1 /* 2 * TCP NV: TCP with Congestion Avoidance 3 * 4 * TCP-NV is a successor of TCP-Vegas that has been developed to 5 * deal with the issues that occur in modern networks. 6 * Like TCP-Vegas, TCP-NV supports true congestion avoidance, 7 * the ability to detect congestion before packet losses occur. 8 * When congestion (queue buildup) starts to occur, TCP-NV 9 * predicts what the cwnd size should be for the current 10 * throughput and it reduces the cwnd proportionally to 11 * the difference between the current cwnd and the predicted cwnd. 12 * 13 * NV is only recommeneded for traffic within a data center, and when 14 * all the flows are NV (at least those within the data center). This 15 * is due to the inherent unfairness between flows using losses to 16 * detect congestion (congestion control) and those that use queue 17 * buildup to detect congestion (congestion avoidance). 18 * 19 * Note: High NIC coalescence values may lower the performance of NV 20 * due to the increased noise in RTT values. In particular, we have 21 * seen issues with rx-frames values greater than 8. 22 * 23 * TODO: 24 * 1) Add mechanism to deal with reverse congestion. 25 */ 26 27 #include <linux/mm.h> 28 #include <linux/module.h> 29 #include <linux/math64.h> 30 #include <net/tcp.h> 31 #include <linux/inet_diag.h> 32 33 /* TCP NV parameters 34 * 35 * nv_pad Max number of queued packets allowed in network 36 * nv_pad_buffer Do not grow cwnd if this closed to nv_pad 37 * nv_reset_period How often (in) seconds)to reset min_rtt 38 * nv_min_cwnd Don't decrease cwnd below this if there are no losses 39 * nv_cong_dec_mult Decrease cwnd by X% (30%) of congestion when detected 40 * nv_ssthresh_factor On congestion set ssthresh to this * <desired cwnd> / 8 41 * nv_rtt_factor RTT averaging factor 42 * nv_loss_dec_factor Decrease cwnd to this (80%) when losses occur 43 * nv_dec_eval_min_calls Wait this many RTT measurements before dec cwnd 44 * nv_inc_eval_min_calls Wait this many RTT measurements before inc cwnd 45 * nv_ssthresh_eval_min_calls Wait this many RTT measurements before stopping 46 * slow-start due to congestion 47 * nv_stop_rtt_cnt Only grow cwnd for this many RTTs after non-congestion 48 * nv_rtt_min_cnt Wait these many RTTs before making congesion decision 49 * nv_cwnd_growth_rate_neg 50 * nv_cwnd_growth_rate_pos 51 * How quickly to double growth rate (not rate) of cwnd when not 52 * congested. One value (nv_cwnd_growth_rate_neg) for when 53 * rate < 1 pkt/RTT (after losses). The other (nv_cwnd_growth_rate_pos) 54 * otherwise. 55 */ 56 57 static int nv_pad __read_mostly = 10; 58 static int nv_pad_buffer __read_mostly = 2; 59 static int nv_reset_period __read_mostly = 5; /* in seconds */ 60 static int nv_min_cwnd __read_mostly = 2; 61 static int nv_cong_dec_mult __read_mostly = 30 * 128 / 100; /* = 30% */ 62 static int nv_ssthresh_factor __read_mostly = 8; /* = 1 */ 63 static int nv_rtt_factor __read_mostly = 128; /* = 1/2*old + 1/2*new */ 64 static int nv_loss_dec_factor __read_mostly = 819; /* => 80% */ 65 static int nv_cwnd_growth_rate_neg __read_mostly = 8; 66 static int nv_cwnd_growth_rate_pos __read_mostly; /* 0 => fixed like Reno */ 67 static int nv_dec_eval_min_calls __read_mostly = 60; 68 static int nv_inc_eval_min_calls __read_mostly = 20; 69 static int nv_ssthresh_eval_min_calls __read_mostly = 30; 70 static int nv_stop_rtt_cnt __read_mostly = 10; 71 static int nv_rtt_min_cnt __read_mostly = 2; 72 73 module_param(nv_pad, int, 0644); 74 MODULE_PARM_DESC(nv_pad, "max queued packets allowed in network"); 75 module_param(nv_reset_period, int, 0644); 76 MODULE_PARM_DESC(nv_reset_period, "nv_min_rtt reset period (secs)"); 77 module_param(nv_min_cwnd, int, 0644); 78 MODULE_PARM_DESC(nv_min_cwnd, "NV will not decrease cwnd below this value" 79 " without losses"); 80 81 /* TCP NV Parameters */ 82 struct tcpnv { 83 unsigned long nv_min_rtt_reset_jiffies; /* when to switch to 84 * nv_min_rtt_new */ 85 s8 cwnd_growth_factor; /* Current cwnd growth factor, 86 * < 0 => less than 1 packet/RTT */ 87 u8 available8; 88 u16 available16; 89 u8 nv_allow_cwnd_growth:1, /* whether cwnd can grow */ 90 nv_reset:1, /* whether to reset values */ 91 nv_catchup:1; /* whether we are growing because 92 * of temporary cwnd decrease */ 93 u8 nv_eval_call_cnt; /* call count since last eval */ 94 u8 nv_min_cwnd; /* nv won't make a ca decision if cwnd is 95 * smaller than this. It may grow to handle 96 * TSO, LRO and interrupt coalescence because 97 * with these a small cwnd cannot saturate 98 * the link. Note that this is different from 99 * the file local nv_min_cwnd */ 100 u8 nv_rtt_cnt; /* RTTs without making ca decision */; 101 u32 nv_last_rtt; /* last rtt */ 102 u32 nv_min_rtt; /* active min rtt. Used to determine slope */ 103 u32 nv_min_rtt_new; /* min rtt for future use */ 104 u32 nv_base_rtt; /* If non-zero it represents the threshold for 105 * congestion */ 106 u32 nv_lower_bound_rtt; /* Used in conjunction with nv_base_rtt. It is 107 * set to 80% of nv_base_rtt. It helps reduce 108 * unfairness between flows */ 109 u32 nv_rtt_max_rate; /* max rate seen during current RTT */ 110 u32 nv_rtt_start_seq; /* current RTT ends when packet arrives 111 * acking beyond nv_rtt_start_seq */ 112 u32 nv_last_snd_una; /* Previous value of tp->snd_una. It is 113 * used to determine bytes acked since last 114 * call to bictcp_acked */ 115 u32 nv_no_cong_cnt; /* Consecutive no congestion decisions */ 116 }; 117 118 #define NV_INIT_RTT U32_MAX 119 #define NV_MIN_CWND 4 120 #define NV_MIN_CWND_GROW 2 121 #define NV_TSO_CWND_BOUND 80 122 123 static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk) 124 { 125 struct tcp_sock *tp = tcp_sk(sk); 126 127 ca->nv_reset = 0; 128 ca->nv_no_cong_cnt = 0; 129 ca->nv_rtt_cnt = 0; 130 ca->nv_last_rtt = 0; 131 ca->nv_rtt_max_rate = 0; 132 ca->nv_rtt_start_seq = tp->snd_una; 133 ca->nv_eval_call_cnt = 0; 134 ca->nv_last_snd_una = tp->snd_una; 135 } 136 137 static void tcpnv_init(struct sock *sk) 138 { 139 struct tcpnv *ca = inet_csk_ca(sk); 140 int base_rtt; 141 142 tcpnv_reset(ca, sk); 143 144 /* See if base_rtt is available from socket_ops bpf program. 145 * It is meant to be used in environments, such as communication 146 * within a datacenter, where we have reasonable estimates of 147 * RTTs 148 */ 149 base_rtt = tcp_call_bpf(sk, BPF_SOCK_OPS_BASE_RTT); 150 if (base_rtt > 0) { 151 ca->nv_base_rtt = base_rtt; 152 ca->nv_lower_bound_rtt = (base_rtt * 205) >> 8; /* 80% */ 153 } else { 154 ca->nv_base_rtt = 0; 155 ca->nv_lower_bound_rtt = 0; 156 } 157 158 ca->nv_allow_cwnd_growth = 1; 159 ca->nv_min_rtt_reset_jiffies = jiffies + 2 * HZ; 160 ca->nv_min_rtt = NV_INIT_RTT; 161 ca->nv_min_rtt_new = NV_INIT_RTT; 162 ca->nv_min_cwnd = NV_MIN_CWND; 163 ca->nv_catchup = 0; 164 