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 20 char _license[] SEC("license") = "GPL"; 21 22 #define clamp(val, lo, hi) min((typeof(val))max(val, lo), hi) 23 #define min(a, b) ((a) < (b) ? (a) : (b)) 24 #define max(a, b) ((a) > (b) ? (a) : (b)) 25 static bool before(__u32 seq1, __u32 seq2) 26 { 27 return (__s32)(seq1-seq2) < 0; 28 } 29 #define after(seq2, seq1) before(seq1, seq2) 30 31 extern __u32 tcp_slow_start(struct tcp_sock *tp, __u32 acked) __ksym; 32 extern void tcp_cong_avoid_ai(struct tcp_sock *tp, __u32 w, __u32 acked) __ksym; 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) /* 4ms */ 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 = 1; 50 static const int beta = 717; /* = 717/1024 (BICTCP_BETA_SCALE) */ 51 static int initial_ssthresh; 52 static const int bic_scale = 41; 53 static int tcp_friendliness = 1; 54 55 static int hystart = 1; 56 static int hystart_detect = HYSTART_ACK_TRAIN | HYSTART_DELAY; 57 static int hystart_low_window = 16; 58 static int hystart_ack_delta_us = 2000; 59 60 static const __u32 cube_rtt_scale = (bic_scale * 10); /* 1024*c/rtt */ 61 static const __u32 beta_scale = 8*(BICTCP_BETA_SCALE+beta) / 3 62 / (BICTCP_BETA_SCALE - beta); 63 /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3 64 * so K = cubic_root( (wmax-cwnd)*rtt/c ) 65 * the unit of K is bictcp_HZ=2^10, not HZ 66 * 67 * c = bic_scale >> 10 68 * rtt = 100ms 69 * 70 * the following code has been designed and tested for 71 * cwnd < 1 million packets 72 * RTT < 100 seconds 73 * HZ < 1,000,00 (corresponding to 10 nano-second) 74 */ 75 76 /* 1/c * 2^2*bictcp_HZ * srtt, 2^40 */ 77 static const __u64 cube_factor = (__u64)(1ull << (10+3*BICTCP_HZ)) 78 / (bic_scale * 10); 79 80 /* BIC TCP Parameters */ 81 struct bpf_bictcp { 82 __u32 cnt; /* increase cwnd by 1 after ACKs */ 83 __u32 last_max_cwnd; /* last maximum snd_cwnd */ 84 __u32 last_cwnd; /* the last snd_cwnd */ 85 __u32 last_time; /* time when updated last_cwnd */ 86 __u32 bic_origin_point;/* origin point of bic function */ 87 __u32 bic_K; /* time to origin point 88 from the beginning of the current epoch */ 89 __u32 delay_min; /* min delay (usec) */ 90 __u32 epoch_start; /* beginning of an epoch */ 91 __u32 ack_cnt; /* number of acks */ 92 __u32 tcp_cwnd; /* estimated tcp cwnd */ 93 __u16 unused; 94 __u8 sample_cnt; /* number of samples to decide curr_rtt */ 95 __u8 found; /* the exit point is found? */ 96 __u32 round_start; /* beginning of each round */ 97 __u32 end_seq; /* end_seq of the round */ 98 __u32 last_ack; /* last time when the ACK spacing is close */ 99 __u32 curr_rtt; /* the minimum rtt of current round */ 100 }; 101 102 static void bictcp_reset(struct bpf_bictcp *ca) 103 { 104 ca->cnt = 0; 105 ca->last_max_cwnd = 0; 106 ca->last_cwnd = 0; 107 ca->last_time = 0; 108 ca->bic_origin_point = 0; 109 ca->bic_K = 0; 110 ca->delay_min = 0; 111 ca->epoch_start = 0; 112 ca->ack_cnt = 0; 113 ca->tcp_cwnd = 0; 114 ca->found = 0; 115 } 116 117 extern unsigned long CONFIG_HZ __kconfig; 118 #define HZ CONFIG_HZ 119 #define USEC_PER_MSEC 1000UL 120 #define USEC_PER_SEC 1000000UL 121 #define USEC_PER_JIFFY (USEC_PER_SEC / HZ) 122 123 static __u64 div64_u64(__u64 dividend, __u64 divisor) 124 { 125 return dividend / divisor; 126 } 127 128 #define div64_ul div64_u64 129 130 #define BITS_PER_U64 (sizeof(__u64) * 8) 131 static int fls64(__u64 x) 132 { 133 int num = BITS_PER_U64 - 1; 134 135 if (x == 0) 136 return 0; 137 138 if (!