1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2009-2013 5 * Swinburne University of Technology, Melbourne, Australia 6 * All rights reserved. 7 * 8 * This software was developed at the Centre for Advanced Internet 9 * Architectures, Swinburne University of Technology, by David Hayes, made 10 * possible in part by a gift from The Cisco University Research Program Fund, 11 * a corporate advised fund of Silicon Valley Community Foundation. Development 12 * and testing were further assisted by a grant from the FreeBSD Foundation. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 23 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 26 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /* 37 * CAIA Delay-Gradient (CDG) congestion control algorithm 38 * 39 * An implemention of the delay-gradient congestion control algorithm proposed 40 * in the following paper: 41 * 42 * D. A. Hayes and G. Armitage, "Revisiting TCP Congestion Control using Delay 43 * Gradients", in IFIP Networking, Valencia, Spain, 9-13 May 2011. 44 * 45 * Developed as part of the NewTCP research project at Swinburne University of 46 * Technology's Centre for Advanced Internet Architectures, Melbourne, 47 * Australia. More details are available at: 48 * http://caia.swin.edu.au/urp/newtcp/ 49 */ 50 51 #include <sys/param.h> 52 #include <sys/hhook.h> 53 #include <sys/kernel.h> 54 #include <sys/khelp.h> 55 #include <sys/limits.h> 56 #include <sys/lock.h> 57 #include <sys/malloc.h> 58 #include <sys/module.h> 59 #include <sys/queue.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 #include <netinet/khelp/h_ertt.h> 79 80 #include <vm/uma.h> 81 82 #define CDG_VERSION "0.1" 83 84 /* Private delay-gradient induced congestion control signal. */ 85 #define CC_CDG_DELAY 0x01000000 86 87 /* NewReno window deflation factor on loss (as a percentage). */ 88 #define RENO_BETA 50 89 90 /* Queue states. */ 91 #define CDG_Q_EMPTY 1 92 #define CDG_Q_RISING 2 93 #define CDG_Q_FALLING 3 94 #define CDG_Q_FULL 4 95 #define CDG_Q_UNKNOWN 9999 96 97 /* Number of bit shifts used in probexp lookup table. */ 98 #define EXP_PREC 15 99 100 /* Largest gradient represented in probexp lookup table. */ 101 #define MAXGRAD 5 102 103 /* 104 * Delay Precision Enhance - number of bit shifts used for qtrend related 105 * integer arithmetic precision. 106 */ 107 #define D_P_E 7 108 109 struct qdiff_sample { 110 long qdiff; 111 STAILQ_ENTRY(qdiff_sample) qdiff_lnk; 112 }; 113 114 struct cdg { 115 long max_qtrend; 116 long min_qtrend; 117 STAILQ_HEAD(minrtts_head, qdiff_sample) qdiffmin_q; 118 STAILQ_HEAD(maxrtts_head, qdiff_sample) qdiffmax_q; 119 long window_incr; 120 /* rttcount for window increase when in congestion avoidance */ 121 long rtt_count; 122 /* maximum measured rtt within an rtt period */ 123 int maxrtt_in_rtt; 124 /* maximum measured rtt within prev rtt period */ 125 int maxrtt_in_prevrtt; 126 /* minimum measured rtt within an rtt period */ 127 int minrtt_in_rtt; 128 /* minimum measured rtt within prev rtt period */ 129 int minrtt_in_prevrtt; 130 /* consecutive congestion episode counter */ 131 uint32_t consec_cong_cnt; 132 /* when tracking a new reno type loss window */ 133 uint32_t shadow_w; 134 /* maximum number of samples in the moving average queue */ 135 int sample_q_size; 136 /* number of samples in the moving average queue */ 137 int num_samples; 138 /* estimate of the queue state of the path */ 139 int queue_state; 140 }; 141 142 /* 143 * Lookup table for: 144 * (1 - exp(-x)) << EXP_PREC, where x = [0,MAXGRAD] in 2^-7 increments 145 * 146 * Note: probexp[0] is set to 10 (not 0) as a safety for very low increase 147 * gradients. 148 */ 149 static const int probexp[641] = { 150 10,255,508,759,1008,1255,1501,1744,1985,2225,2463,2698,2932,3165,3395,3624, 151 3850,4075,4299,4520,4740,4958,5175,5389,5602,5814,6024,6232,6438,6643,6846, 152 7048,7248,7447,7644,7839,8033,8226,8417,8606,8794,8981,9166,9350,9532,9713, 153 9892,10070,10247,10422,10596,10769,10940,11110,11278,11445,11611,11776,11939, 154 12101,12262,12422,12580,12737,12893,13048,13201,13354,13505,13655,13803,13951, 155 14097,14243,14387,14530,14672,14813,14952,15091,15229,15365,15500,15635,15768, 156 15900,16032,16162,16291,16419,16547,16673,16798,16922,17046,17168,17289,17410, 157 17529,17648,17766,17882,17998,18113,18227,18340,18453,18564,18675,18784,18893, 158 19001,19108,19215,19320,19425,19529,19632,19734,19835,19936,20036,20135,20233, 159 20331,20427,20523,20619,20713,20807,20900,20993,21084,21175,21265,21355,21444, 160 21532,21619,21706,21792,21878,21962,22046,22130,22213,22295,22376,22457,22537, 161 22617,22696,22774,22852,22929,23006,23082,23157,23232,23306,23380,23453,23525, 162 23597,23669,23739,23810,23879,23949,24017,24085,24153,24220,24286,24352,24418, 163 24483,24547,24611,24675,24738,24800,24862,24924,24985,25045,25106,25165,25224, 164 25283,25341,25399,25456,25513,25570,25626,25681,25737,25791,25846,25899,25953, 165 26006,26059,26111,26163,26214,26265,26316,26366,26416,26465,26514,26563,26611, 166 26659,26707,26754,26801,26847,26893,26939,26984,27029,27074,27118,27162,27206, 167 27249,27292,27335,27377,27419,27460,27502,27543,27583,27624,27664,27703,27743, 168 27782,27821,27859,27897,27935,27973,28010,28047,28084,28121,28157,28193,28228, 169 28263,28299,28333,28368,28402,28436,28470,28503,28536,28569,28602,28634,28667, 170 28699,28730,28762,28793,28824,28854,28885,28915,28945,28975,29004,29034,29063, 171 29092,29120,29149,29177,29205,29232,29260,29287,29314,29341,29368,29394,29421, 172 29447,29472,29498,29524,29549,29574,29599,29623,29648,29672,29696,29720,29744, 173 29767,29791,29814,29837,29860,29882,29905,29927,29949,29971,29993,30014,30036, 174 30057,30078,30099,30120,30141,30161,30181,30201,30221,30241,30261,30280,30300, 175 30319,30338,30357,30376,30394,30413,30431,30449,30467,30485,30503,30521,30538, 176 30555,30573,30590,30607,30624,30640,30657,30673,30690,30706,30722,30738,30753, 177 30769,30785,30800,30815,30831,30846,30861,30876,30890,30905,30919,30934,30948, 178 30962,30976,30990,31004,31018,31031,31045,31058,31072,31085,31098,31111,31124, 179 