1/* 2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.3 3 * Home page: 4 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC 5 * This is from the implementation of CUBIC TCP in 6 * Sangtae Ha, Injong Rhee and Lisong Xu, 7 * "CUBIC: A New TCP-Friendly High-Speed TCP Variant" 8 * in ACM SIGOPS Operating System Review, July 2008. 9 * Available from: 10 * http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf 11 * 12 * CUBIC integrates a new slow start algorithm, called HyStart. 13 * The details of HyStart are presented in 14 * Sangtae Ha and Injong Rhee, 15 * "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008. 16 * Available from: 17 * http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf 18 * 19 * All testing results are available from: 20 * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing 21 * 22 * Unless CUBIC is enabled and congestion window is large 23 * this behaves the same as the original Reno. 24 */ 25 26#include <linux/mm.h> 27#include <linux/module.h> 28#include <linux/math64.h> 29#include <net/tcp.h> 30 31#define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation 32 * max_cwnd = snd_cwnd * beta 33 */ 34#define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ 35 36/* Two methods of hybrid slow start */ 37#define HYSTART_ACK_TRAIN 0x1 38#define HYSTART_DELAY 0x2 39 40/* Number of delay samples for detecting the increase of delay */ 41#define HYSTART_MIN_SAMPLES 8 42#define HYSTART_DELAY_MIN (4U<<3) 43#define HYSTART_DELAY_MAX (16U<<3) 44#define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) 45 46static int fast_convergence __read_mostly = 1; 47static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ 48static int initial_ssthresh __read_mostly; 49static int bic_scale __read_mostly = 41; 50static int tcp_friendliness __read_mostly = 1; 51 52static int hystart __read_mostly = 1; 53static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY; 54static int hystart_low_window __read_mostly = 16; 55static int hystart_ack_delta __read_mostly = 2; 56 57static u32 cube_rtt_scale __read_mostly; 58static u32 beta_scale __read_mostly; 59static u64 cube_factor __read_mostly; 60 61/* Note parameters that are used for precomputing scale factors are read-only */ 62module_param(fast_convergence, int, 0644); 63MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence"); 64module_param(beta, int, 0644); 65MODULE_PARM_DESC(beta, "beta for multiplicative increase"); 66module_param(initial_ssthresh, int, 0644); 67MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold"); 68module_param(bic_scale, int, 0444); 69MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)"); 70module_param(tcp_friendliness, int, 0644); 71MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness"); 72module_param(hystart, int, 0644); 73MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm"); 74module_param(hystart_detect, int, 0644); 75MODULE_PARM_DESC(hystart_detect, "hyrbrid slow start detection mechanisms" 76 " 1: packet-train 2: delay 3: both packet-train and delay"); 77module_param(hystart_low_window, int, 0644); 78MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start"); 79module_param(hystart_ack_delta, int, 0644); 80MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)"); 81 82/* BIC TCP Parameters */ 83struct bictcp { 84 u32 cnt; /* increase cwnd by 1 after ACKs */ 85 u32 last_max_cwnd; /* last maximum snd_cwnd */ 86 u32 loss_cwnd; /* congestion window at last loss */ 87 u32 last_cwnd; /* the last snd_cwnd */ 88 u32 last_time; /* time when updated last_cwnd */ 89 u32 bic_origin_point;/* origin point of bic function */ 90 u32 bic_K; /* time to origin point 91 from the beginning of the current epoch */ 92 u32 delay_min; /* min delay (msec << 3) */ 93 u32 epoch_start; /* beginning of an epoch */ 94 u32 ack_cnt; /* number of acks */ 95 u32 tcp_cwnd; /* estimated tcp cwnd */ 96 u16 unused; 97 u8 sample_cnt; /* number of samples to decide curr_rtt */ 98 u8 found; /* the exit