1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * TCP CUBIC: Binary Increase Congestion control for TCP v2.3 4 * Home page: 5 * http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC 6 * This is from the implementation of CUBIC TCP in 7 * Sangtae Ha, Injong Rhee and Lisong Xu, 8 * "CUBIC: A New TCP-Friendly High-Speed TCP Variant" 9 * in ACM SIGOPS Operating System Review, July 2008. 10 * Available from: 11 * http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf 12 * 13 * CUBIC integrates a new slow start algorithm, called HyStart. 14 * The details of HyStart are presented in 15 * Sangtae Ha and Injong Rhee, 16 * "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008. 17 * Available from: 18 * http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf 19 * 20 * All testing results are available from: 21 * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing 22 * 23 * Unless CUBIC is enabled and congestion window is large 24 * this behaves the same as the original Reno. 25 */ 26 27 #include <linux/mm.h> 28 #include <linux/btf.h> 29 #include <linux/btf_ids.h> 30 #include <linux/module.h> 31 #include <linux/math64.h> 32 #include <net/tcp.h> 33 34 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation 35 * max_cwnd = snd_cwnd * beta 36 */ 37 #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ 38 39 /* Two methods of hybrid slow start */ 40 #define HYSTART_ACK_TRAIN 0x1 41 #define HYSTART_DELAY 0x2 42 43 /* Number of delay samples for detecting the increase of delay */ 44 #define HYSTART_MIN_SAMPLES 8 45 #define HYSTART_DELAY_MIN (4000U) /* 4 ms */ 46 #define HYSTART_DELAY_MAX (16000U) /* 16 ms */ 47 #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) 48 49 static int fast_convergence __read_mostly = 1; 50 static int beta __read_mostly = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ 51 static int initial_ssthresh __read_mostly; 52 static int bic_scale __read_mostly = 41; 53 static int tcp_friendliness __read_mostly = 1; 54 55 static int hystart __read_mostly = 1; 56 static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY; 57 static int hystart_low_window __read_mostly = 16; 58 static int hystart_ack_delta_us __read_mostly = 2000; 59 60 static u32 cube_rtt_scale __read_mostly; 61 static u32 beta_scale __read_mostly; 62 static u64 cube_factor __read_mostly; 63 64 /* Note parameters that are used for precomputing scale factors are read-only */ 65 module_param(fast_convergence, int, 0644); 66 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence"); 67 module_param(beta, int, 0644); 68 MODULE_PARM_DESC(beta, "beta for multiplicative increase"); 69 module_param(initial_ssthresh, int, 0644); 70 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold"); 71 module_param(bic_scale, int, 0444); 72 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)"); 73 module_param(tcp_friendliness, int, 0644); 74 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness"); 75 module_param(hystart, int, 0644); 76 MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm"); 77 module_param(hystart_detect, int, 0644); 78 MODULE_PARM_DESC(hystart_detect, "hybrid slow start detection mechanisms" 79 " 1: packet-train 2: delay 3: both packet-train and delay"); 80 module_param(hystart_low_window, int, 0644); 81 MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start"); 82 module_param(hystart_ack_delta_us, int, 0644); 83 MODULE_PARM_DESC(hystart_ack_delta_us, "spacing