31137,31149,31162,31174,31187,31199,31211,31223,31235,31247,31259,31271,31283, 180 31294,31306,31317,31328,31339,31351,31362,31373,31383,31394,31405,31416,31426, 181 31436,31447,31457,31467,31477,31487,31497,31507,31517,31527,31537,31546,31556, 182 31565,31574,31584,31593,31602,31611,31620,31629,31638,31647,31655,31664,31673, 183 31681,31690,31698,31706,31715,31723,31731,31739,31747,31755,31763,31771,31778, 184 31786,31794,31801,31809,31816,31824,31831,31838,31846,31853,31860,31867,31874, 185 31881,31888,31895,31902,31908,31915,31922,31928,31935,31941,31948,31954,31960, 186 31967,31973,31979,31985,31991,31997,32003,32009,32015,32021,32027,32033,32038, 187 32044,32050,32055,32061,32066,32072,32077,32083,32088,32093,32098,32104,32109, 188 32114,32119,32124,32129,32134,32139,32144,32149,32154,32158,32163,32168,32173, 189 32177,32182,32186,32191,32195,32200,32204,32209,32213,32217,32222,32226,32230, 190 32234,32238,32242,32247,32251,32255,32259,32263,32267,32270,32274,32278,32282, 191 32286,32290,32293,32297,32301,32304,32308,32311,32315,32318,32322,32325,32329, 192 32332,32336,32339,32342,32346,32349,32352,32356,32359,32362,32365,32368,32371, 193 32374,32377,32381,32384,32387,32389,32392,32395,32398,32401,32404,32407,32410, 194 32412,32415,32418,32421,32423,32426,32429,32431,32434,32437,32439,32442,32444, 195 32447,32449,32452,32454,32457,32459,32461,32464,32466,32469,32471,32473,32476, 196 32478,32480,32482,32485,32487,32489,32491,32493,32495,32497,32500,32502,32504, 197 32506,32508,32510,32512,32514,32516,32518,32520,32522,32524,32526,32527,32529, 198 32531,32533,32535,32537,32538,32540,32542,32544,32545,32547}; 199 200 static uma_zone_t qdiffsample_zone; 201 static int ertt_id; 202 203 VNET_DEFINE_STATIC(uint32_t, cdg_alpha_inc); 204 VNET_DEFINE_STATIC(uint32_t, cdg_beta_delay); 205 VNET_DEFINE_STATIC(uint32_t, cdg_beta_loss); 206 VNET_DEFINE_STATIC(uint32_t, cdg_smoothing_factor); 207 VNET_DEFINE_STATIC(uint32_t, cdg_exp_backoff_scale); 208 VNET_DEFINE_STATIC(uint32_t, cdg_consec_cong); 209 VNET_DEFINE_STATIC(uint32_t, cdg_hold_backoff); 210 #define V_cdg_alpha_inc VNET(cdg_alpha_inc) 211 #define V_cdg_beta_delay VNET(cdg_beta_delay) 212 #define V_cdg_beta_loss VNET(cdg_beta_loss) 213 #define V_cdg_smoothing_factor VNET(cdg_smoothing_factor) 214 #define V_cdg_exp_backoff_scale VNET(cdg_exp_backoff_scale) 215 #define V_cdg_consec_cong VNET(cdg_consec_cong) 216 #define V_cdg_hold_backoff VNET(cdg_hold_backoff) 217 218 /* Function prototypes. */ 219 static int cdg_mod_init(void); 220 static int cdg_mod_destroy(void); 221 static void cdg_conn_init(struct cc_var *ccv); 222 static int cdg_cb_init(struct cc_var *ccv, void *ptr); 223 static void cdg_cb_destroy(struct cc_var *ccv); 224 static void cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type); 225 static void cdg_ack_received(struct cc_var *ccv, uint16_t ack_type); 226 static size_t cdg_data_sz(void); 227 228 struct cc_algo cdg_cc_algo = { 229 .name = "cdg", 230 .mod_init = cdg_mod_init, 231 .ack_received = cdg_ack_received, 232 .cb_destroy = cdg_cb_destroy, 