point is found? */ 99 u32 round_start; /* beginning of each round */ 100 u32 end_seq; /* end_seq of the round */ 101 u32 last_ack; /* last time when the ACK spacing is close */ 102 u32 curr_rtt; /* the minimum rtt of current round */ 103}; 104 105static inline void bictcp_reset(struct bictcp *ca) 106{ 107 ca->cnt = 0; 108 ca->last_max_cwnd = 0; 109 ca->last_cwnd = 0; 110 ca->last_time = 0; 111 ca->bic_origin_point = 0; 112 ca->bic_K = 0; 113 ca->delay_min = 0; 114 ca->epoch_start = 0; 115 ca->ack_cnt = 0; 116 ca->tcp_cwnd = 0; 117 ca->found = 0; 118} 119 120static inline u32 bictcp_clock(void) 121{ 122#if HZ < 1000 123 return ktime_to_ms(ktime_get_real()); 124#else 125 return jiffies_to_msecs(jiffies); 126#endif 127} 128 129static inline void bictcp_hystart_reset(struct sock *sk) 130{ 131 struct tcp_sock *tp = tcp_sk(sk); 132 struct bictcp *ca = inet_csk_ca(sk); 133 134 ca->round_start = ca->last_ack = bictcp_clock(); 135 ca->end_seq = tp->snd_nxt; 136 ca->curr_rtt = 0; 137 ca->sample_cnt = 0; 138} 139 140static void bictcp_init(struct sock *sk) 141{ 142 struct bictcp *ca = inet_csk_ca(sk); 143 144 bictcp_reset(ca); 145 ca->loss_cwnd = 0; 146 147 if (hystart) 148 bictcp_hystart_reset(sk); 149 150 if (!hystart && initial_ssthresh) 151 tcp_sk(sk)->snd_ssthresh = initial_ssthresh; 152} 153 154static void bictcp_cwnd_event(struct sock *sk, enum tcp_ca_event event) 155{ 156 if (event == CA_EVENT_TX_START) { 157 struct bictcp *ca = inet_csk_ca(sk); 158 u32 now = tcp_time_stamp; 159 s32 delta; 160 161 delta = now - tcp_sk(sk)->lsndtime; 162 163 /* We were application limited (idle) for a while. 164 * Shift epoch_start to keep cwnd growth to cubic curve. 165 */ 166 if (ca->epoch_start && delta > 0) { 167 ca->epoch_start += delta; 168 if (after(ca->epoch_start, now)) 169 ca->epoch_start = now; 170 } 171 return; 172 } 173} 174 175/* calculate the cubic root of x using a table lookup followed by one 176 * Newton-Raphson iteration. 177 * Avg err ~= 0.195% 178 */ 179static u32 cubic_root(u64 a) 180{ 181 u32 x, b, shift; 182 /* 183 * cbrt(x) MSB values for x MSB values in [0..63]. 184 * Precomputed then refined by hand - Willy Tarreau 185 * 186 * For x in [0..63], 187 * v = cbrt(x << 18) - 1 188 * cbrt(x) = (v[x] + 10) >> 6 189 */ 190 static const u8 v[] = { 191 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, 192 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, 193 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, 194 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, 195 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, 196 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, 197 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, 198 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, 199 }; 200 201 b = fls64(a); 202 if (b < 7) { 203 /* a in [0..63] */ 204 return ((u32)v[(u32)a] + 35) >> 6; 205 } 206 207 b = ((b * 84) >> 8) - 1; 208 shift = (a >> (b * 3)); 209 210 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; 211 212 /* 213 * Newton-Raphson iteration 214 * 2 215 * x = ( 2 * x + a / x ) / 3 216 * k+1 k k 217 */ 218 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); 219 x = ((x * 341) >> 10); 220 return x; 221} 222 223/* 224 * Compute congestion window to use. 225 */ 226static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked) 227{ 228 u32 delta, bic_target, max_cnt; 229 u64 offs, t; 230 231 ca->ack_cnt += acked; /* count the number of ACKed packets */ 232 233 if (ca->last_cwnd == cwnd && 234 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32) 235 return; 236 237 /* The CUBIC function can update ca->cnt at most once per jiffy. 