1 // SPDX-License-Identifier: GPL-2.0-only 2 3 /* WARNING: This implementation is not necessarily the same 4 * as the tcp_cubic.c. The purpose is mainly for testing 5 * the kernel BPF logic. 6 * 7 * Highlights: 8 * 1. CONFIG_HZ .kconfig map is used. 9 * 2. In bictcp_update(), calculation is changed to use usec 10 * resolution (i.e. USEC_PER_JIFFY) instead of using jiffies. 11 * Thus, usecs_to_jiffies() is not used in the bpf_cubic.c. 12 * 3. In bitctcp_update() [under tcp_friendliness], the original 13 * "while (ca->ack_cnt > delta)" loop is changed to the equivalent 14 * "ca->ack_cnt / delta" operation. 15 */ 16 17 #include "bpf_tracing_net.h" 18 #include <bpf/bpf_tracing.h> 19 #include <errno.h> 20 21 char _license[] SEC("license") = "GPL"; 22 23 #define clamp(val, lo, hi) min((typeof(val))max(val, lo), hi) 24 25 extern __u32 tcp_slow_start(struct tcp_sock *tp, __u32 acked) __ksym; 26 extern void tcp_cong_avoid_ai(struct tcp_sock *tp, __u32 w, __u32 acked) __ksym; 27 28 #define BICTCP_BETA_SCALE 1024 /* Scale factor beta calculation 29 * max_cwnd = snd_cwnd * beta 30 */ 31 #define BICTCP_HZ 10 /* BIC HZ 2^10 = 1024 */ 32 33 /* Two methods of hybrid slow start */ 34 #define HYSTART_ACK_TRAIN 0x1 35 #define HYSTART_DELAY 0x2 36 37 /* Number of delay samples for detecting the increase of delay */ 38 #define HYSTART_MIN_SAMPLES 8 39 #define HYSTART_DELAY_MIN (4000U) /* 4ms */ 40 #define HYSTART_DELAY_MAX (16000U) /* 16 ms */ 41 #define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX) 42 43 static int fast_convergence = 1; 44 static const int beta = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ 45 static int initial_ssthresh; 46 static const int bic_scale = 41; 47 static int tcp_friendliness = 1; 48 49 static int hystart = 1; 50 static int hystart_detect = HYSTART_ACK_TRAIN | HYSTART_DELAY; 51 static int hystart_low_window = 16; 52 static int hystart_ack_delta_us = 2000; 53 54 static const __u32 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ 55 static const __u32 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 56 / (BICTCP_BETA_SCALE - beta); 57 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 58 * so K = cubic_root( (wmax-cwnd)*rtt/c ) 59 * the unit of K is bictcp_HZ=2^10, not HZ 60 * 61 * c = bic_scale >> 10 62 * rtt = 100ms 63 * 64 * the following code has been designed and tested for 65 * cwnd < 1 million packets 66 * RTT < 100 seconds 67 * HZ < 1,000,00 (corresponding to 10 nano-second) 68 */ 69 70 /* 1/c * 2^2*bictcp_HZ * srtt, 2^40 */ 71 static const __u64 cube_factor = (__u64)(1ull << (10+3*BICTCP_HZ)) 72 / (bic_scale * 10); 73 74 /* BIC TCP Parameters */ 75 struct bpf_bictcp { 76 __u32 cnt; /* increase cwnd by 1 after ACKs */ 77 __u32 last_max_cwnd; /* last maximum snd_cwnd */ 78 __u32 last_cwnd; /* the last snd_cwnd */ 79 __u32 last_time; /* time when updated last_cwnd */ 80 __u32 bic_origin_point;/* origin point of bic function */ 81 __u32 bic_K; /* time to origin point 82 from the beginning of the current epoch */ 83 __u32 delay_min; /* min delay (usec) */ 84 __u32 epoch_start; /* beginning of an epoch */ 85 __u32 ack_cnt; /* number of acks */ 86 __u32 tcp_cwnd; /* estimated tcp cwnd */ 87 __u16 unused; 88 __u8 sample_cnt; /* number of samples to decide curr_rtt */ 89 __u8 found; /* the exit point is found? */ 90 __u32 round_start; /* beginning of each round */ 91 __u32 end_seq; /* end_seq of the round */ 92 __u32 last_ack; /* last time when the ACK spacing is close */ 93 __u32 curr_rtt; /* the minimum rtt of current round */ 94 }; 95 96 static void bictcp_reset(struct bpf_bictcp *ca) 97 { 98 ca->cnt = 0; 99 ca->last_max_cwnd = 0; 100 ca->last_cwnd = 0; 101 ca->last_time = 0; 102 ca->bic_origin_point = 0; 103 ca->bic_K = 0; 104 ca->delay_min = 0; 105 ca->epoch_start = 0; 106 ca->ack_cnt = 0; 107 ca->tcp_cwnd = 0; 108 ca->found = 0; 109 } 110 111 extern unsigned long CONFIG_HZ __kconfig; 112 #define HZ CONFIG_HZ 113 #define USEC_PER_MSEC 1000UL 114 #define USEC_PER_SEC 1000000UL 115 #define USEC_PER_JIFFY (USEC_PER_SEC / HZ) 116 117 static __u64 div64_u64(__u64 dividend, __u64 divisor) 118 { 119 return dividend / divisor; 120 } 121 122 #define div64_ul div64_u64 123 124 #define BITS_PER_U64 (sizeof(__u64) * 8) 125 static int fls64(__u64 x) 126 { 127 int num = BITS_PER_U64 - 1; 128 129 if (x == 0) 130 return 0; 131 132 if (!(x & (~0ull << (BITS_PER_U64-32)))) { 133 num -= 32; 134 x <<= 32; 135 } 136 if (!(x & (~0ull << (BITS_PER_U64-16)))) { 137 num -= 16; 138 x <<= 16; 139 } 140 if (!(x & (~0ull << (BITS_PER_U64-8)))) { 141 num -= 8; 142 x <<= 8; 143 } 144 if (!(x & (~0ull << (BITS_PER_U64-4)))) { 145 num -= 4; 146 x <<= 4; 147 } 148 if (!(x & (~0ull << (BITS_PER_U64-2)))) { 149 num -= 2; 150 x <<= 2; 151 } 152 if (!(x & (~0ull << (BITS_PER_U64-1)))) 153 num -= 1; 154 155 return num + 1; 156 } 157 158 static __u32 bictcp_clock_us(const struct sock *sk) 159 { 160 return tcp_sk(sk)->tcp_mstamp; 161 } 162 163 static void bictcp_hystart_reset(struct sock *sk) 164 { 165 struct tcp_sock *tp = tcp_sk(sk); 166 struct bpf_bictcp *ca = inet_csk_ca(sk); 167 168 ca->round_start = ca->last_ack = bictcp_clock_us(sk); 169 ca->end_seq = tp->snd_nxt; 170 ca->curr_rtt = ~0U; 171 ca->sample_cnt = 0; 172 } 173 174 bool nodelay_init_reject = false; 175 bool nodelay_cwnd_event_tx_start_reject = false; 176 177 SEC("struct_ops") 178 void BPF_PROG(bpf_cubic_init, struct sock *sk) 179 { 180 struct bpf_bictcp *ca = inet_csk_ca(sk); 181 int true_val = 1, ret; 182 183 ret = bpf_setsockopt(sk, SOL_TCP, TCP_NODELAY, &true_val, sizeof(true_val)); 184 if (ret == -EOPNOTSUPP) 185 nodelay_init_reject = true; 186 187 bictcp_reset(ca); 188 189 if (hystart) 190 bictcp_hystart_reset(sk); 191 192 if (!hystart && initial_ssthresh) 193 tcp_sk(sk)->snd_ssthresh = initial_ssthresh; 194 } 195 196 SEC("struct_ops") 197 void BPF_PROG(bpf_cubic_cwnd_event_tx_start, struct sock *sk) 198 { 199 struct bpf_bictcp *ca = inet_csk_ca(sk); 200 __u32 now = tcp_jiffies32; 201 int true_val = 1, ret; 202 __s32 delta; 203 204 ret = bpf_setsockopt(sk, SOL_TCP, TCP_NODELAY, &true_val, sizeof(true_val)); 205 if (ret == -EOPNOTSUPP) 206 nodelay_cwnd_event_tx_start_reject = true; 207 208 delta = now - tcp_sk(sk)->lsndtime; 209 210 /* We were application limited (idle) for a while. 211 * Shift epoch_start to keep cwnd growth to cubic curve. 212 */ 213 if (ca->epoch_start && delta > 0) { 214 ca->epoch_start += delta; 215 if (after(ca->epoch_start, now)) 216 ca->epoch_start = now; 217 } 218 } 219 220 /* 221 * cbrt(x) MSB values for x MSB values in [0..63]. 