233 .cb_init = cdg_cb_init, 234 .conn_init = cdg_conn_init, 235 .cong_signal = cdg_cong_signal, 236 .mod_destroy = cdg_mod_destroy, 237 .cc_data_sz = cdg_data_sz, 238 .post_recovery = newreno_cc_post_recovery, 239 .after_idle = newreno_cc_after_idle, 240 }; 241 242 /* Vnet created and being initialised. */ 243 static void 244 cdg_init_vnet(const void *unused __unused) 245 { 246 247 V_cdg_alpha_inc = 0; 248 V_cdg_beta_delay = 70; 249 V_cdg_beta_loss = 50; 250 V_cdg_smoothing_factor = 8; 251 V_cdg_exp_backoff_scale = 3; 252 V_cdg_consec_cong = 5; 253 V_cdg_hold_backoff = 5; 254 } 255 256 static int 257 cdg_mod_init(void) 258 { 259 VNET_ITERATOR_DECL(v); 260 261 ertt_id = khelp_get_id("ertt"); 262 if (ertt_id <= 0) 263 return (EINVAL); 264 265 qdiffsample_zone = uma_zcreate("cdg_qdiffsample", 266 sizeof(struct qdiff_sample), NULL, NULL, NULL, NULL, 0, 0); 267 268 VNET_LIST_RLOCK(); 269 VNET_FOREACH(v) { 270 CURVNET_SET(v); 271 cdg_init_vnet(NULL); 272 CURVNET_RESTORE(); 273 } 274 VNET_LIST_RUNLOCK(); 275 return (0); 276 } 277 278 static int 279 cdg_mod_destroy(void) 280 { 281 282 uma_zdestroy(qdiffsample_zone); 283 return (0); 284 } 285 286 static size_t 287 cdg_data_sz(void) 288 { 289 return (sizeof(struct cdg)); 290 } 291 292 static int 293 cdg_cb_init(struct cc_var *ccv, void *ptr) 294 { 295 struct cdg *cdg_data; 296 297 INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp)); 298 if (ptr == NULL) { 299 cdg_data = malloc(sizeof(struct cdg), M_CC_MEM, M_NOWAIT); 300 if (cdg_data == NULL) 301 return (ENOMEM); 302 } else { 303 cdg_data = ptr; 304 } 305 cdg_data->shadow_w = 0; 306 cdg_data->max_qtrend = 0; 307 cdg_data->min_qtrend = 0; 308 cdg_data->queue_state = CDG_Q_UNKNOWN; 309 cdg_data->maxrtt_in_rtt = 0; 310 cdg_data->maxrtt_in_prevrtt = 0; 311 cdg_data->minrtt_in_rtt = INT_MAX; 312 cdg_data->minrtt_in_prevrtt = 0; 313 cdg_data->window_incr = 0; 314 cdg_data->rtt_count = 0; 315 cdg_data->consec_cong_cnt = 0; 316 cdg_data->sample_q_size = V_cdg_smoothing_factor; 317 cdg_data->num_samples = 0; 318 STAILQ_INIT(&cdg_data->qdiffmin_q); 319 STAILQ_INIT(&cdg_data->qdiffmax_q); 320 321 ccv->cc_data = cdg_data; 322 323 return (0); 324 } 325 326 static void 327 cdg_conn_init(struct cc_var *ccv) 328 { 329 struct cdg *cdg_data = ccv->cc_data; 330 331 /* 332 * Initialise the shadow_cwnd in case we are competing with loss based 333 * flows from the start 334 */ 335 cdg_data->shadow_w = CCV(ccv, snd_cwnd); 336 } 337 338 static void 339 cdg_cb_destroy(struct cc_var *ccv) 340 { 341 struct cdg *cdg_data; 342 struct qdiff_sample *qds, *qds_n; 343 344 cdg_data = ccv->cc_data; 345 346 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q); 347 while (qds != NULL) { 348 qds_n = STAILQ_NEXT(qds, qdiff_lnk); 349 uma_zfree(qdiffsample_zone,qds); 350 qds = qds_n; 351 } 352 353 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q); 354 while (qds != NULL) { 355 qds_n = STAILQ_NEXT(qds, qdiff_lnk); 356 uma_zfree(qdiffsample_zone,qds); 357 qds = qds_n; 358 } 359 360 free(ccv->cc_data, M_CC_MEM); 361 } 362 363 static int 