238 * On all cwnd reduction events, ca->epoch_start is set to 0, 239 * which will force a recalculation of ca->cnt. 240 */ 241 if (ca->epoch_start && tcp_time_stamp == ca->last_time) 242 goto tcp_friendliness; 243 244 ca->last_cwnd = cwnd; 245 ca->last_time = tcp_time_stamp; 246 247 if (ca->epoch_start == 0) { 248 ca->epoch_start = tcp_time_stamp; /* record beginning */ 249 ca->ack_cnt = acked; /* start counting */ 250 ca->tcp_cwnd = cwnd; /* syn with cubic */ 251 252 if (ca->last_max_cwnd <= cwnd) { 253 ca->bic_K = 0; 254 ca->bic_origin_point = cwnd; 255 } else { 256 /* Compute new K based on 257 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) 258 */ 259 ca->bic_K = cubic_root(cube_factor 260 * (ca->last_max_cwnd - cwnd)); 261 ca->bic_origin_point = ca->last_max_cwnd; 262 } 263 } 264 265 /* cubic function - calc*/ 266 /* calculate c * time^3 / rtt, 267 * while considering overflow in calculation of time^3 268 * (so time^3 is done by using 64 bit) 269 * and without the support of division of 64bit numbers 270 * (so all divisions are done by using 32 bit) 271 * also NOTE the unit of those veriables 272 * time = (t - K) / 2^bictcp_HZ 273 * c = bic_scale >> 10 274 * rtt = (srtt >> 3) / HZ 275 * !!! The following code does not have overflow problems, 276 * if the cwnd < 1 million packets !!! 277 */ 278 279 t = (s32)(tcp_time_stamp - ca->epoch_start); 280 t += msecs_to_jiffies(ca->delay_min >> 3); 281 /* change the unit from HZ to bictcp_HZ */ 282 t <<= BICTCP_HZ; 283 do_div(t, HZ); 284 285 if (t < ca->bic_K) /* t - K */ 286 offs = ca->bic_K - t; 287 else 288 offs = t - ca->bic_K; 289 290 /* c/rtt * (t-K)^3 */ 291 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); 292 if (t < ca->bic_K) /* below origin*/ 293 bic_target = ca->bic_origin_point - delta; 294 else /* above origin*/ 295 bic_target = ca->bic_origin_point + delta; 296 297 /* cubic function - calc bictcp_cnt*/ 298 if (bic_target > cwnd) { 299 ca->cnt = cwnd / (bic_target - cwnd); 300 } else { 301 ca->cnt = 100 * cwnd; /* very small increment*/ 302 } 303 304 /* 305 * The initial growth of cubic function may be too conservative 306 * when the available bandwidth is still unknown. 307 */ 308 if (ca->last_max_cwnd == 0 && ca->cnt > 20) 309 ca->cnt = 20; /* increase cwnd 5% per RTT */ 310 311tcp_friendliness: 312 /* TCP Friendly */ 313 if (tcp_friendliness) { 314 u32 scale = beta_scale; 315 316 delta = (cwnd * scale) >> 3; 317 while (ca->ack_cnt > delta) { /* update tcp cwnd */ 318 ca->ack_cnt -= delta; 319 ca->tcp_cwnd++; 320 } 321 322 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ 323 delta = ca->tcp_cwnd - cwnd; 324 max_cnt = cwnd / delta; 325 if (ca->cnt > max_cnt) 326 ca->cnt = max_cnt; 327 } 328 } 329 330 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per 331 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. 332 */ 333 ca->cnt = max(ca->cnt, 2U); 334} 335 336static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 337{ 338 struct tcp_sock *tp = tcp_sk(sk); 339 struct bictcp *ca = inet_csk_ca(sk); 340 341 if (!tcp_is_cwnd_limited(sk)) 342 return; 343 344 if (tcp_in_slow_start(tp)) { 345 if (hystart && after(ack, ca->end_seq)) 346 bictcp_hystart_reset(sk); 347 acked = tcp_slow_start(tp, acked); 348 if (!acked) 349 return; 350 } 351 bictcp_update(ca, tp->snd_cwnd, acked); 352 tcp_cong_avoid_ai(tp, ca->cnt, acked); 353} 354 355static u32 bictcp_recalc_ssthresh(struct sock *sk) 356{ 357 const struct tcp_sock *tp = tcp_sk(sk); 358 struct bictcp *ca = inet_csk_ca(sk); 359 360 ca->epoch_start = 0; /* end of epoch */ 361 362 /* Wmax and fast convergence */ 363 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) 364 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) 365 / (2 * BICTCP_BETA_SCALE); 366 else 367 ca->last_max_cwnd = tp->snd_cwnd; 368 369 ca->loss_cwnd = tp->snd_cwnd; 370 371 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); 372} 373 374static u32 bictcp_undo_cwnd(struct sock *sk) 375{ 376 struct bictcp *ca = inet_csk_ca(sk); 377 378 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd); 379} 380 381static void bictcp_state(struct sock *sk, u8 new_state) 382{ 383 if (new_state == TCP_CA_Loss) { 384 bictcp_reset(inet_csk_ca(sk)); 385 bictcp_hystart_reset(sk); 386 } 387} 388 389static void hystart_update(struct sock *sk, u32 