1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * TCP Vegas congestion control 4 * 5 * This is based on the congestion detection/avoidance scheme described in 6 * Lawrence S. Brakmo and Larry L. Peterson. 7 * "TCP Vegas: End to end congestion avoidance on a global internet." 8 * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480, 9 * October 1995. Available from: 10 * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps 11 * 12 * See http://www.cs.arizona.edu/xkernel/ for their implementation. 13 * The main aspects that distinguish this implementation from the 14 * Arizona Vegas implementation are: 15 * o We do not change the loss detection or recovery mechanisms of 16 * Linux in any way. Linux already recovers from losses quite well, 17 * using fine-grained timers, NewReno, and FACK. 18 * o To avoid the performance penalty imposed by increasing cwnd 19 * only every-other RTT during slow start, we increase during 20 * every RTT during slow start, just like Reno. 21 * o Largely to allow continuous cwnd growth during slow start, 22 * we use the rate at which ACKs come back as the "actual" 23 * rate, rather than the rate at which data is sent. 24 * o To speed convergence to the right rate, we set the cwnd 25 * to achieve the right ("actual") rate when we exit slow start. 26 * o To filter out the noise caused by delayed ACKs, we use the 27 * minimum RTT sample observed during the last RTT to calculate 28 * the actual rate. 29 * o When the sender re-starts from idle, it waits until it has 30 * received ACKs for an entire flight of new data before making 31 * a cwnd adjustment decision. The original Vegas implementation 32 * assumed senders never went idle. 33 */ 34 35 #include <linux/mm.h> 36 #include <linux/module.h> 37 #include <linux/skbuff.h> 38 #include <linux/inet_diag.h> 39 40 #include <net/tcp.h> 41 42 #include "tcp_vegas.h" 43 44 static int alpha = 2; 45 static int beta = 4; 46 static int gamma = 1; 47 48 module_param(alpha, int, 0644); 49 MODULE_PARM_DESC(alpha, "lower bound of packets in network"); 50 module_param(beta, int, 0644); 51 MODULE_PARM_DESC(beta, "upper bound of packets in network"); 52 module_param(gamma, int, 0644); 53 MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)"); 54 55 /* There are several situations when we must "re-start" Vegas: 56 * 57 * o when a connection is established 58 * o after an RTO 59 * o after fast recovery 60 * o when we send a packet and there is no outstanding 61 * unacknowledged data (restarting an idle connection) 62 * 63 * In these circumstances we cannot do a Vegas calculation at the 64 * end of the first RTT, because any calculation we do is using 65 * stale info -- both the saved cwnd and congestion feedback are 66 * stale. 67 * 68 * Instead we must wait until the completion of an RTT during 69 * which we actually receive ACKs. 70 */ 71 static void vegas_enable(struct sock *sk) 72 { 73 const struct tcp_sock *tp = tcp_sk(sk); 74 struct vegas *vegas = inet_csk_ca(sk); 75 76 /* Begin taking Vegas samples next time we send something. */ 77 vegas->doing_vegas_now = 1; 78 79 /* Set the beginning of the next send window. */ 80 vegas->beg_snd_nxt = tp->snd_nxt; 81 82 vegas->cntRTT = 0; 83 vegas->minRTT = 0x7fffffff; 84 } 85 86 /* Stop taking Vegas samples for now. */ 87 static inline void vegas_disable(struct sock *sk) 88 { 89 struct vegas *vegas = inet_csk_ca(sk); 90 91 vegas->doing_vegas_now = 0; 92 } 93 94 void tcp_vegas_init(struct sock *sk) 95 { 96 struct vegas *vegas = inet_csk_ca(sk); 97 98 vegas->baseRTT = 0x7fffffff; 99 vegas_enable(sk); 100 } 101 EXPORT_SYMBOL_GPL(tcp_vegas_init); 102 103 /* Do RTT sampling needed for Vegas. 