364 cdg_beta_handler(SYSCTL_HANDLER_ARGS) 365 { 366 int error; 367 uint32_t new; 368 369 new = *(uint32_t *)arg1; 370 error = sysctl_handle_int(oidp, &new, 0, req); 371 if (error == 0 && req->newptr != NULL) { 372 if (new == 0 || new > 100) 373 error = EINVAL; 374 else 375 *(uint32_t *)arg1 = new; 376 } 377 378 return (error); 379 } 380 381 static int 382 cdg_exp_backoff_scale_handler(SYSCTL_HANDLER_ARGS) 383 { 384 int error; 385 uint32_t new; 386 387 new = *(uint32_t *)arg1; 388 error = sysctl_handle_int(oidp, &new, 0, req); 389 if (error == 0 && req->newptr != NULL) { 390 if (new < 1) 391 error = EINVAL; 392 else 393 *(uint32_t *)arg1 = new; 394 } 395 396 return (error); 397 } 398 399 static inline uint32_t 400 cdg_window_decrease(struct cc_var *ccv, unsigned long owin, unsigned int beta) 401 { 402 403 return ((ulmin(CCV(ccv, snd_wnd), owin) * beta) / 100); 404 } 405 406 /* 407 * Window increase function 408 * This window increase function is independent of the initial window size 409 * to ensure small window flows are not discriminated against (i.e. fairness). 410 * It increases at 1pkt/rtt like Reno for alpha_inc rtts, and then 2pkts/rtt for 411 * the next alpha_inc rtts, etc. 412 */ 413 static void 414 cdg_window_increase(struct cc_var *ccv, int new_measurement) 415 { 416 struct cdg *cdg_data; 417 int incr, s_w_incr; 418 419 cdg_data = ccv->cc_data; 420 incr = s_w_incr = 0; 421 422 if (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)) { 423 /* Slow start. */ 424 incr = CCV(ccv, t_maxseg); 425 s_w_incr = incr; 426 cdg_data->window_incr = cdg_data->rtt_count = 0; 427 } else { 428 /* Congestion avoidance. */ 429 if (new_measurement) { 430 s_w_incr = CCV(ccv, t_maxseg); 431 if (V_cdg_alpha_inc == 0) { 432 incr = CCV(ccv, t_maxseg); 433 } else { 434 if (++cdg_data->rtt_count >= V_cdg_alpha_inc) { 435 cdg_data->window_incr++; 436 cdg_data->rtt_count = 0; 437 } 438 incr = CCV(ccv, t_maxseg) * 439 cdg_data->window_incr; 440 } 441 } 442 } 443 444 if (cdg_data->shadow_w > 0) 445 cdg_data->shadow_w = ulmin(cdg_data->shadow_w + s_w_incr, 446 TCP_MAXWIN << CCV(ccv, snd_scale)); 447 448 CCV(ccv, snd_cwnd) = ulmin(CCV(ccv, snd_cwnd) + incr, 449 TCP_MAXWIN << CCV(ccv, snd_scale)); 450 } 451 452 static void 453 cdg_cong_signal(struct cc_var *ccv, uint32_t signal_type) 454 { 455 struct cdg *cdg_data = ccv->cc_data; 456 457 switch(signal_type) { 458 case CC_CDG_DELAY: 459 CCV(ccv, snd_ssthresh) = cdg_window_decrease(ccv, 460 CCV(ccv, snd_cwnd), V_cdg_beta_delay); 461 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 462 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 463 cdg_data->window_incr = cdg_data->rtt_count = 0; 464 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 465 break; 466 case CC_NDUPACK: 467 /* 468 * If already responding to congestion OR we have guessed no 469 * queue in the path is full. 470 */ 471 if (IN_CONGRECOVERY(CCV(ccv, t_flags)) || 472 cdg_data->queue_state < CDG_Q_FULL) { 473 CCV(ccv, snd_ssthresh) = CCV(ccv, snd_cwnd); 474 CCV(ccv, snd_recover) = CCV(ccv, snd_max); 475 } else { 476 /* 477 * Loss is likely to be congestion related. We have 478 * inferred a queue full state, so have shadow window 479 * react to loss as NewReno would. 