222 * Precomputed then refined by hand - Willy Tarreau 223 * 224 * For x in [0..63], 225 * v = cbrt(x << 18) - 1 226 * cbrt(x) = (v[x] + 10) >> 6 227 */ 228 static const __u8 v[] = { 229 /* 0x00 */ 0, 54, 54, 54, 118, 118, 118, 118, 230 /* 0x08 */ 123, 129, 134, 138, 143, 147, 151, 156, 231 /* 0x10 */ 157, 161, 164, 168, 170, 173, 176, 179, 232 /* 0x18 */ 181, 185, 187, 190, 192, 194, 197, 199, 233 /* 0x20 */ 200, 202, 204, 206, 209, 211, 213, 215, 234 /* 0x28 */ 217, 219, 221, 222, 224, 225, 227, 229, 235 /* 0x30 */ 231, 232, 234, 236, 237, 239, 240, 242, 236 /* 0x38 */ 244, 245, 246, 248, 250, 251, 252, 254, 237 }; 238 239 /* calculate the cubic root of x using a table lookup followed by one 240 * Newton-Raphson iteration. 241 * Avg err ~= 0.195% 242 */ 243 static __u32 cubic_root(__u64 a) 244 { 245 __u32 x, b, shift; 246 247 if (a < 64) { 248 /* a in [0..63] */ 249 return ((__u32)v[(__u32)a] + 35) >> 6; 250 } 251 252 b = fls64(a); 253 b = ((b * 84) >> 8) - 1; 254 shift = (a >> (b * 3)); 255 256 /* it is needed for verifier's bound check on v */ 257 if (shift >= 64) 258 return 0; 259 260 x = ((__u32)(((__u32)v[shift] + 10) << b)) >> 6; 261 262 /* 263 * Newton-Raphson iteration 264 * 2 265 * x = ( 2 * x + a / x ) / 3 266 * k+1 k k 267 */ 268 x = (2 * x + (__u32)div64_u64(a, (__u64)x * (__u64)(x - 1))); 269 x = ((x * 341) >> 10); 270 return x; 271 } 272 273 /* 274 * Compute congestion window to use. 275 */ 276 static void bictcp_update(struct bpf_bictcp *ca, __u32 cwnd, __u32 acked) 277 { 278 __u32 delta, bic_target, max_cnt; 279 __u64 offs, t; 280 281 ca->ack_cnt += acked; /* count the number of ACKed packets */ 282 283 if (ca->last_cwnd == cwnd && 284 (__s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) 285 return; 286 287 /* The CUBIC function can update ca->cnt at most once per jiffy. 288 * On all cwnd reduction events, ca->epoch_start is set to 0, 289 * which will force a recalculation of ca->cnt. 290 */ 291 if (ca->epoch_start && tcp_jiffies32 == ca->last_time) 292 goto tcp_friendliness; 293 294 ca->last_cwnd = cwnd; 295 ca->last_time = tcp_jiffies32; 296 297 if (ca->epoch_start == 0) { 298 ca->epoch_start = tcp_jiffies32; /* record beginning */ 299 ca->ack_cnt = acked; /* start counting */ 300 ca->tcp_cwnd = cwnd; /* syn with cubic */ 301 302 if (ca->last_max_cwnd <= cwnd) { 303 ca->bic_K = 0; 304 ca->bic_origin_point = cwnd; 305 } else { 306 /* Compute new K based on 307 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) 308 */ 309 ca->bic_K = cubic_root(cube_factor 310 * (ca->last_max_cwnd - cwnd)); 311 ca->bic_origin_point = ca->last_max_cwnd; 312 } 313 } 314 315 /* cubic function - calc*/ 316 /* calculate c * time^3 / rtt, 317 * while considering overflow in calculation of time^3 318 * (so time^3 is done by using 64 bit) 319 * and without the support of division of 64bit numbers 320 * (so all divisions are done by using 32 bit) 321 * also NOTE the unit of those variables 322 * time = (t - K) / 2^bictcp_HZ 323 * c = bic_scale >> 10 324 * rtt = (srtt >> 3) / HZ 325 * !!! The following code does not have overflow problems, 326 * if the cwnd < 1 million packets !!! 