between ack's indicating train (usecs)"); 84 85 /* BIC TCP Parameters */ 86 struct bictcp { 87 u32 cnt; /* increase cwnd by 1 after ACKs */ 88 u32 last_max_cwnd; /* last maximum snd_cwnd */ 89 u32 last_cwnd; /* the last snd_cwnd */ 90 u32 last_time; /* time when updated last_cwnd */ 91 u32 bic_origin_point;/* origin point of bic function */ 92 u32 bic_K; /* time to origin point 93 from the beginning of the current epoch */ 94 u32 delay_min; /* min delay (usec) */ 95 u32 epoch_start; /* beginning of an epoch */ 96 u32 ack_cnt; /* number of acks */ 97 u32 tcp_cwnd; /* estimated tcp cwnd */ 98 u16 unused; 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 }; 106 107 static inline void bictcp_reset(struct bictcp *ca) 108 { 109 memset(ca, 0, offsetof(struct bictcp, unused)); 110 ca->found = 0; 111 } 112 113 static inline u32 bictcp_clock_us(const struct sock *sk) 114 { 115 return tcp_sk(sk)->tcp_mstamp; 116 } 117 118 static inline void bictcp_hystart_reset(struct sock *sk) 119 { 120 struct tcp_sock *tp = tcp_sk(sk); 121 struct bictcp *ca = inet_csk_ca(sk); 122 123 ca->round_start = ca->last_ack = bictcp_clock_us(sk); 124 ca->end_seq = tp->snd_nxt; 125 ca->curr_rtt = ~0U; 126 ca->sample_cnt = 0; 127 } 128 129 __bpf_kfunc static void cubictcp_init(struct sock *sk) 130 { 131 struct bictcp *ca = inet_csk_ca(sk); 132 133 bictcp_reset(ca); 134 135 if (hystart) 136 bictcp_hystart_reset(sk); 137 138 if (!hystart && initial_ssthresh) 139 tcp_sk(sk)->snd_ssthresh = initial_ssthresh; 140 } 141 142 __bpf_kfunc static void cubictcp_cwnd_event_tx_start(struct sock *sk) 143 { 144 struct bictcp *ca = inet_csk_ca(sk); 145 u32 now = tcp_jiffies32; 146 s32 delta; 147 148 delta = now - tcp_sk(sk)->lsndtime; 149 150 /* We were application limited (idle) for a while. 151 * Shift epoch_start to keep cwnd growth to cubic curve. 152 */ 153 if (ca->epoch_start && delta > 0) { 154 ca->epoch_start += delta; 155 if (after(ca->epoch_start, now)) 156 ca->epoch_start = now; 157 } 158 } 159 160 /* calculate the cubic root of x using a table lookup followed by one 161 * Newton-Raphson iteration. 162 * Avg err ~= 0.195% 163 */ 164 static u32 cubic_root(u64 a) 165 { 166 u32 x, b, shift; 167 /* 168 * cbrt(x) MSB values for x MSB values in [0..63]. 169 * Precomputed then refined by hand - Willy Tarreau 170 * 171 * For x in [0..63], 172 * v = cbrt(x << 18) - 1 173 * cbrt(x) = (v[x] + 10) >> 6 174 */ 175 static const u8 v[] = { 176 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, 177 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, 178 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, 179 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, 180 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, 181 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, 182 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, 183 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, 184 }; 185 186 b = fls64(a); 187 if (b < 7) { 188 /* a in [0..63] */ 189 return ((u32)v[(u32)a] + 35) >> 6; 190 } 191 192 b = ((b * 84) >> 8) - 1; 193 shift = (a >> (b * 3)); 194 195 x = ((u32)(((u32)v[shift] + 10) << b)) >> 6; 196 197 /* 198 * Newton-Raphson iteration 199 * 2 200 * x = ( 2 * x + a / x ) / 3 201 * k+1 k k 202 */ 203 x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1))); 204 x = ((x * 341) >> 10); 205 return x; 206 } 207 208 /* 209 * Compute congestion window to use. 210 */ 211 static inline void bictcp_update(struct bictcp *ca, u32 cwnd, u32 acked) 212 { 213 u32 delta, bic_target, max_cnt; 214 u64 offs, t; 215 216 ca->ack_cnt += acked; /* count the number of ACKed packets */ 217 218 if (ca->last_cwnd == cwnd && 219 (s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) 220 return; 221 222 /* The CUBIC function can update ca->cnt at most once per jiffy. 