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#define ACK_RATIO_SHIFT 4 97#define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT) 98 u16 delayed_ack; /* estimate the ratio of Packets/ACKs << 4 */ 99 u8 sample_cnt; /* number of samples to decide curr_rtt */ 100 u8 found; /* the exit point is found? */ 101 u32 round_start; /* beginning of each round */ 102 u32 end_seq; /* end_seq of the round */ 103 u32 last_ack; /* last time when the ACK spacing is close */ 104 u32 curr_rtt; /* the minimum rtt of current round */ 105}; 106 107static inline void bictcp_reset(struct bictcp *ca) 108{ 109 ca->cnt = 0; 110 ca->last_max_cwnd = 0; 111 ca->last_cwnd = 0; 112 ca->last_time = 0; 113 ca->bic_origin_point = 0; 114 ca->bic_K = 0; 115 ca->delay_min = 0; 116 ca->epoch_start = 0; 117 ca->delayed_ack = 2 << ACK_RATIO_SHIFT; 118 ca->ack_cnt = 0; 119 ca->tcp_cwnd = 0; 120 ca->found = 0; 121} 122 123static inline u32 bictcp_clock(void) 124{ 125#if HZ < 1000 126 return ktime_to_ms(ktime_get_real()); 127#else 128 return jiffies_to_msecs(jiffies); 129#endif 130} 131 132static inline void bictcp_hystart_reset(struct sock *sk) 133{ 134 struct tcp_sock *tp = tcp_sk(sk); 135 struct bictcp *ca = inet_csk_ca(sk); 136 137 ca->round_start = ca->last_ack = bictcp_clock(); 138 ca->end_seq = tp->snd_nxt; 139 ca->curr_rtt = 0; 140 ca->sample_cnt = 0; 141} 142 143static void bictcp_init(struct sock *sk) 144{ 145 struct bictcp *ca = inet_csk_ca(sk); 146 147 bictcp_reset(ca); 148 ca->loss_cwnd = 0; 149 150 if (hystart) 151 bictcp_hystart_reset(sk); 152 153 if (!hystart && initial_ssthresh) 154 tcp_sk(sk)->snd_ssthresh = initial_ssthresh; 155} 156 157/* calculate the cubic root of x using a table lookup followed by one 158 * Newton-Raphson iteration. 159 * Avg err ~= 0.195% 160 */ 161static u32 cubic_root(u64 a) 162{ 163 u32 x, b, shift; 164 /* 165 * cbrt(x) MSB values for x MSB values in [0..63]. 166 * Precomputed then refined by hand - Willy Tarreau 167 * 168 * For x in [0..63], 169 * v = cbrt(x << 18) - 1 170 * cbrt(x) = (v[x] + 10) >> 6 171 */ 172 static const u8 v[] = { 173 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, 174 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, 175 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, 176 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, 177 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, 178 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, 179 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, 180 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, 181 }; 182 183 b = fls64(a); 184 if (b < 7) { 185 /* a in [0..63] */ 186 return ((u32)v[(u32)a] + 35) >> 6; 187 } 188 189 b = ((b * 84) >> 8) - 1; 190 shift = (a >> (b * 3)); 191 192 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; 193 194 /* 195 * Newton-Raphson iteration 196 * 2 197 * x = ( 2 * x + a / x ) / 3 198 * k+1 k k 199 */ 200 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); 201 x = ((x * 341) >> 10); 202 return x; 203} 204 205/* 206 * Compute congestion window to use. 207 */ 208static inline void bictcp_update(struct bictcp *ca, u32 cwnd) 209{ 210 u32 delta, bic_target, max_cnt; 211 u64 offs, t; 212 213 ca->ack_cnt++; /* count the number of ACKs */ 214 215 if (ca->last_cwnd == cwnd && 216 (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32) 217 return; 218 219 ca->last_cwnd = cwnd; 220 ca->last_time = tcp_time_stamp; 221 222 if (ca->epoch_start == 0) { 223 ca->epoch_start = tcp_time_stamp; /* record beginning */ 224 ca->ack_cnt = 1; /* start counting */ 225 ca->tcp_cwnd = cwnd; /* syn with cubic */ 226 227 if (ca->last_max_cwnd <= cwnd) { 228 ca->bic_K = 0; 229 ca->bic_origin_point = cwnd; 230 } else { 231 /* Compute new K based on 232 