 274 } 275 276 if (backoff) { 277 /* 278 * Update shadow_w before delay based backoff. 279 */ 280 if (chd_data->loss_compete || 281 qdly > chd_data->prev_backoff_qdly) { 282 /* 283 * Delay is higher than when we backed off previously, 284 * so it is possible that this flow is competing with 285 * loss based flows. 286 */ 287 chd_data->shadow_w = max(CCV(ccv, snd_cwnd), 288 chd_data->shadow_w); 289 } else { 290 /* 291 * Reset shadow_w, as it is probable that this flow is 292 * not competing with loss based flows at the moment. 293 */ 294 chd_data->shadow_w = 0; 295 } 296 297 chd_data->prev_backoff_qdly = qdly; 298 /* 299 * Send delay-based congestion signal to the congestion signal 300 * handler. 301 */ 302 chd_cong_signal(ccv, CC_CHD_DELAY); 303 304 } else if (ack_type == CC_ACK) 305 chd_window_increase(ccv, new_measurement); 306 } 307 308 static void 309 chd_cb_destroy(struct cc_var *ccv) 310 { 311 free(ccv->cc_data, M_CC_MEM); 312 } 313 314 size_t 315 chd_data_sz(void) 316 { 317 return (sizeof(struct chd)); 318 } 319 320 static int 321 chd_cb_init(struct cc_var *ccv, void *ptr) 322 { 323 struct chd *chd_data; 324 325 INP_WLOCK_ASSERT(tptoinpcb(ccv->tp)); 326 if (ptr == NULL) { 327 chd_data = malloc(sizeof(struct chd), M_CC_MEM, M_NOWAIT); 328 if (chd_data == NULL) 329 return (ENOMEM); 330 } else 331 chd_data = ptr; 332 333 chd_data->shadow_w = 0; 334 ccv->cc_data = chd_data; 335 336 return (0); 337 } 338 339 static void 340 chd_cong_signal(struct cc_var *ccv, ccsignal_t signal_type) 341 { 342 struct ertt *e_t; 343 struct chd *chd_data; 344 int qdly; 345 346 e_t = khelp_get_osd(&CCV(ccv, t_osd), ertt_id); 347 chd_data = ccv->cc_data; 348 qdly = imax(e_t->rtt, chd_data->maxrtt_in_rtt) - e_t->minrtt; 349 350 switch((int)signal_type) { 351 case CC_CHD_DELAY: 352 chd_window_decrease(ccv); /* Set new ssthresh. */ 353 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 354 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 355 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 356 break; 357 358 case CC_NDUPACK: /* Packet loss. */ 359 /* 360 * Only react to loss as a congestion signal if qdly > 361 * V_chd_qthresh. If qdly is less than qthresh, presume that 362 * this is a non congestion related loss. If qdly is greater 363 * than qthresh, assume that we are competing with loss based 364 * tcp flows and restore window from any unnecessary backoffs, 365 * before the decrease. 366 */ 367 if (!IN_RECOVERY(CCV(ccv, t_flags)) && qdly > V_chd_qthresh) { 368 if (chd_data->loss_compete) { 369 CCV(ccv, snd_cwnd) = max(CCV(ccv, snd_cwnd), 370 chd_data->shadow_w); 371 } 372 chd_window_decrease(ccv); 373 } else { 374 /* 375 * This loss isn't congestion related, or already 376 * recovering from congestion. 377 */ 378 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 379 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 380 } 381 382 if (chd_data->shadow_w > 0) { 383 chd_data->shadow_w = max(chd_data->shadow_w / 384 CCV(ccv, t_maxseg) / 2, 2) * CCV(ccv, t_maxseg); 385 } 386 ENTER_FASTRECOVERY(CCV(ccv, t_flags)); 387 break; 388 389 default: 390 newreno_cc_cong_signal(ccv, signal_type); 391 break; 392 } 393 } 394 395 static void 396 chd_conn_init(struct cc_var *ccv) 397 { 398 struct chd *chd_data; 399 400 chd_data = ccv->cc_data; 401 chd_data->prev_backoff_qdly = 0; 402 chd_data->maxrtt_in_rtt = 0; 403 chd_data->loss_compete = 0; 404 /* 405 * Initialise the shadow_cwnd to be equal to snd_cwnd in case we are 406 * competing with loss based flows from the start. 407 */ 408 chd_data->shadow_w = CCV(ccv, snd_cwnd); 409 } 410 411 static int 412 chd_mod_init(void) 413 { 414 415 ertt_id = khelp_get_id("ertt"); 416 if (ertt_id <= 0) { 417 printf("%s: h_ertt module not found\n", __func__); 418 return (ENOENT); 419 } 420 return (0); 421 } 422 423 static int 424 chd_loss_fair_handler(SYSCTL_HANDLER_ARGS) 425 { 426 int error; 427 uint32_t new; 428 429 new = V_chd_loss_fair; 430 error = sysctl_handle_int(oidp, &new, 0, req); 431 if (error == 0 && req->newptr != NULL) { 432 if (new > 1) 433 error = EINVAL; 434 else 435 V_chd_loss_fair = new; 436 } 437 438 return (error); 439 } 440 441 static int 442 chd_pmax_handler(SYSCTL_HANDLER_ARGS) 443 { 444 int error; 445 uint32_t new; 446 447 new = V_chd_pmax; 448 error = sysctl_handle_int(oidp, &new, 0, req); 449 if (error == 0 && req->newptr != NULL) { 450 if (new == 0 || new > 100) 451 error = EINVAL; 452 else 453 V_chd_pmax = new; 454 } 455 456 return (error); 457 } 458 459 static int 460 chd_qthresh_handler(SYSCTL_HANDLER_ARGS) 461 { 462 int error; 463 uint32_t new; 464 465 new = V_chd_qthresh; 466 error = sysctl_handle_int(oidp, &new, 0, req); 467 if (error == 0 && req->newptr != NULL) { 468 if (new <= V_chd_qmin) 469 error = EINVAL; 470 else 471 V_chd_qthresh = new; 472 } 473 474 return (error); 475 } 476 477 SYSCTL_DECL(_net_inet_tcp_cc_chd); 478 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, chd, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 479 "CAIA Hamilton delay-based congestion control related settings"); 480 481 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, loss_fair, 482 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 483 &VNET_NAME(chd_loss_fair), 1, &chd_loss_fair_handler, 484 "IU", "Flag to enable shadow window functionality."); 485 486 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, pmax, 487 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 488 &VNET_NAME(chd_pmax), 5, &chd_pmax_handler, 489 "IU", "Per RTT maximum backoff probability as a percentage"); 490 491 SYSCTL_PROC(_net_inet_tcp_cc_chd, OID_AUTO, queue_threshold, 492 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 493 &VNET_NAME(chd_qthresh), 20, &chd_qthresh_handler, 494 "IU", "Queueing congestion threshold in ticks"); 495 496 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, queue_min, 497 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_qmin), 5, 498 "Minimum queueing delay threshold in ticks"); 499 500 SYSCTL_UINT(_net_inet_tcp_cc_chd, OID_AUTO, use_max, 501 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(chd_use_max), 1, 502 "Use the maximum RTT seen within the measurement period (RTT) " 503 "as the basic delay measurement for the algorithm."); 504 505 DECLARE_CC_MODULE(chd, &chd_cc_algo); 506 MODULE_VERSION(chd, 2); 507 MODULE_DEPEND(chd, ertt, 1, 1, 1); 508