(x & (~0ull << (BITS_PER_U64-32)))) { 139 num -= 32; 140 x <<= 32; 141 } 142 if (!(x & (~0ull << (BITS_PER_U64-16)))) { 143 num -= 16; 144 x <<= 16; 145 } 146 if (!(x & (~0ull << (BITS_PER_U64-8)))) { 147 num -= 8; 148 x <<= 8; 149 } 150 if (!(x & (~0ull << (BITS_PER_U64-4)))) { 151 num -= 4; 152 x <<= 4; 153 } 154 if (!(x & (~0ull << (BITS_PER_U64-2)))) { 155 num -= 2; 156 x <<= 2; 157 } 158 if (!(x & (~0ull << (BITS_PER_U64-1)))) 159 num -= 1; 160 161 return num + 1; 162 } 163 164 static __u32 bictcp_clock_us(const struct sock *sk) 165 { 166 return tcp_sk(sk)->tcp_mstamp; 167 } 168 169 static void bictcp_hystart_reset(struct sock *sk) 170 { 171 struct tcp_sock *tp = tcp_sk(sk); 172 struct bpf_bictcp *ca = inet_csk_ca(sk); 173 174 ca->round_start = ca->last_ack = bictcp_clock_us(sk); 175 ca->end_seq = tp->snd_nxt; 176 ca->curr_rtt = ~0U; 177 ca->sample_cnt = 0; 178 } 179 180 SEC("struct_ops") 181 void BPF_PROG(bpf_cubic_init, struct sock *sk) 182 { 183 struct bpf_bictcp *ca = inet_csk_ca(sk); 184 185 bictcp_reset(ca); 186 187 if (hystart) 188 bictcp_hystart_reset(sk); 189 190 if (!hystart && initial_ssthresh) 191 tcp_sk(sk)->snd_ssthresh = initial_ssthresh; 192 } 193 194 SEC("struct_ops") 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 bpf_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 __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 void bictcp_update(struct bpf_bictcp *ca, __u32 cwnd, __u32 acked) 273 { 274 __u32 delta, bic_target, max_cnt; 275 __u64 offs, t; 276 277 ca->ack_cnt += acked; /* count the number of ACKed packets */ 278 279 if (ca->last_cwnd == cwnd && 280 (__s32)(tcp_jiffies32 - ca->last_time) <= HZ / 32) 281 return; 282 283 /* The CUBIC function can update ca->cnt at most once per jiffy. 284 * On all cwnd reduction events, ca->epoch_start is set to 0, 285 * which will force a recalculation of ca->cnt. 286 */ 287 if (ca->epoch_start && tcp_jiffies32 == ca->last_time) 288 goto tcp_friendliness; 289 290 ca->last_cwnd = cwnd; 291 ca->last_time = tcp_jiffies32; 292 293 if (ca->epoch_start == 0) { 294 ca->epoch_start = tcp_jiffies32; /* record beginning */ 295 ca->ack_cnt = acked; /* start counting */ 296 ca->tcp_cwnd = cwnd; /* syn with cubic */ 297 298 if (ca->last_max_cwnd <= cwnd) { 299 ca->bic_K = 0; 300 ca->bic_origin_point = cwnd; 301 } else { 302 /* Compute new K based on 303 * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ) 304 */ 305 ca->bic_K = cubic_root(cube_factor 306 * (ca->last_max_cwnd - cwnd)); 307 ca->bic_origin_point = ca->last_max_cwnd; 308 } 309 } 310 311 /* cubic function - calc*/ 312 /* calculate c * time^3 / rtt, 313 * while considering overflow in calculation of time^3 314 * (so time^3 is done by using 64 bit) 315 * and without the support of division of 64bit numbers 316 * (so all divisions are done by using 32 bit) 317 * also NOTE the unit of those variables 318 * time = (t - K) / 2^bictcp_HZ 319 * c = bic_scale >> 10 320 * rtt = (srtt >> 3) / HZ 321 * !!! The following code does not have overflow problems, 322 * if the cwnd < 1 million packets !!! 