ca->cwnd_growth_factor = 0; 165 } 166 167 /* If provided, apply upper (base_rtt) and lower (lower_bound_rtt) 168 * bounds to RTT. 169 */ 170 inline u32 nv_get_bounded_rtt(struct tcpnv *ca, u32 val) 171 { 172 if (ca->nv_lower_bound_rtt > 0 && val < ca->nv_lower_bound_rtt) 173 return ca->nv_lower_bound_rtt; 174 else if (ca->nv_base_rtt > 0 && val > ca->nv_base_rtt) 175 return ca->nv_base_rtt; 176 else 177 return val; 178 } 179 180 static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked) 181 { 182 struct tcp_sock *tp = tcp_sk(sk); 183 struct tcpnv *ca = inet_csk_ca(sk); 184 u32 cnt; 185 186 if (!tcp_is_cwnd_limited(sk)) 187 return; 188 189 /* Only grow cwnd if NV has not detected congestion */ 190 if (!ca->nv_allow_cwnd_growth) 191 return; 192 193 if (tcp_in_slow_start(tp)) { 194 acked = tcp_slow_start(tp, acked); 195 if (!acked) 196 return; 197 } 198 199 if (ca->cwnd_growth_factor < 0) { 200 cnt = tp->snd_cwnd << -ca->cwnd_growth_factor; 201 tcp_cong_avoid_ai(tp, cnt, acked); 202 } else { 203 cnt = max(4U, tp->snd_cwnd >> ca->cwnd_growth_factor); 204 tcp_cong_avoid_ai(tp, cnt, acked); 205 } 206 } 207 208 static u32 tcpnv_recalc_ssthresh(struct sock *sk) 209 { 210 const struct tcp_sock *tp = tcp_sk(sk); 211 212 return max((tp->snd_cwnd * nv_loss_dec_factor) >> 10, 2U); 213 } 214 215 static void tcpnv_state(struct sock *sk, u8 new_state) 216 { 217 struct tcpnv *ca = inet_csk_ca(sk); 218 219 if (new_state == TCP_CA_Open && ca->nv_reset) { 220 tcpnv_reset(ca, sk); 221 } else if (new_state == TCP_CA_Loss || new_state == TCP_CA_CWR || 222 new_state == TCP_CA_Recovery) { 223 ca->nv_reset = 1; 224 ca->nv_allow_cwnd_growth = 0; 225 if (new_state == TCP_CA_Loss) { 226 /* Reset cwnd growth factor to Reno value */ 227 if (ca->cwnd_growth_factor > 0) 228 ca->cwnd_growth_factor = 0; 229 /* Decrease growth rate if allowed */ 230 if (nv_cwnd_growth_rate_neg > 0 && 231 ca->cwnd_growth_factor > -8) 232 ca->cwnd_growth_factor--; 233 } 234 } 235 } 236 237 /* Do congestion avoidance calculations for TCP-NV 238 */ 239 static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample) 240 { 241 const struct inet_connection_sock *icsk = inet_csk(sk); 242 struct tcp_sock *tp = tcp_sk(sk); 243 struct tcpnv *ca = inet_csk_ca(sk); 244 unsigned long now = jiffies; 245 s64 rate64 = 0; 246 u32 rate, max_win, cwnd_by_slope; 247 u32 avg_rtt; 248 u32 bytes_acked = 0; 249 250 /* Some calls are for duplicates without timetamps */ 251 if (sample->rtt_us < 0) 252 return; 253 254 /* If not in TCP_CA_Open or TCP_CA_Disorder states, skip. */ 255 if (icsk->icsk_ca_state != TCP_CA_Open && 256 icsk->icsk_ca_state != TCP_CA_Disorder) 257 return; 258 259 /* Stop cwnd growth if we were in catch up mode */ 260 if (ca->nv_catchup && tp->snd_cwnd >= nv_min_cwnd) { 261 ca->nv_catchup = 0; 262 ca->nv_allow_cwnd_growth = 0; 263 } 264 265 bytes_acked = tp->snd_una - ca->nv_last_snd_una; 266 ca->nv_last_snd_una = tp->snd_una; 267 268 if (sample->in_flight == 0) 269 return; 270 271 /* Calculate moving average of RTT */ 272 if (nv_rtt_factor > 0) { 273 if (ca->nv_last_rtt > 0) { 274 avg_rtt = (((u64)sample->rtt_us) * nv_rtt_factor + 275 ((u64)ca->nv_last_rtt) 276 * (256 - nv_rtt_factor)) >> 8; 277 } else { 278 avg_rtt = sample->rtt_us; 279 ca->nv_min_rtt = avg_rtt << 1; 280 } 281 ca->nv_last_rtt = avg_rtt; 282 } else { 283 avg_rtt = sample->rtt_us; 284 } 285 286 /* rate in 100's bits per second */ 287 rate64 = ((u64)sample->in_flight) * 8000000; 288 rate = (u32)div64_u64(rate64, (u64)(avg_rtt * 100)); 289 290 /* Remember the maximum rate seen during this RTT 291 * Note: It may be more than one RTT. This function should be 292 * called at least nv_dec_eval_min_calls times. 