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) 233 */ 234 ca->bic_K = cubic_root(cube_factor 235 * (ca->last_max_cwnd - cwnd)); 236 ca->bic_origin_point = ca->last_max_cwnd; 237 } 238 } 239 240 /* cubic function - calc*/ 241 /* calculate c * time^3 / rtt, 242 * while considering overflow in calculation of time^3 243 * (so time^3 is done by using 64 bit) 244 * and without the support of division of 64bit numbers 245 * (so all divisions are done by using 32 bit) 246 * also NOTE the unit of those veriables 247 * time = (t - K) / 2^bictcp_HZ 248 * c = bic_scale >> 10 249 * rtt = (srtt >> 3) / HZ 250 * !!! The following code does not have overflow problems, 251 * if the cwnd < 1 million packets !!! 252 */ 253 254 t = (s32)(tcp_time_stamp - ca->epoch_start); 255 t += msecs_to_jiffies(ca->delay_min >> 3); 256 /* change the unit from HZ to bictcp_HZ */ 257 t <<= BICTCP_HZ; 258 do_div(t, HZ); 259 260 if (t < ca->bic_K) /* t - K */ 261 offs = ca->bic_K - t; 262 else 263 offs = t - ca->bic_K; 264 265 /* c/rtt * (t-K)^3 */ 266 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); 267 if (t < ca->bic_K) /* below origin*/ 268 bic_target = ca->bic_origin_point - delta; 269 else /* above origin*/ 270 bic_target = ca->bic_origin_point + delta; 271 272 /* cubic function - calc bictcp_cnt*/ 273 if (bic_target > cwnd) { 274 ca->cnt = cwnd / (bic_target - cwnd); 275 } else { 276 ca->cnt = 100 * cwnd; /* very small increment*/ 277 } 278 279 /* 280 * The initial growth of cubic function may be too conservative 281 * when the available bandwidth is still unknown. 282 */ 283 if (ca->last_max_cwnd == 0 && ca->cnt > 20) 284 ca->cnt = 20; /* increase cwnd 5% per RTT */ 285 286 /* TCP Friendly */ 287 if (tcp_friendliness) { 288 u32 scale = beta_scale; 289 290 delta = (cwnd * scale) >> 3; 291 while (ca->ack_cnt > delta) { /* update tcp cwnd */ 292 ca->ack_cnt -= delta; 293 ca->tcp_cwnd++; 294 } 295 296 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ 297 delta = ca->tcp_cwnd - cwnd; 298 max_cnt = cwnd / delta; 299 if (ca->cnt > max_cnt) 300 ca->cnt = max_cnt; 301 } 302 } 303 304 ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack; 305 if (ca->cnt == 0) /* cannot be zero */ 306 ca->cnt = 1; 307} 308 309static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 310{ 311 struct tcp_sock *tp = tcp_sk(sk); 312 struct bictcp *ca = inet_csk_ca(sk); 313 314 if (!tcp_is_cwnd_limited(sk)) 315 return; 316 317 if (tp->snd_cwnd <= tp->snd_ssthresh) { 318 if (hystart && after(ack, ca->end_seq)) 319 bictcp_hystart_reset(sk); 320 tcp_slow_start(tp, acked); 321 } else { 322 bictcp_update(ca, tp->snd_cwnd); 323 tcp_cong_avoid_ai(tp, ca->cnt); 324 } 325} 326 327static u32 bictcp_recalc_ssthresh(struct sock *sk) 328{ 329 const struct tcp_sock *tp = tcp_sk(sk); 330 struct bictcp *ca = inet_csk_ca(sk); 331 332 ca->epoch_start = 0; /* end of epoch */ 333 334 /* Wmax and fast convergence */ 335 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) 336 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) 337 / (2 * BICTCP_BETA_SCALE); 338 else 339 ca->last_max_cwnd = tp->snd_cwnd; 340 341 ca->loss_cwnd = tp->snd_cwnd; 342 343 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); 344} 345 346static u32 bictcp_undo_cwnd(struct sock *sk) 347{ 348 struct bictcp *ca = inet_csk_ca(sk); 349 350 return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd); 351} 352 353static void bictcp_state(struct sock *sk, u8 new_state) 354{ 355 if (new_state == TCP_CA_Loss) { 356 bictcp_reset(inet_csk_ca(sk)); 357 bictcp_hystart_reset(sk); 358 } 359} 360 361static void hystart_update(struct sock *sk, u32 delay) 362{ 363 struct tcp_sock *tp = tcp_sk(sk); 364 struct bictcp *ca = inet_csk_ca(sk); 365 366 if (!