327 */ 328 329 t = (__s32)(tcp_jiffies32 - ca->epoch_start) * USEC_PER_JIFFY; 330 t += ca->delay_min; 331 /* change the unit from usec to bictcp_HZ */ 332 t <<= BICTCP_HZ; 333 t /= USEC_PER_SEC; 334 335 if (t < ca->bic_K) /* t - K */ 336 offs = ca->bic_K - t; 337 else 338 offs = t - ca->bic_K; 339 340 /* c/rtt * (t-K)^3 */ 341 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); 342 if (t < ca->bic_K) /* below origin*/ 343 bic_target = ca->bic_origin_point - delta; 344 else /* above origin*/ 345 bic_target = ca->bic_origin_point + delta; 346 347 /* cubic function - calc bictcp_cnt*/ 348 if (bic_target > cwnd) { 349 ca->cnt = cwnd / (bic_target - cwnd); 350 } else { 351 ca->cnt = 100 * cwnd; /* very small increment*/ 352 } 353 354 /* 355 * The initial growth of cubic function may be too conservative 356 * when the available bandwidth is still unknown. 357 */ 358 if (ca->last_max_cwnd == 0 && ca->cnt > 20) 359 ca->cnt = 20; /* increase cwnd 5% per RTT */ 360 361 tcp_friendliness: 362 /* TCP Friendly */ 363 if (tcp_friendliness) { 364 __u32 scale = beta_scale; 365 __u32 n; 366 367 /* update tcp cwnd */ 368 delta = (cwnd * scale) >> 3; 369 if (ca->ack_cnt > delta && delta) { 370 n = ca->ack_cnt / delta; 371 ca->ack_cnt -= n * delta; 372 ca->tcp_cwnd += n; 373 } 374 375 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ 376 delta = ca->tcp_cwnd - cwnd; 377 max_cnt = cwnd / delta; 378 if (ca->cnt > max_cnt) 379 ca->cnt = max_cnt; 380 } 381 } 382 383 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per 384 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. 385 */ 386 ca->cnt = max(ca->cnt, 2U); 387 } 388 389 SEC("struct_ops") 390 void BPF_PROG(bpf_cubic_cong_avoid, struct sock *sk, __u32 ack, __u32 acked) 391 { 392 struct tcp_sock *tp = tcp_sk(sk); 393 struct bpf_bictcp *ca = inet_csk_ca(sk); 394 395 if (!tcp_is_cwnd_limited(sk)) 396 return; 397 398 if (tcp_in_slow_start(tp)) { 399 if (hystart && after(ack, ca->end_seq)) 400 bictcp_hystart_reset(sk); 401 acked = tcp_slow_start(tp, acked); 402 if (!acked) 403 return; 404 } 405 bictcp_update(ca, tp->snd_cwnd, acked); 406 tcp_cong_avoid_ai(tp, ca->cnt, acked); 407 } 408 409 SEC("struct_ops") 410 __u32 BPF_PROG(bpf_cubic_recalc_ssthresh, struct sock *sk) 411 { 412 const struct tcp_sock *tp = tcp_sk(sk); 413 struct bpf_bictcp *ca = inet_csk_ca(sk); 414 415 ca->epoch_start = 0; /* end of epoch */ 416 417 /* Wmax and fast convergence */ 418 if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence) 419 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta)) 420 / (2 * BICTCP_BETA_SCALE); 421 else 422 ca->last_max_cwnd = tp->snd_cwnd; 423 424 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U); 425 } 426 427 SEC("struct_ops") 428 void BPF_PROG(bpf_cubic_state, struct sock *sk, __u8 new_state) 429 { 430 if (new_state == TCP_CA_Loss) { 431 bictcp_reset(inet_csk_ca(sk)); 432 bictcp_hystart_reset(sk); 433 } 434 } 435 436 #define GSO_MAX_SIZE 65536 437 438 /* Account for TSO/GRO delays. 439 * Otherwise short RTT flows could get too small ssthresh, since during 440 * slow start we begin with small TSO packets and ca->delay_min would 441 * not account for long aggregation delay when TSO packets get bigger. 442 * Ideally even with a very small RTT we would like to have at least one 443 * TSO packet being sent and received by GRO, and another one in qdisc layer. 