293 */ 294 if (ca->nv_rtt_max_rate < rate) 295 ca->nv_rtt_max_rate = rate; 296 297 /* We have valid information, increment counter */ 298 if (ca->nv_eval_call_cnt < 255) 299 ca->nv_eval_call_cnt++; 300 301 /* Apply bounds to rtt. Only used to update min_rtt */ 302 avg_rtt = nv_get_bounded_rtt(ca, avg_rtt); 303 304 /* update min rtt if necessary */ 305 if (avg_rtt < ca->nv_min_rtt) 306 ca->nv_min_rtt = avg_rtt; 307 308 /* update future min_rtt if necessary */ 309 if (avg_rtt < ca->nv_min_rtt_new) 310 ca->nv_min_rtt_new = avg_rtt; 311 312 /* nv_min_rtt is updated with the minimum (possibley averaged) rtt 313 * seen in the last sysctl_tcp_nv_reset_period seconds (i.e. a 314 * warm reset). This new nv_min_rtt will be continued to be updated 315 * and be used for another sysctl_tcp_nv_reset_period seconds, 316 * when it will be updated again. 317 * In practice we introduce some randomness, so the actual period used 318 * is chosen randomly from the range: 319 * [sysctl_tcp_nv_reset_period*3/4, sysctl_tcp_nv_reset_period*5/4) 320 */ 321 if (time_after_eq(now, ca->nv_min_rtt_reset_jiffies)) { 322 unsigned char rand; 323 324 ca->nv_min_rtt = ca->nv_min_rtt_new; 325 ca->nv_min_rtt_new = NV_INIT_RTT; 326 get_random_bytes(&rand, 1); 327 ca->nv_min_rtt_reset_jiffies = 328 now + ((nv_reset_period * (384 + rand) * HZ) >> 9); 329 /* Every so often we decrease ca->nv_min_cwnd in case previous 330 * value is no longer accurate. 331 */ 332 ca->nv_min_cwnd = max(ca->nv_min_cwnd / 2, NV_MIN_CWND); 333 } 334 335 /* Once per RTT check if we need to do congestion avoidance */ 336 if (before(ca->nv_rtt_start_seq, tp->snd_una)) { 337 ca->nv_rtt_start_seq = tp->snd_nxt; 338 if (ca->nv_rtt_cnt < 0xff) 339 /* Increase counter for RTTs without CA decision */ 340 ca->nv_rtt_cnt++; 341 342 /* If this function is only called once within an RTT 343 * the cwnd is probably too small (in some cases due to 344 * tso, lro or interrupt coalescence), so we increase 345 * ca->nv_min_cwnd. 346 */ 347 if (ca->nv_eval_call_cnt == 1 && 348 bytes_acked >= (ca->nv_min_cwnd - 1) * tp->mss_cache && 349 ca->nv_min_cwnd < (NV_TSO_CWND_BOUND + 1)) { 350 ca->nv_min_cwnd = min(ca->nv_min_cwnd 351 + NV_MIN_CWND_GROW, 352 NV_TSO_CWND_BOUND + 1); 353 ca->nv_rtt_start_seq = tp->snd_nxt + 354 ca->nv_min_cwnd * tp->mss_cache; 355 ca->nv_eval_call_cnt = 0; 356 ca->nv_allow_cwnd_growth = 1; 357 return; 358 } 359 360 /* Find the ideal cwnd for current rate from slope 361 * slope = 80000.0 * mss / nv_min_rtt 362 * cwnd_by_slope = nv_rtt_max_rate / slope 363 */ 364 cwnd_by_slope = (u32) 365 div64_u64(((u64)ca->nv_rtt_max_rate) * ca->nv_min_rtt, 366 (u64)(80000 * tp->mss_cache)); 367 max_win = cwnd_by_slope + nv_pad; 368 369 /* If cwnd > max_win, decrease cwnd 370 * if cwnd < max_win, grow cwnd 371 * else leave the same 372 */ 373 if (tp->snd_cwnd > max_win) { 374 /* there is congestion, check that it is ok 375 * to make a CA decision 376 * 1. We should have at least nv_dec_eval_min_calls 377 * data points before making a CA decision 378 * 2. We only make a congesion decision after 379 * nv_rtt_min_cnt