(ca->found & hystart_detect)) { 367 u32 now = bictcp_clock(); 368 369 /* first detection parameter - ack-train detection */ 370 if ((s32)(now - ca->last_ack) <= hystart_ack_delta) { 371 ca->last_ack = now; 372 if ((s32)(now - ca->round_start) > ca->delay_min >> 4) 373 ca->found |= HYSTART_ACK_TRAIN; 374 } 375 376 /* obtain the minimum delay of more than sampling packets */ 377 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { 378 if (ca->curr_rtt == 0 || ca->curr_rtt > delay) 379 ca->curr_rtt = delay; 380 381 ca->sample_cnt++; 382 } else { 383 if (ca->curr_rtt > ca->delay_min + 384 HYSTART_DELAY_THRESH(ca->delay_min>>4)) 385 ca->found |= HYSTART_DELAY; 386 } 387 /* 388 * Either one of two conditions are met, 389 * we exit from slow start immediately. 390 */ 391 if (ca->found & hystart_detect) 392 tp->snd_ssthresh = tp->snd_cwnd; 393 } 394} 395 396/* Track delayed acknowledgment ratio using sliding window 397 * ratio = (15*ratio + sample) / 16 398 */ 399static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us) 400{ 401 const struct inet_connection_sock *icsk = inet_csk(sk); 402 const struct tcp_sock *tp = tcp_sk(sk); 403 struct bictcp *ca = inet_csk_ca(sk); 404 u32 delay; 405 406 if (icsk->icsk_ca_state == TCP_CA_Open) { 407 u32 ratio = ca->delayed_ack; 408 409 ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT; 410 ratio += cnt; 411 412 ca->delayed_ack = clamp(ratio, 1U, ACK_RATIO_LIMIT); 413 } 414 415 /* Some calls are for duplicates without timetamps */ 416 if (rtt_us < 0) 417 return; 418 419 /* Discard delay samples right after fast recovery */ 420 if (ca->epoch_start && (s32)(tcp_time_stamp - ca->epoch_start) < HZ) 421 return; 422 423 delay = (rtt_us << 3) / USEC_PER_MSEC; 424 if (delay == 0) 425 delay = 1; 426 427 /* first time call or link delay decreases */ 428 if (ca->delay_min == 0 || ca->delay_min > delay) 429 ca->delay_min = delay; 430 431 /* hystart triggers when cwnd is larger than some threshold */ 432 if (hystart && tp->snd_cwnd <= tp->snd_ssthresh && 433 tp->snd_cwnd >= hystart_low_window) 434 hystart_update(sk, delay); 435} 436 437static struct tcp_congestion_ops cubictcp __read_mostly = { 438 .init = bictcp_init, 439 .ssthresh = bictcp_recalc_ssthresh, 440 .cong_avoid = bictcp_cong_avoid, 441 .set_state = bictcp_state, 442 .undo_cwnd = bictcp_undo_cwnd, 443 .pkts_acked = bictcp_acked, 444 .owner = THIS_MODULE, 445 .name = "cubic", 446}; 447 448static int __init cubictcp_register(void) 449{ 450 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); 451 452 /* Precompute a bunch of the scaling factors that are used per-packet 453 * based on SRTT of 100ms 454 */ 455 456 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 457 / (BICTCP_BETA_SCALE - beta); 458 459 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ 460 461 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 462 * so K = cubic_root( (wmax-cwnd)*rtt/c ) 463 * the unit of K is bictcp_HZ=2^10, not HZ 464 * 465 * c = bic_scale >> 10 466 * rtt = 100ms 467 * 468 * the following code has been designed and tested for 469 * cwnd < 1 million packets 470 * RTT < 100 seconds 471 * HZ < 1,000,00 (corresponding to 10 nano-second) 472 */ 473 474 /* 1/c * 2^2*bictcp_HZ * srtt */ 475 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ 476 477 /* divide by bic_scale and by constant Srtt (100ms) */ 478 do_div(cube_factor, bic_scale * 10); 479 480 return tcp_register_congestion_control(&cubictcp); 481} 482 483static void __exit cubictcp_unregister(void) 484{ 485 tcp_unregister_congestion_control(&cubictcp); 486} 487 488module_init(cubictcp_register); 489module_exit(cubictcp_unregister); 490 491MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); 492MODULE_LICENSE("GPL"); 493MODULE_DESCRIPTION("CUBIC TCP"); 494MODULE_VERSION("2.3"); 495