323 */ 324 325 t = (__s32)(tcp_jiffies32 - ca->epoch_start) * USEC_PER_JIFFY; 326 t += ca->delay_min; 327 /* change the unit from usec to bictcp_HZ */ 328 t <<= BICTCP_HZ; 329 t /= USEC_PER_SEC; 330 331 if (t < ca->bic_K) /* t - K */ 332 offs = ca->bic_K - t; 333 else 334 offs = t - ca->bic_K; 335 336 /* c/rtt * (t-K)^3 */ 337 delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ); 338 if (t < ca->bic_K) /* below origin*/ 339 bic_target = ca->bic_origin_point - delta; 340 else /* above origin*/ 341 bic_target = ca->bic_origin_point + delta; 342 343 /* cubic function - calc bictcp_cnt*/ 344 if (bic_target > cwnd) { 345 ca->cnt = cwnd / (bic_target - cwnd); 346 } else { 347 ca->cnt = 100 * cwnd; /* very small increment*/ 348 } 349 350 /* 351 * The initial growth of cubic function may be too conservative 352 * when the available bandwidth is still unknown. 353 */ 354 if (ca->last_max_cwnd == 0 && ca->cnt > 20) 355 ca->cnt = 20; /* increase cwnd 5% per RTT */ 356 357 tcp_friendliness: 358 /* TCP Friendly */ 359 if (tcp_friendliness) { 360 __u32 scale = beta_scale; 361 __u32 n; 362 363 /* update tcp cwnd */ 364 delta = (cwnd * scale) >> 3; 365 if (ca->ack_cnt > delta && delta) { 366 n = ca->ack_cnt / delta; 367 ca->ack_cnt -= n * delta; 368 ca->tcp_cwnd += n; 369 } 370 371 if (ca->tcp_cwnd > cwnd) { /* if bic is slower than tcp */ 372 delta = ca->tcp_cwnd - cwnd; 373 max_cnt = cwnd / delta; 374 if (ca->cnt > max_cnt) 375 ca->cnt = max_cnt; 376 } 377 } 378 379 /* The maximum rate of cwnd increase CUBIC allows is 1 packet per 380 * 2 packets ACKed, meaning cwnd grows at 1.5x per RTT. 381 */ 382 ca->cnt = max(ca->cnt, 2U); 383 } 384 385 SEC("struct_ops") 386 void BPF_PROG(bpf_cubic_cong_avoid, struct sock *sk, __u32 ack, __u32 acked) 387 { 388 struct tcp_sock *tp = tcp_sk(sk); 389 struct bpf_bictcp *ca = inet_csk_ca(sk); 390 391 if (!tcp_is_cwnd_limited(sk)) 392 return; 393 394 if (tcp_in_slow_start(tp)) { 395 if (hystart && after(ack, ca->end_seq)) 396 bictcp_hystart_reset(sk); 397 acked = tcp_slow_start(tp, acked); 398 if (!acked) 399 return; 400 } 401 bictcp_update(ca, tp->snd_cwnd, acked); 402 tcp_cong_avoid_ai(tp, ca->cnt, acked); 403 } 404 405 SEC("struct_ops") 406 __u32 BPF_PROG(bpf_cubic_recalc_ssthresh, struct sock *sk) 407 { 408 const struct tcp_sock *tp = tcp_sk(sk); 409 struct bpf_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 SEC("struct_ops") 424 void BPF_PROG(bpf_cubic_state, struct sock *sk, __u8 new_state) 425 { 426 if (new_state == TCP_CA_Loss) { 427 bictcp_reset(inet_csk_ca(sk)); 428 bictcp_hystart_reset(sk); 429 } 430 } 431 432 #define GSO_MAX_SIZE 65536 433 434 /* Account for TSO/GRO delays. 435 * Otherwise short RTT flows could get too small ssthresh, since during 436 * slow start we begin with small TSO packets and ca->delay_min would 437 * not account for long aggregation delay when TSO packets get bigger. 438 * Ideally even with a very small RTT we would like to have at least one 439 * TSO packet being sent and received by GRO, and another one in qdisc layer. 