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 by this (50%) 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 = 512; /* => 50% */ 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_rtt_max_rate; /* max rate seen during current RTT */ 105 u32 nv_rtt_start_seq; /* current RTT ends when packet arrives 106 * acking beyond nv_rtt_start_seq */ 107 u32 nv_last_snd_una; /* Previous value of tp->snd_una. It is 108 * used to determine bytes acked since last 109 * call to bictcp_acked */ 110 u32 nv_no_cong_cnt; /* Consecutive no congestion decisions */ 111 }; 112 113 #define NV_INIT_RTT U32_MAX 114 #define NV_MIN_CWND 4 115 #define NV_MIN_CWND_GROW 2 116 #define NV_TSO_CWND_BOUND 80 117 118 static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk) 119 { 120 struct tcp_sock *tp = tcp_sk(sk); 121 122 ca->nv_reset = 0; 123 ca->nv_no_cong_cnt = 0; 124 ca->nv_rtt_cnt = 0; 125 ca->nv_last_rtt = 0; 126 ca->nv_rtt_max_rate = 0; 127 ca->nv_rtt_start_seq = tp->snd_una; 128 ca->nv_eval_call_cnt = 0; 129 ca->nv_last_snd_una = tp->snd_una; 130 } 131 132 static void tcpnv_init(struct sock *sk) 133 { 134 struct tcpnv *ca = inet_csk_ca(sk); 135 136 tcpnv_reset(ca, sk); 137 138 ca->nv_allow_cwnd_growth = 1; 139 ca->nv_min_rtt_reset_jiffies = jiffies + 2 * HZ; 140 ca->nv_min_rtt = NV_INIT_RTT; 141 ca->nv_min_rtt_new = NV_INIT_RTT; 142 ca->nv_min_cwnd = NV_MIN_CWND; 143 ca->nv_catchup = 0; 144 ca->cwnd_growth_factor = 0; 145 } 146 147 static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked) 148 { 149 struct tcp_sock *tp = tcp_sk(sk); 150 struct tcpnv *ca = inet_csk_ca(sk); 151 u32 cnt; 152 153 if (!tcp_is_cwnd_limited(sk)) 154 return; 155 156 /* Only grow cwnd if NV has not detected congestion */ 157 if (!ca->nv_allow_cwnd_growth) 158 return; 159 160 if (tcp_in_slow_start(tp)) { 161 acked = tcp_slow_start(tp, acked); 162 if (!acked) 163 return; 164 } 165 166 if (ca->cwnd_growth_factor < 0) { 167 cnt = tp->snd_cwnd << -ca->cwnd_growth_factor; 168 tcp_cong_avoid_ai(tp, cnt, acked); 169 } else { 170 cnt = max(4U, tp->snd_cwnd >> ca->cwnd_growth_factor); 171 tcp_cong_avoid_ai(tp, cnt, acked); 172 } 173 } 174 175 static u32 tcpnv_recalc_ssthresh(struct sock *sk) 176 { 177 const struct tcp_sock *tp = tcp_sk(sk); 178 179 return max((tp->snd_cwnd * nv_loss_dec_factor) >> 10, 2U); 180 } 181 182 static void tcpnv_state(struct sock *sk, u8 new_state) 183 { 184 struct tcpnv *ca = inet_csk_ca(sk); 185 186 if (new_state == TCP_CA_Open && ca->nv_reset) { 187 tcpnv_reset(ca, sk); 188 } else if (new_state == TCP_CA_Loss || new_state == TCP_CA_CWR || 189 new_state == TCP_CA_Recovery) { 190 ca->nv_reset = 1; 191 ca->nv_allow_cwnd_growth = 0; 192 if (new_state == TCP_CA_Loss) { 193 /* Reset cwnd growth factor to Reno value */ 194 if (ca->cwnd_growth_factor > 0) 195 ca->cwnd_growth_factor = 0; 196 /* Decrease growth rate if allowed */ 197 if (nv_cwnd_growth_rate_neg > 0 && 198 ca->cwnd_growth_factor > -8) 199 ca->cwnd_growth_factor--; 200 } 201 } 202 } 203 204 /* Do congestion avoidance calculations for TCP-NV 205 */ 206 static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample) 207 { 208 const struct inet_connection_sock *icsk = inet_csk(sk); 209 struct tcp_sock *tp = tcp_sk(sk); 210 struct tcpnv *ca = inet_csk_ca(sk); 211 unsigned long now = jiffies; 212 s64 rate64 = 0; 213 u32 rate, max_win, cwnd_by_slope; 214 u32 avg_rtt; 215 u32 bytes_acked = 0; 216 217 /* Some calls are for duplicates without timetamps */ 218 if (sample->rtt_us < 0) 219 return; 220 221 /* If not in TCP_CA_Open or TCP_CA_Disorder states, skip. */ 222 if (icsk->icsk_ca_state != TCP_CA_Open && 223 icsk->icsk_ca_state != TCP_CA_Disorder) 224 return; 225 226 /* Stop cwnd growth if we were in catch up mode */ 227 if (ca->nv_catchup && tp->snd_cwnd >= nv_min_cwnd) { 228 ca->nv_catchup = 0; 229 ca->nv_allow_cwnd_growth = 0; 230 } 231 232 bytes_acked = tp->snd_una - ca->nv_last_snd_una; 233 ca->nv_last_snd_una = tp->snd_una; 234 235 if (sample->in_flight == 0) 236 return; 237 238 /* Calculate moving average of RTT */ 239 if (nv_rtt_factor > 0) { 240 if (ca->nv_last_rtt > 0) { 241 avg_rtt = (((u64)sample->rtt_us) * nv_rtt_factor + 242 ((u64)ca->nv_last_rtt) 243 * (256 - nv_rtt_factor)) >> 8; 244 } else { 245 avg_rtt = sample->rtt_us; 246 ca->nv_min_rtt = avg_rtt << 1; 247 } 248 ca->nv_last_rtt = avg_rtt; 249 } else { 250 avg_rtt = sample->rtt_us; 251 } 252 253 /* rate in 100's bits per second */ 254 rate64 = ((u64)sample->in_flight) * 8000000; 255 rate = (u32)div64_u64(rate64, (u64)(avg_rtt * 100)); 256 257 /* Remember the maximum rate seen during this RTT 258 * Note: It may be more than one RTT. This function should be 259 * called at least nv_dec_eval_min_calls times. 260 */ 261 if (ca->nv_rtt_max_rate < rate) 262 ca->nv_rtt_max_rate = rate; 263 264 /* We have valid information, increment counter */ 265 if (ca->nv_eval_call_cnt < 255) 266 ca->nv_eval_call_cnt++; 267 268 /* update min rtt if necessary */ 269 if (avg_rtt < ca->nv_min_rtt) 270 ca->nv_min_rtt = avg_rtt; 271 272 /* update future min_rtt if necessary */ 273 if (avg_rtt < ca->nv_min_rtt_new) 274 ca->nv_min_rtt_new = avg_rtt; 275 276 /* nv_min_rtt is updated with the minimum (possibley averaged) rtt 277 * seen in the last sysctl_tcp_nv_reset_period seconds (i.e. a 278 * warm reset). This new nv_min_rtt will be continued to be updated 279 * and be used for another sysctl_tcp_nv_reset_period seconds, 280 * when it will be updated again. 