480 */ 481 if (cdg_data->shadow_w > 0) 482 cdg_data->shadow_w = cdg_window_decrease(ccv, 483 cdg_data->shadow_w, RENO_BETA); 484 485 CCV(ccv, snd_ssthresh) = max(cdg_data->shadow_w, 486 cdg_window_decrease(ccv, CCV(ccv, snd_cwnd), 487 V_cdg_beta_loss)); 488 489 cdg_data->window_incr = cdg_data->rtt_count = 0; 490 } 491 ENTER_RECOVERY(CCV(ccv, t_flags)); 492 break; 493 default: 494 newreno_cc_cong_signal(ccv, signal_type); 495 break; 496 } 497 } 498 499 /* 500 * Using a negative exponential probabilistic backoff so that sources with 501 * varying RTTs which share the same link will, on average, have the same 502 * probability of backoff over time. 503 * 504 * Prob_backoff = 1 - exp(-qtrend / V_cdg_exp_backoff_scale), where 505 * V_cdg_exp_backoff_scale is the average qtrend for the exponential backoff. 506 */ 507 static inline int 508 prob_backoff(long qtrend) 509 { 510 int backoff, idx, p; 511 512 backoff = (qtrend > ((MAXGRAD * V_cdg_exp_backoff_scale) << D_P_E)); 513 514 if (!backoff) { 515 if (V_cdg_exp_backoff_scale > 1) 516 idx = (qtrend + V_cdg_exp_backoff_scale / 2) / 517 V_cdg_exp_backoff_scale; 518 else 519 idx = qtrend; 520 521 /* Backoff probability proportional to rate of queue growth. */ 522 p = (INT_MAX / (1 << EXP_PREC)) * probexp[idx]; 523 backoff = (random() < p); 524 } 525 526 return (backoff); 527 } 528 529 static inline void 530 calc_moving_average(struct cdg *cdg_data, long qdiff_max, long qdiff_min) 531 { 532 struct qdiff_sample *qds; 533 534 ++cdg_data->num_samples; 535 if (cdg_data->num_samples > cdg_data->sample_q_size) { 536 /* Minimum RTT. */ 537 qds = STAILQ_FIRST(&cdg_data->qdiffmin_q); 538 cdg_data->min_qtrend = cdg_data->min_qtrend + 539 (qdiff_min - qds->qdiff) / cdg_data->sample_q_size; 540 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmin_q, qdiff_lnk); 541 qds->qdiff = qdiff_min; 542 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, qdiff_lnk); 543 544 /* Maximum RTT. */ 545 qds = STAILQ_FIRST(&cdg_data->qdiffmax_q); 546 cdg_data->max_qtrend = cdg_data->max_qtrend + 547 (qdiff_max - qds->qdiff) / cdg_data->sample_q_size; 548 STAILQ_REMOVE_HEAD(&cdg_data->qdiffmax_q, qdiff_lnk); 549 qds->qdiff = qdiff_max; 550 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, qdiff_lnk); 551 --cdg_data->num_samples; 552 } else { 553 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT); 554 if (qds != NULL) { 555 cdg_data->min_qtrend = cdg_data->min_qtrend + 556 qdiff_min / cdg_data->sample_q_size; 557 qds->qdiff = qdiff_min; 558 STAILQ_INSERT_TAIL(&cdg_data->qdiffmin_q, qds, 559 qdiff_lnk); 560 } 561 562 qds = uma_zalloc(qdiffsample_zone, M_NOWAIT); 563 if (qds) { 564 cdg_data->max_qtrend = cdg_data->max_qtrend + 565 qdiff_max / cdg_data->sample_q_size; 566 qds->qdiff = qdiff_max; 567 STAILQ_INSERT_TAIL(&cdg_data->qdiffmax_q, qds, 568 qdiff_lnk); 569 } 570 } 571 } 572 573 static void 574 cdg_ack_received(struct