440 * We apply another 100% factor because @rate is doubled at this point. 441 * We cap the cushion to 1ms. 442 */ 443 static __u32 hystart_ack_delay(struct sock *sk) 444 { 445 unsigned long rate; 446 447 rate = sk->sk_pacing_rate; 448 if (!rate) 449 return 0; 450 return min((__u64)USEC_PER_MSEC, 451 div64_ul((__u64)GSO_MAX_SIZE * 4 * USEC_PER_SEC, rate)); 452 } 453 454 static void hystart_update(struct sock *sk, __u32 delay) 455 { 456 struct tcp_sock *tp = tcp_sk(sk); 457 struct bpf_bictcp *ca = inet_csk_ca(sk); 458 __u32 threshold; 459 460 if (hystart_detect & HYSTART_ACK_TRAIN) { 461 __u32 now = bictcp_clock_us(sk); 462 463 /* first detection parameter - ack-train detection */ 464 if ((__s32)(now - ca->last_ack) <= hystart_ack_delta_us) { 465 ca->last_ack = now; 466 467 threshold = ca->delay_min + hystart_ack_delay(sk); 468 469 /* Hystart ack train triggers if we get ack past 470 * ca->delay_min/2. 471 * Pacing might have delayed packets up to RTT/2 472 * during slow start. 473 */ 474 if (sk->sk_pacing_status == SK_PACING_NONE) 475 threshold >>= 1; 476 477 if ((__s32)(now - ca->round_start) > threshold) { 478 ca->found = 1; 479 tp->snd_ssthresh = tp->snd_cwnd; 480 } 481 } 482 } 483 484 if (hystart_detect & HYSTART_DELAY) { 485 /* obtain the minimum delay of more than sampling packets */ 486 if (ca->curr_rtt > delay) 487 ca->curr_rtt = delay; 488 if (ca->sample_cnt < HYSTART_MIN_SAMPLES) { 489 ca->sample_cnt++; 490 } else { 491 if (ca->curr_rtt > ca->delay_min + 492 HYSTART_DELAY_THRESH(ca->delay_min >> 3)) { 493 ca->found = 1; 494 tp->snd_ssthresh = tp->snd_cwnd; 495 } 496 } 497 } 498 } 499 500 int bpf_cubic_acked_called = 0; 501 502 SEC("struct_ops") 503 void BPF_PROG(bpf_cubic_acked, struct sock *sk, const struct ack_sample *sample) 504 { 505 const struct tcp_sock *tp = tcp_sk(sk); 506 struct bpf_bictcp *ca = inet_csk_ca(sk); 507 __u32 delay; 508 509 bpf_cubic_acked_called = 1; 510 /* Some calls are for duplicates without timestamps */ 511 if (sample->rtt_us < 0) 512 return; 513 514 /* Discard delay samples right after fast recovery */ 515 if (ca->epoch_start && (__s32)(tcp_jiffies32 - ca->epoch_start) < HZ) 516 return; 517 518 delay = sample->rtt_us; 519 if (delay == 0) 520 delay = 1; 521 522 /* first time call or link delay decreases */ 523 if (ca->delay_min == 0 || ca->delay_min > delay) 524 ca->delay_min = delay; 525 526 /* hystart triggers when cwnd is larger than some threshold */ 527 if (!ca->found && tcp_in_slow_start(tp) && hystart && 528 tp->snd_cwnd >= hystart_low_window) 529 hystart_update(sk, delay); 530 } 531 532 extern __u32 tcp_reno_undo_cwnd(struct sock *sk) __ksym; 533 534 SEC("struct_ops") 535 __u32 BPF_PROG(bpf_cubic_undo_cwnd, struct sock *sk) 536 { 537 return tcp_reno_undo_cwnd(sk); 538 } 539 540 SEC(".struct_ops") 541 struct tcp_congestion_ops cubic = { 542 .init = (void *)bpf_cubic_init, 543 .ssthresh = (void *)bpf_cubic_recalc_ssthresh, 544 .cong_avoid = (void *)bpf_cubic_cong_avoid, 545 .set_state = (void *)bpf_cubic_state, 546 .undo_cwnd = (void *)bpf_cubic_undo_cwnd, 547 .cwnd_event = (void *)bpf_cubic_cwnd_event, 548 .pkts_acked = (void *)bpf_cubic_acked, 549 .name = "bpf_cubic", 550 }; 551