223 * On all cwnd reduction events, ca->epoch_start is set to 0, 224 * which will force a recalculation of ca->cnt. 225 */ 226 if (ca->epoch_start && tcp_jiffies32 == ca->last_time) 227 goto tcp_friendliness; 228 229 ca->last_cwnd = cwnd; 230 ca->last_time = tcp_jiffies32; 231 232 if (ca->epoch_start == 0) { 233 ca->epoch_start = tcp_jiffies32; /* record beginning */ 234 ca->ack_cnt = acked; /* start counting */ 235 ca->tcp_cwnd = cwnd; /* syn with cubic */ 236 237 if (ca->last_max_cwnd <= cwnd) { 238 ca->bic_K = 0; 239 ca->bic_origin_point = cwnd; 240 } else { 241 /* Compute new K based on 242 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) 243 */ 244 ca->bic_K = cubic_root(cube_factor 245 * (ca->last_max_cwnd - cwnd)); 246 ca->bic_origin_point = ca->last_max_cwnd; 247 } 248 } 249 250 /* cubic function - calc*/ 251 /* calculate c * time^3 / rtt, 252 * while considering overflow in calculation of time^3 253 * (so time^3 is done by using 64 bit) 254 * and without the support of division of 64bit numbers 255 * (so all divisions are done by using 32 bit) 256 * also NOTE the unit of those veriables 257 * time = (t - K) / 2^bictcp_HZ 258 * c = bic_scale >> 10 259 * rtt = (srtt >> 3) / HZ 260 * !!! The following code does not have overflow problems, 261 * if the cwnd < 1 million packets !!! 262 */ 263 264 t = (s32)(tcp_jiffies32 - ca->epoch_start); 265 t += usecs_to_jiffies(ca->delay_min); 266 /* change the unit from HZ to bictcp_HZ */ 267 t <<= BICTCP_HZ; 268 do_div(t, HZ); 269 270 if (t < ca->bic_K) /* t - K */ 271 offs = ca->bic_K - t; 272 else 273 offs = t - ca->bic_K; 274 275 /* c/rtt * (t-K)^3 */ 276 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); 277 if (t < ca->bic_K) /* below origin*/ 278 bic_target = ca->bic_origin_point - delta; 279 else /* above origin*/ 280 bic_target = ca->bic_origin_point + delta; 281 282 /* cubic function - calc bictcp_cnt*/ 283 if (bic_target > cwnd) { 284 ca->cnt = cwnd / (bic_target - cwnd); 285 } else { 286 ca->cnt = 100 * cwnd; /* very small increment*/ 287 } 288 289 /* 290 * The initial growth of cubic function may be too conservative 291 * when the available bandwidth is still unknown. 292 */ 293 if (ca->last_max_cwnd == 0 && ca->cnt > 20) 294 ca->cnt = 20; /* increase cwnd 5% per RTT */ 295 296 tcp_friendliness: 297 /* TCP Friendly */ 298 if (tcp_friendliness) { 299 u32 scale = beta_scale; 300 301 delta = (cwnd * scale) >> 3; 302 while (ca->ack_cnt > delta) { /* update tcp cwnd */ 303 ca->ack_cnt -= delta; 304 ca->tcp_cwnd++; 305 } 306 307 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ 308 delta = ca->tcp_cwnd - cwnd; 309 max_cnt = cwnd / delta; 310 if (ca->cnt > max_cnt) 311 ca->cnt = max_cnt; 312 } 313 } 314 315 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per 316 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. 