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