281 * In practice we introduce some randomness, so the actual period used 282 * is chosen randomly from the range: 283 * [sysctl_tcp_nv_reset_period*3/4, sysctl_tcp_nv_reset_period*5/4) 284 */ 285 if (time_after_eq(now, ca->nv_min_rtt_reset_jiffies)) { 286 unsigned char rand; 287 288 ca->nv_min_rtt = ca->nv_min_rtt_new; 289 ca->nv_min_rtt_new = NV_INIT_RTT; 290 get_random_bytes(&rand, 1); 291 ca->nv_min_rtt_reset_jiffies = 292 now + ((nv_reset_period * (384 + rand) * HZ) >> 9); 293 /* Every so often we decrease ca->nv_min_cwnd in case previous 294 * value is no longer accurate. 295 */ 296 ca->nv_min_cwnd = max(ca->nv_min_cwnd / 2, NV_MIN_CWND); 297 } 298 299 /* Once per RTT check if we need to do congestion avoidance */ 300 if (before(ca->nv_rtt_start_seq, tp->snd_una)) { 301 ca->nv_rtt_start_seq = tp->snd_nxt; 302 if (ca->nv_rtt_cnt < 0xff) 303 /* Increase counter for RTTs without CA decision */ 304 ca->nv_rtt_cnt++; 305 306 /* If this function is only called once within an RTT 307 * the cwnd is probably too small (in some cases due to 308 * tso, lro or interrupt coalescence), so we increase 309 * ca->nv_min_cwnd. 310 */ 311 if (ca->nv_eval_call_cnt == 1 && 312 bytes_acked >= (ca->nv_min_cwnd - 1) * tp->mss_cache && 313 ca->nv_min_cwnd < (NV_TSO_CWND_BOUND + 1)) { 314 ca->nv_min_cwnd = min(ca->nv_min_cwnd 315 + NV_MIN_CWND_GROW, 316 NV_TSO_CWND_BOUND + 1); 317 ca->nv_rtt_start_seq = tp->snd_nxt + 318 ca->nv_min_cwnd * tp->mss_cache; 319 ca->nv_eval_call_cnt = 0; 320 ca->nv_allow_cwnd_growth = 1; 321 return; 322 } 323 324 /* Find the ideal cwnd for current rate from slope 325 * slope = 80000.0 * mss / nv_min_rtt 326 * cwnd_by_slope = nv_rtt_max_rate / slope 327 */ 328 cwnd_by_slope = (u32) 329 div64_u64(((u64)ca->nv_rtt_max_rate) * ca->nv_min_rtt, 330 (u64)(80000 * tp->mss_cache)); 331 max_win = cwnd_by_slope + nv_pad; 332 333 /* If cwnd > max_win, decrease cwnd 334 * if cwnd < max_win, grow cwnd 335 * else leave the same 336 */ 337 if (tp->snd_cwnd > max_win) { 338 /* there is congestion, check that it is ok 339 * to make a CA decision 340 * 1. We should have at least nv_dec_eval_min_calls 341 * data points before making a CA decision 342 * 2. We only make a congesion decision after 343 * nv_rtt_min_cnt RTTs 344 */ 345 if (ca->nv_rtt_cnt < nv_rtt_min_cnt) { 346 return; 347 } else if (tp->snd_ssthresh == TCP_INFINITE_SSTHRESH) { 348 if (ca->nv_eval_call_cnt < 349 nv_ssthresh_eval_min_calls) 350 return; 351 /* otherwise we will decrease cwnd */ 352 } else if (ca->nv_eval_call_cnt < 353 nv_dec_eval_min_calls) { 354 if (ca->nv_allow_cwnd_growth && 355 ca->nv_rtt_cnt > nv_stop_rtt_cnt) 356 ca->nv_allow_cwnd_growth = 0; 357 return; 358 } 359 360 /* We have enough data to determine we are congested */ 361 ca->nv_allow_cwnd_growth = 0; 362 tp->snd_ssthresh = 363 (nv_ssthresh_factor * max_win) >> 3; 364 if (tp->snd_cwnd - max_win > 2) { 365 /* gap > 2, we do exponential cwnd decrease */ 366 int dec; 367 368 dec = max(2U, ((tp->snd_cwnd - max_win) * 369 nv_cong_dec_mult) >> 7); 370 tp->snd_cwnd -= dec; 371 } else if (nv_cong_dec_mult > 0) { 372 tp->snd_cwnd = max_win; 373 } 374 if (ca->cwnd_growth_factor > 0) 375 ca->cwnd_growth_factor = 0; 376 ca->nv_no_cong_cnt = 0; 377 } else if (tp->snd_cwnd <= max_win - nv_pad_buffer) { 378 /* There is no congestion, grow cwnd if allowed*/ 379 if (ca->nv_eval_call_cnt < nv_inc_eval_min_calls) 380 return; 381 382 ca->nv_allow_cwnd_growth = 1; 383 ca->nv_no_cong_cnt++; 384 if (ca->cwnd_growth_factor < 0 && 385 nv_cwnd_growth_rate_neg > 0 && 386 ca->nv_no_cong_cnt > nv_cwnd_growth_rate_neg) { 387 ca->cwnd_growth_factor++; 388 ca->nv_no_cong_cnt = 0; 389 } else if (ca->cwnd_growth_factor >= 0 && 390 nv_cwnd_growth_rate_pos > 0 && 391 ca->nv_no_cong_cnt > 392 nv_cwnd_growth_rate_pos) { 393 ca->cwnd_growth_factor++; 394 ca->nv_no_cong_cnt = 0; 395 } 396 } else { 397 /* cwnd is in-between, so do nothing */ 398 return; 399 } 400 401 /* update state */ 402 ca->nv_eval_call_cnt = 0; 403 ca->nv_rtt_cnt = 0; 404 ca->nv_rtt_max_rate = 0; 405 406 /* Don't want to make cwnd < nv_min_cwnd 407 * (it wasn't before, if it is now is because nv 408 * decreased it). 409 */ 410 if (tp->snd_cwnd < nv_min_cwnd) 411 tp->snd_cwnd = nv_min_cwnd; 412 } 413 } 414 415 /* Extract info for Tcp socket info provided via netlink */ 416 static size_t tcpnv_get_info(struct sock *sk, u32 ext, int *attr, 417 union tcp_cc_info *info) 418 { 419 const struct tcpnv *ca = inet_csk_ca(sk); 420 421 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 422 info->vegas.tcpv_enabled = 1; 423 info->vegas.tcpv_rttcnt = ca->nv_rtt_cnt; 424 info->vegas.tcpv_rtt = ca->nv_last_rtt; 425 info->vegas.tcpv_minrtt = ca->nv_min_rtt; 426 427 *attr = INET_DIAG_VEGASINFO; 428 return sizeof(struct tcpvegas_info); 429 } 430 return 0; 431 } 432 433 static struct tcp_congestion_ops tcpnv __read_mostly = { 434 .init = tcpnv_init, 435 .ssthresh = tcpnv_recalc_ssthresh, 436 .cong_avoid = tcpnv_cong_avoid, 437 .set_state = tcpnv_state, 438 .undo_cwnd = tcp_reno_undo_cwnd, 439 .pkts_acked = tcpnv_acked, 440 .get_info = tcpnv_get_info, 441 442 .owner = THIS_MODULE, 443 .name = "nv", 444 }; 445 446 static int __init tcpnv_register(void) 447 { 448 BUILD_BUG_ON(sizeof(struct tcpnv) > ICSK_CA_PRIV_SIZE); 449 450 return tcp_register_congestion_control(&tcpnv); 451 } 452 453 static void __exit tcpnv_unregister(void) 454 { 455 tcp_unregister_congestion_control(&tcpnv); 456 } 457 458 module_init(tcpnv_register); 459 module_exit(tcpnv_unregister); 460 461 MODULE_AUTHOR("Lawrence Brakmo"); 462 MODULE_LICENSE("GPL"); 463 MODULE_DESCRIPTION("TCP NV"); 464 MODULE_VERSION("1.0"); 465