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