317 */ 318 ca->cnt = max(ca->cnt, 2U); 319 } 320 321 __bpf_kfunc static void cubictcp_cong_avoid(struct sock *sk, u32 ack, u32 acked) 322 { 323 struct tcp_sock *tp = tcp_sk(sk); 324 struct bictcp *ca = inet_csk_ca(sk); 325 326 if (!tcp_is_cwnd_limited(sk)) 327 return; 328 329 if (tcp_in_slow_start(tp)) { 330 acked = tcp_slow_start(tp, acked); 331 if (!acked) 332 return; 333 } 334 bictcp_update(ca, tcp_snd_cwnd(tp), acked); 335 tcp_cong_avoid_ai(tp, ca->cnt, acked); 336 } 337 338 __bpf_kfunc static u32 cubictcp_recalc_ssthresh(struct sock *sk) 339 { 340 const struct tcp_sock *tp = tcp_sk(sk); 341 struct bictcp *ca = inet_csk_ca(sk); 342 343 ca->epoch_start = 0; /* end of epoch */ 344 345 /* Wmax and fast convergence */ 346 if (tcp_snd_cwnd(tp) < ca->last_max_cwnd && fast_convergence) 347 ca->last_max_cwnd = (tcp_snd_cwnd(tp) * (BICTCP_BETA_SCALE + beta)) 348 / (2 * BICTCP_BETA_SCALE); 349 else 350 ca->last_max_cwnd = tcp_snd_cwnd(tp); 351 352 return max((tcp_snd_cwnd(tp) * beta) / BICTCP_BETA_SCALE, 2U); 353 } 354 355 __bpf_kfunc static void cubictcp_state(struct sock *sk, u8 new_state) 356 { 357 if (new_state == TCP_CA_Loss) { 358 bictcp_reset(inet_csk_ca(sk)); 359 bictcp_hystart_reset(sk); 360 } 361 } 362 363 /* Account for TSO/GRO delays. 364 * Otherwise short RTT flows could get too small ssthresh, since during 365 * slow start we begin with small TSO packets and ca->delay_min would 366 * not account for long aggregation delay when TSO packets get bigger. 367 * Ideally even with a very small RTT we would like to have at least one 368 * TSO packet being sent and received by GRO, and another one in qdisc layer. 369 * We apply another 100% factor because @rate is doubled at this point. 370 * We cap the cushion to 1ms. 371 */ 372 static u32 hystart_ack_delay(const struct sock *sk) 373 { 374 unsigned long rate; 375 376 rate = READ_ONCE(sk->sk_pacing_rate); 377 if (!rate) 378 return 0; 379 return min_t(u64, USEC_PER_MSEC, 380 div64_ul((u64)sk->sk_gso_max_size * 4 * USEC_PER_SEC, rate)); 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 u32 threshold; 388 389 if (after(tp->snd_una, ca->end_seq)) 390 bictcp_hystart_reset(sk); 391 392 /* hystart triggers when cwnd is larger than some threshold */ 393 if (tcp_snd_cwnd(tp) < hystart_low_window) 394 return; 395 396 if (hystart_detect & HYSTART_ACK_TRAIN) { 397 u32 now = bictcp_clock_us(sk); 398 399 /* first detection parameter - ack-train detection */ 400 if ((s32)(now - ca->last_ack) <= hystart_ack_delta_us) { 401 ca->last_ack = now; 402 403 threshold = ca->delay_min + hystart_ack_delay(sk); 404 405 /* Hystart ack train triggers if we get ack past 406 * ca->delay_min/2. 407 * Pacing might have delayed packets up to RTT/2 408 * during slow start. 409 */ 410 if (sk->sk_pacing_status == SK_PACING_NONE) 411 threshold >>= 1; 412 413 if ((s32)(now - ca->round_start) > threshold) { 414 ca->found = 1; 415 pr_debug("hystart_ack_train (%u > %u) delay_min %u (+ ack_delay %u) cwnd %u\n", 416 now - ca->round_start, threshold, 417 ca->delay_min, hystart_ack_delay(sk), tcp_snd_cwnd(tp)); 418 NET_INC_STATS(sock_net(sk), 419 LINUX_MIB_TCPHYSTARTTRAINDETECT); 420 NET_ADD_STATS(sock_net(sk), 421 LINUX_MIB_TCPHYSTARTTRAINCWND, 422 tcp_snd_cwnd(tp)); 423 tp->snd_ssthresh = tcp_snd_cwnd(tp); 424 } 425 } 426 } 427 428 if (hystart_detect & HYSTART_DELAY) { 429 /* obtain the minimum delay of more than sampling packets */ 430 if (ca->curr_rtt > delay) 431 ca->curr_rtt = delay; 432 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { 433 ca->sample_cnt++; 434 } else { 435 if (ca->curr_rtt > ca->delay_min + 436 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { 437 ca->found = 1; 438 NET_INC_STATS(sock_net(sk), 439 LINUX_MIB_TCPHYSTARTDELAYDETECT); 440 NET_ADD_STATS(sock_net(sk), 441 LINUX_MIB_TCPHYSTARTDELAYCWND, 442 tcp_snd_cwnd(tp)); 443 tp->snd_ssthresh = tcp_snd_cwnd(tp); 444 } 445 } 446 } 447 } 448 449 __bpf_kfunc static void cubictcp_acked(struct sock *sk, const struct ack_sample *sample) 450 { 451 const struct tcp_sock *tp = tcp_sk(sk); 452 struct bictcp *ca = inet_csk_ca(sk); 453 u32 delay; 454 455 /* Some calls are for duplicates without timetamps */ 456 if (sample->rtt_us < 0) 457 return; 458 459 /* Discard delay samples right after fast recovery */ 460 if (ca->epoch_start && (s32)(tcp_jiffies32 - ca->epoch_start) < HZ) 461 return; 462 463 delay = sample->rtt_us; 464 if (delay == 0) 465 delay = 1; 466 467 /* first time call or link delay decreases */ 468 if (ca->delay_min == 0 || ca->delay_min > delay) 469 ca->delay_min = delay; 470 471 if (!ca->found && tcp_in_slow_start(tp) && hystart) 472 hystart_update(sk, delay); 473 } 474 475 static struct tcp_congestion_ops cubictcp __read_mostly = { 476 .init = cubictcp_init, 477 .ssthresh = cubictcp_recalc_ssthresh, 478 .cong_avoid = cubictcp_cong_avoid, 479 .set_state = cubictcp_state, 480 .undo_cwnd = tcp_reno_undo_cwnd, 481 .cwnd_event_tx_start = cubictcp_cwnd_event_tx_start, 482 .pkts_acked = cubictcp_acked, 483 .owner = THIS_MODULE, 484 .name = "cubic", 485 }; 486 487 BTF_KFUNCS_START(tcp_cubic_check_kfunc_ids) 488 BTF_ID_FLAGS(func, cubictcp_init) 489 BTF_ID_FLAGS(func, cubictcp_recalc_ssthresh) 490 BTF_ID_FLAGS(func, cubictcp_cong_avoid) 491 BTF_ID_FLAGS(func, cubictcp_state) 492 BTF_ID_FLAGS(func, cubictcp_cwnd_event_tx_start) 493 BTF_ID_FLAGS(func, cubictcp_acked) 494 BTF_KFUNCS_END(tcp_cubic_check_kfunc_ids) 495 496 static const struct btf_kfunc_id_set tcp_cubic_kfunc_set = { 497 .owner = THIS_MODULE, 498 .set = &tcp_cubic_check_kfunc_ids, 499 }; 500 501 static int __init cubictcp_register(void) 502 { 503 int ret; 504 505 BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE); 506 507 /* Precompute a bunch of the scaling factors that are used per-packet 508 * based on SRTT of 100ms 509 */ 510 511 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 512 / (BICTCP_BETA_SCALE - beta); 513 514 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ 515 516 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 517 * so K = cubic_root( (wmax-cwnd)*rtt/c ) 518 * the unit of K is bictcp_HZ=2^10, not HZ 519 * 520 * c = bic_scale >> 10 521 * rtt = 100ms 522 * 523 * the following code has been designed and tested for 524 * cwnd < 1 million packets 525 * RTT < 100 seconds 526 * HZ < 1,000,00 (corresponding to 10 nano-second) 527 */ 528 529 /* 1/c * 2^2*bictcp_HZ * srtt */ 530 cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */ 531 532 /* divide by bic_scale and by constant Srtt (100ms) */ 533 do_div(cube_factor, bic_scale * 10); 534 535 ret = register_btf_kfunc_id_set(BPF_PROG_TYPE_STRUCT_OPS, &tcp_cubic_kfunc_set); 536 if (ret < 0) 537 return ret; 538 return tcp_register_congestion_control(&cubictcp); 539 } 540 541 static void __exit cubictcp_unregister(void) 542 { 543 tcp_unregister_congestion_control(&cubictcp); 544 } 545 546 module_init(cubictcp_register); 547 module_exit(cubictcp_unregister); 548 549 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger"); 550 MODULE_LICENSE("GPL"); 551 MODULE_DESCRIPTION("CUBIC TCP"); 552 MODULE_VERSION("2.3"); 553