delay) 390{ 391 struct tcp_sock *tp = tcp_sk(sk); 392 struct bictcp *ca = inet_csk_ca(sk); 393 394 if (ca->found & hystart_detect) 395 return; 396 397 if (hystart_detect & HYSTART_ACK_TRAIN) { 398 u32 now = bictcp_clock(); 399 400 /* first detection parameter - ack-train detection */ 401 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) { 402 ca->last_ack = now; 403 if ((s32)(now - ca->round_start) > ca->delay_min >> 4) { 404 ca->found |= HYSTART_ACK_TRAIN; 405 NET_INC_STATS_BH(sock_net(sk), 406 LINUX_MIB_TCPHYSTARTTRAINDETECT); 407 NET_ADD_STATS_BH(sock_net(sk), 408 LINUX_MIB_TCPHYSTARTTRAINCWND, 409 tp->snd_cwnd); 410 tp->snd_ssthresh = tp->snd_cwnd; 411 } 412 } 413 } 414 415 if (hystart_detect & HYSTART_DELAY) { 416 /* obtain the minimum delay of more than sampling packets */ 417 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { 418 if (ca->curr_rtt == 0 || ca->curr_rtt > delay) 419 ca->curr_rtt = delay; 420 421 ca->sample_cnt++; 422 } else { 423 if (ca->curr_rtt > ca->delay_min + 424 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { 425 ca->found |= HYSTART_DELAY; 426 NET_INC_STATS_BH(sock_net(sk), 427 LINUX_MIB_TCPHYSTARTDELAYDETECT); 428 NET_ADD_STATS_BH(sock_net(sk), 429 LINUX_MIB_TCPHYSTARTDELAYCWND, 430 tp->snd_cwnd); 431 tp->snd_ssthresh = tp->snd_cwnd; 432 } 433 } 434 } 435} 436 437/* Track delayed acknowledgment ratio using sliding window 438 * ratio = (15*ratio + sample) / 16 439 */ 440static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us) 441{ 442 const struct tcp_sock *tp = tcp_sk(sk); 443 struct bictcp *ca = inet_csk_ca(sk); 444 u32 delay; 445 446 /* Some calls are for duplicates without timetamps */ 447 if (rtt_us < 0) 448 return; 449 450 /* Discard delay samples right after fast recovery */ 451 if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ) 452 return; 453 454 delay = (rtt_us << 3) / USEC_PER_MSEC; 455 if (delay == 0) 456 delay = 1; 457 458 /* first time call or link delay decreases */ 459 if (ca->delay_min == 0 || ca->delay_min > delay) 460 ca->delay_min = delay; 461 462 /* hystart triggers when cwnd is larger than some threshold */ 463 if (hystart && tcp_in_slow_start(tp) && 464 tp->snd_cwnd >= hystart_low_window) 465 hystart_update(sk, delay); 466} 467 468static struct tcp_congestion_ops cubictcp __read_mostly = { 469 .init = bictcp_init, 470 .ssthresh = bictcp_recalc_ssthresh, 471 .cong_avoid = bictcp_cong_avoid, 472 .set_state = bictcp_state, 473 .undo_cwnd = bictcp_undo_cwnd, 474 .cwnd_event = bictcp_cwnd_event, 475 .pkts_acked = bictcp_acked, 476 .owner = THIS_MODULE, 477 .name = "cubic", 478}; 479 480static int __init cubictcp_register(void) 481{ 482 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); 483 484 /* Precompute a bunch of the scaling factors that are used per-packet 485 * based on SRTT of 100ms 486 */ 487 488 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 489 / (BICTCP_BETA_SCALE - beta); 490 491 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ 492 493 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 494 * so K = cubic_root( (wmax-cwnd)*rtt/c ) 495 * the unit of K is bictcp_HZ=2^10, not HZ 496 * 497 * c = bic_scale >> 10 498 * rtt = 100ms 499 * 500 * the following code has been designed and tested for 501 * cwnd < 1 million packets 502 * RTT < 100 seconds 503 * HZ < 1,000,00 (corresponding to 10 nano-second) 504 */ 505 506 /* 1/c * 2^2*bictcp_HZ * srtt */ 507 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ 508 509 /* divide by bic_scale and by constant Srtt (100ms) */ 510 do_div(cube_factor, bic_scale * 10); 511 512 return tcp_register_congestion_control(&cubictcp); 513} 514 515static void __exit cubictcp_unregister(void) 516{ 517 tcp_unregister_congestion_control(&cubictcp); 518} 519 520module_init(cubictcp_register); 521module_exit(cubictcp_unregister); 522 523MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); 524MODULE_LICENSE("GPL"); 525MODULE_DESCRIPTION("CUBIC TCP"); 526MODULE_VERSION("2.3"); 527