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