cc_var *ccv, uint16_t ack_type) 575 { 576 struct cdg *cdg_data; 577 struct ertt *e_t; 578 long qdiff_max, qdiff_min; 579 int congestion, new_measurement, slowstart; 580 581 cdg_data = ccv->cc_data; 582 e_t = (struct ertt *)khelp_get_osd(&CCV(ccv, t_osd), ertt_id); 583 new_measurement = e_t->flags & ERTT_NEW_MEASUREMENT; 584 congestion = 0; 585 cdg_data->maxrtt_in_rtt = imax(e_t->rtt, cdg_data->maxrtt_in_rtt); 586 cdg_data->minrtt_in_rtt = imin(e_t->rtt, cdg_data->minrtt_in_rtt); 587 588 if (new_measurement) { 589 slowstart = (CCV(ccv, snd_cwnd) <= CCV(ccv, snd_ssthresh)); 590 /* 591 * Update smoothed gradient measurements. Since we are only 592 * using one measurement per RTT, use max or min rtt_in_rtt. 593 * This is also less noisy than a sample RTT measurement. Max 594 * RTT measurements can have trouble due to OS issues. 595 */ 596 if (cdg_data->maxrtt_in_prevrtt) { 597 qdiff_max = ((long)(cdg_data->maxrtt_in_rtt - 598 cdg_data->maxrtt_in_prevrtt) << D_P_E ); 599 qdiff_min = ((long)(cdg_data->minrtt_in_rtt - 600 cdg_data->minrtt_in_prevrtt) << D_P_E ); 601 602 if (cdg_data->sample_q_size == 0) { 603 cdg_data->max_qtrend = qdiff_max; 604 cdg_data->min_qtrend = qdiff_min; 605 } else 606 calc_moving_average(cdg_data, qdiff_max, qdiff_min); 607 608 /* Probabilistic backoff with respect to gradient. */ 609 if (slowstart && qdiff_min > 0) 610 congestion = prob_backoff(qdiff_min); 611 else if (cdg_data->min_qtrend > 0) 612 congestion = prob_backoff(cdg_data->min_qtrend); 613 else if (slowstart && qdiff_max > 0) 614 congestion = prob_backoff(qdiff_max); 615 else if (cdg_data->max_qtrend > 0) 616 congestion = prob_backoff(cdg_data->max_qtrend); 617 618 /* Update estimate of queue state. */ 619 if (cdg_data->min_qtrend > 0 && 620 cdg_data->max_qtrend <= 0) { 621 cdg_data->queue_state = CDG_Q_FULL; 622 } else if (cdg_data->min_qtrend >= 0 && 623 cdg_data->max_qtrend < 0) { 624 cdg_data->queue_state = CDG_Q_EMPTY; 625 cdg_data->shadow_w = 0; 626 } else if (cdg_data->min_qtrend > 0 && 627 cdg_data->max_qtrend > 0) { 628 cdg_data->queue_state = CDG_Q_RISING; 629 } else if (cdg_data->min_qtrend < 0 && 630 cdg_data->max_qtrend < 0) { 631 cdg_data->queue_state = CDG_Q_FALLING; 632 } 633 634 if (cdg_data->min_qtrend < 0 || 635 cdg_data->max_qtrend < 0) 636 cdg_data->consec_cong_cnt = 0; 637 } 638 639 cdg_data->minrtt_in_prevrtt = cdg_data->minrtt_in_rtt; 640 cdg_data->minrtt_in_rtt = INT_MAX; 641 cdg_data->maxrtt_in_prevrtt = cdg_data->maxrtt_in_rtt; 642 cdg_data->maxrtt_in_rtt = 0; 643 e_t->flags &= ~ERTT_NEW_MEASUREMENT; 644 } 645 646 if (congestion) { 647 cdg_data->consec_cong_cnt++; 648 if (!IN_RECOVERY(CCV(ccv, t_flags))) { 649 if (cdg_data->consec_cong_cnt <= V_cdg_consec_cong) 650 cdg_cong_signal(ccv, CC_CDG_DELAY); 651 else 652 /* 653 * We have been backing off but the queue is not 654 * falling. Assume we are competing with 655 * loss-based flows and don't back off for the 656 * next V_cdg_hold_backoff RTT periods. 