RTTs 380 */ 381 if (ca->nv_rtt_cnt < nv_rtt_min_cnt) { 382 return; 383 } else if (tp->snd_ssthresh == TCP_INFINITE_SSTHRESH) { 384 if (ca->nv_eval_call_cnt < 385 nv_ssthresh_eval_min_calls) 386 return; 387 /* otherwise we will decrease cwnd */ 388 } else if (ca->nv_eval_call_cnt < 389 nv_dec_eval_min_calls) { 390 if (ca->nv_allow_cwnd_growth && 391 ca->nv_rtt_cnt > nv_stop_rtt_cnt) 392 ca->nv_allow_cwnd_growth = 0; 393 return; 394 } 395 396 /* We have enough data to determine we are congested */ 397 ca->nv_allow_cwnd_growth = 0; 398 tp->snd_ssthresh = 399 (nv_ssthresh_factor * max_win) >> 3; 400 if (tp->snd_cwnd - max_win > 2) { 401 /* gap > 2, we do exponential cwnd decrease */ 402 int dec; 403 404 dec = max(2U, ((tp->snd_cwnd - max_win) * 405 nv_cong_dec_mult) >> 7); 406 tp->snd_cwnd -= dec; 407 } else if (nv_cong_dec_mult > 0) { 408 tp->snd_cwnd = max_win; 409 } 410 if (ca->cwnd_growth_factor > 0) 411 ca->cwnd_growth_factor = 0; 412 ca->nv_no_cong_cnt = 0; 413 } else if (tp->snd_cwnd <= max_win - nv_pad_buffer) { 414 /* There is no congestion, grow cwnd if allowed*/ 415 if (ca->nv_eval_call_cnt < nv_inc_eval_min_calls) 416 return; 417 418 ca->nv_allow_cwnd_growth = 1; 419 ca->nv_no_cong_cnt++; 420 if (ca->cwnd_growth_factor < 0 && 421 nv_cwnd_growth_rate_neg > 0 && 422 ca->nv_no_cong_cnt > nv_cwnd_growth_rate_neg) { 423 ca->cwnd_growth_factor++; 424 ca->nv_no_cong_cnt = 0; 425 } else if (ca->cwnd_growth_factor >= 0 && 426 nv_cwnd_growth_rate_pos > 0 && 427 ca->nv_no_cong_cnt > 428 nv_cwnd_growth_rate_pos) { 429 ca->cwnd_growth_factor++; 430 ca->nv_no_cong_cnt = 0; 431 } 432 } else { 433 /* cwnd is in-between, so do nothing */ 434 return; 435 } 436 437 /* update state */ 438 ca->nv_eval_call_cnt = 0; 439 ca->nv_rtt_cnt = 0; 440 ca->nv_rtt_max_rate = 0; 441 442 /* Don't want to make cwnd < nv_min_cwnd 443 * (it wasn't before, if it is now is because nv 444 * decreased it). 445 */ 446 if (tp->snd_cwnd < nv_min_cwnd) 447 tp->snd_cwnd = nv_min_cwnd; 448 } 449 } 450 451 /* Extract info for Tcp socket info provided via netlink */ 452 static size_t tcpnv_get_info(struct sock *sk, u32 ext, int *attr, 453 union tcp_cc_info *info) 454 { 455 const struct tcpnv *ca = inet_csk_ca(sk); 456 457 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 458 info->vegas.tcpv_enabled = 1; 459 info->vegas.tcpv_rttcnt = ca->nv_rtt_cnt; 460 info->vegas.tcpv_rtt = ca->nv_last_rtt; 461 info->vegas.tcpv_minrtt = ca->nv_min_rtt; 462 463 *attr = INET_DIAG_VEGASINFO; 464 return sizeof(struct tcpvegas_info); 465 } 466 return 0; 467 } 468 469 static struct tcp_congestion_ops tcpnv __read_mostly = { 470 .init = tcpnv_init, 471 .ssthresh = tcpnv_recalc_ssthresh, 472 .cong_avoid = tcpnv_cong_avoid, 473 .set_state = tcpnv_state, 474 .undo_cwnd = tcp_reno_undo_cwnd, 475 .pkts_acked = tcpnv_acked, 476 .get_info = tcpnv_get_info, 477 478 .owner = THIS_MODULE, 479 .name = "nv", 480 }; 481 482 static int __init tcpnv_register(void) 483 { 484 BUILD_BUG_ON(sizeof(struct tcpnv) > ICSK_CA_PRIV_SIZE); 485 486 return tcp_register_congestion_control(&tcpnv); 487 } 488 489 static void __exit tcpnv_unregister(void) 490 { 491 tcp_unregister_congestion_control(&tcpnv); 492 } 493 494 module_init(tcpnv_register); 495 module_exit(tcpnv_unregister); 496 497 MODULE_AUTHOR("Lawrence Brakmo"); 498 MODULE_LICENSE("GPL"); 499 MODULE_DESCRIPTION("TCP NV"); 500 MODULE_VERSION("1.0"); 501