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