444 * We apply another 100% factor because @rate is doubled at this point. 445 * We cap the cushion to 1ms. 446 */ 447 static __u32 hystart_ack_delay(struct sock *sk) 448 { 449 unsigned long rate; 450 451 rate = sk->sk_pacing_rate; 452 if (!rate) 453 return 0; 454 return min((__u64)USEC_PER_MSEC, 455 div64_ul((__u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate)); 456 } 457 458 static void hystart_update(struct sock *sk, __u32 delay) 459 { 460 struct tcp_sock *tp = tcp_sk(sk); 461 struct bpf_bictcp *ca = inet_csk_ca(sk); 462 __u32 threshold; 463 464 if (hystart_detect & HYSTART_ACK_TRAIN) { 465 __u32 now = bictcp_clock_us(sk); 466 467 /* first detection parameter - ack-train detection */ 468 if ((__s32)(now - ca->last_ack) <= hystart_ack_delta_us) { 469 ca->last_ack = now; 470 471 threshold = ca->delay_min + hystart_ack_delay(sk); 472 473 /* Hystart ack train triggers if we get ack past 474 * ca->delay_min/2. 475 * Pacing might have delayed packets up to RTT/2 476 * during slow start. 477 */ 478 if (sk->sk_pacing_status == SK_PACING_NONE) 479 threshold >>= 1; 480 481 if ((__s32)(now - ca->round_start) > threshold) { 482 ca->found = 1; 483 tp->snd_ssthresh = tp->snd_cwnd; 484 } 485 } 486 } 487 488 if (hystart_detect & HYSTART_DELAY) { 489 /* obtain the minimum delay of more than sampling packets */ 490 if (ca->curr_rtt > delay) 491 ca->curr_rtt = delay; 492 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { 493 ca->sample_cnt++; 494 } else { 495 if (ca->curr_rtt > ca->delay_min + 496 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { 497 ca->found = 1; 498 tp->snd_ssthresh = tp->snd_cwnd; 499 } 500 } 501 } 502 } 503 504 int bpf_cubic_acked_called = 0; 505 506 SEC("struct_ops") 507 void BPF_PROG(bpf_cubic_acked, struct sock *sk, const struct ack_sample *sample) 508 { 509 const struct tcp_sock *tp = tcp_sk(sk); 510 struct bpf_bictcp *ca = inet_csk_ca(sk); 511 __u32 delay; 512 513 bpf_cubic_acked_called = 1; 514 /* Some calls are for duplicates without timestamps */ 515 if (sample->rtt_us < 0) 516 return; 517 518 /* Discard delay samples right after fast recovery */ 519 if (ca->epoch_start && (__s32)(tcp_jiffies32 - ca->epoch_start) < HZ) 520 return; 521 522 delay = sample->rtt_us; 523 if (delay == 0) 524 delay = 1; 525 526 /* first time call or link delay decreases */ 527 if (ca->delay_min == 0 || ca->delay_min > delay) 528 ca->delay_min = delay; 529 530 /* hystart triggers when cwnd is larger than some threshold */ 531 if (!ca->found && tcp_in_slow_start(tp) && hystart && 532 tp->snd_cwnd >= hystart_low_window) 533 hystart_update(sk, delay); 534 } 535 536 extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym; 537 538 SEC("struct_ops") 539 __u32 BPF_PROG(bpf_cubic_undo_cwnd, struct sock *sk) 540 { 541 return tcp_reno_undo_cwnd(sk); 542 } 543 544 SEC(".struct_ops") 545 struct tcp_congestion_ops cubic = { 546 .init = (void *)bpf_cubic_init, 547 .ssthresh = (void *)bpf_cubic_recalc_ssthresh, 548 .cong_avoid = (void *)bpf_cubic_cong_avoid, 549 .set_state = (void *)bpf_cubic_state, 550 .undo_cwnd = (void *)bpf_cubic_undo_cwnd, 551 .cwnd_event_tx_start = (void *)bpf_cubic_cwnd_event_tx_start, 552 .pkts_acked = (void *)bpf_cubic_acked, 553 .name = "bpf_cubic", 554 }; 555