657 */ 658 if (cdg_data->consec_cong_cnt >= 659 V_cdg_consec_cong + V_cdg_hold_backoff) 660 cdg_data->consec_cong_cnt = 0; 661 662 /* Won't see effect until 2nd RTT. */ 663 cdg_data->maxrtt_in_prevrtt = 0; 664 /* 665 * Resync shadow window in case we are competing with a 666 * loss based flow 667 */ 668 cdg_data->shadow_w = ulmax(CCV(ccv, snd_cwnd), 669 cdg_data->shadow_w); 670 } 671 } else if (ack_type == CC_ACK) 672 cdg_window_increase(ccv, new_measurement); 673 } 674 675 /* When a vnet is created and being initialised, init the per-stack CDG vars. */ 676 VNET_SYSINIT(cdg_init_vnet, SI_SUB_PROTO_BEGIN, SI_ORDER_FIRST, 677 cdg_init_vnet, NULL); 678 679 SYSCTL_DECL(_net_inet_tcp_cc_cdg); 680 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, cdg, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 681 "CAIA delay-gradient congestion control related settings"); 682 683 SYSCTL_STRING(_net_inet_tcp_cc_cdg, OID_AUTO, version, 684 CTLFLAG_RD, CDG_VERSION, sizeof(CDG_VERSION) - 1, 685 "Current algorithm/implementation version number"); 686 687 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, alpha_inc, 688 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_alpha_inc), 0, 689 "Increment the window increase factor alpha by 1 MSS segment every " 690 "alpha_inc RTTs during congestion avoidance mode."); 691 692 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_delay, 693 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 694 &VNET_NAME(cdg_beta_delay), 70, &cdg_beta_handler, "IU", 695 "Delay-based window decrease factor as a percentage " 696 "(on delay-based backoff, w = w * beta_delay / 100)"); 697 698 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, beta_loss, 699 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 700 &VNET_NAME(cdg_beta_loss), 50, &cdg_beta_handler, "IU", 701 "Loss-based window decrease factor as a percentage " 702 "(on loss-based backoff, w = w * beta_loss / 100)"); 703 704 SYSCTL_PROC(_net_inet_tcp_cc_cdg, OID_AUTO, exp_backoff_scale, 705 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 706 &VNET_NAME(cdg_exp_backoff_scale), 2, &cdg_exp_backoff_scale_handler, "IU", 707 "Scaling parameter for the probabilistic exponential backoff"); 708 709 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, smoothing_factor, 710 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_smoothing_factor), 8, 711 "Number of samples used for moving average smoothing (0 = no smoothing)"); 712 713 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_consec_cong, 714 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_consec_cong), 5, 715 "Number of consecutive delay-gradient based congestion episodes which will " 716 "trigger loss based CC compatibility"); 717 718 SYSCTL_UINT(_net_inet_tcp_cc_cdg, OID_AUTO, loss_compete_hold_backoff, 719 CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(cdg_hold_backoff), 5, 720 "Number of consecutive delay-gradient based congestion episodes to hold " 721 "the window backoff for loss based CC compatibility"); 722 723 DECLARE_CC_MODULE(cdg, &cdg_cc_algo); 724 MODULE_VERSION(cdg, 2); 725 MODULE_DEPEND(cdg, ertt, 1, 1, 1); 726