104 * Basically we: 105 * o min-filter RTT samples from within an RTT to get the current 106 * propagation delay + queuing delay (we are min-filtering to try to 107 * avoid the effects of delayed ACKs) 108 * o min-filter RTT samples from a much longer window (forever for now) 109 * to find the propagation delay (baseRTT) 110 */ 111 void tcp_vegas_pkts_acked(struct sock *sk, const struct ack_sample *sample) 112 { 113 struct vegas *vegas = inet_csk_ca(sk); 114 u32 vrtt; 115 116 if (sample->rtt_us < 0) 117 return; 118 119 /* Never allow zero rtt or baseRTT */ 120 vrtt = sample->rtt_us + 1; 121 122 /* Filter to find propagation delay: */ 123 if (vrtt < vegas->baseRTT) 124 vegas->baseRTT = vrtt; 125 126 /* Find the min RTT during the last RTT to find 127 * the current prop. delay + queuing delay: 128 */ 129 vegas->minRTT = min(vegas->minRTT, vrtt); 130 vegas->cntRTT++; 131 } 132 EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked); 133 134 void tcp_vegas_state(struct sock *sk, u8 ca_state) 135 { 136 if (ca_state == TCP_CA_Open) 137 vegas_enable(sk); 138 else 139 vegas_disable(sk); 140 } 141 EXPORT_SYMBOL_GPL(tcp_vegas_state); 142 143 /* 144 * If the connection is idle and we are restarting, 145 * then we don't want to do any Vegas calculations 146 * until we get fresh RTT samples. So when we 147 * restart, we reset our Vegas state to a clean 148 * slate. After we get acks for this flight of 149 * packets, _then_ we can make Vegas calculations 150 * again. 151 */ 152 void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event) 153 { 154 if (event == CA_EVENT_CWND_RESTART) 155 tcp_vegas_init(sk); 156 } 157 EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event); 158 159 void tcp_vegas_cwnd_event_tx_start(struct sock *sk) 160 { 161 tcp_vegas_init(sk); 162 } 163 EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event_tx_start); 164 165 static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp) 166 { 167 return min(tp->snd_ssthresh, tcp_snd_cwnd(tp)); 168 } 169 170 static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 acked) 171 { 172 struct tcp_sock *tp = tcp_sk(sk); 173 struct vegas *vegas = inet_csk_ca(sk); 174 175 if (!vegas->doing_vegas_now) { 176 tcp_reno_cong_avoid(sk, ack, acked); 177 return; 178 } 179 180 if (after(ack, vegas->beg_snd_nxt)) { 181 /* Do the Vegas once-per-RTT cwnd adjustment. */ 182 183 /* Save the extent of the current window so we can use this 184 * at the end of the next RTT. 185 */ 186 vegas->beg_snd_nxt = tp->snd_nxt; 187 188 /* We do the Vegas calculations only if we got enough RTT 189 * samples that we can be reasonably sure that we got 190 * at least one RTT sample that wasn't from a delayed ACK. 191 * If we only had 2 samples total, 192 * then that means we're getting only 1 ACK per RTT, which 193 * means they're almost certainly delayed ACKs. 194 * If we have 3 samples, we should be OK. 195 */ 196 197 if (vegas->cntRTT <= 2) { 198 /* We don't have enough RTT samples to do the Vegas 199 * calculation, so we'll behave like Reno. 200 */ 201 tcp_reno_cong_avoid(sk, ack, acked); 202 } else { 203 u32 rtt, diff; 204 u64 target_cwnd; 205 206 /* We have enough RTT samples, so, using the Vegas 207 * algorithm, we determine if we should increase or 208 * decrease cwnd, and by how much. 209 */ 210 211 /* Pluck out the RTT we are using for the Vegas 212 * calculations. This is the min RTT seen during the 213 * last RTT. Taking the min filters out the effects 214 * of delayed ACKs, at the cost of noticing congestion 215 * a bit later. 216 */ 217 rtt = vegas->minRTT; 218 219 /* Calculate the cwnd we should have, if we weren't 220 * going too fast. 221 * 222 * This is: 223 * (actual rate in segments) * baseRTT 224 */ 225 target_cwnd = (u64)tcp_snd_cwnd(tp) * vegas->baseRTT; 226 do_div(target_cwnd, rtt); 227 228 /* Calculate the difference between the window we had, 229 * and the window we would like to have. This quantity 230 * is the "Diff" from the Arizona Vegas papers. 231 */ 232 diff = tcp_snd_cwnd(tp) * (rtt-vegas->baseRTT) / vegas->baseRTT; 233 234 if (diff > gamma && tcp_in_slow_start(tp)) { 235 /* Going too fast. Time to slow down 236 * and switch to congestion avoidance. 237 */ 238 239 /* Set cwnd to match the actual rate 240 * exactly: 241 * cwnd = (actual rate) * baseRTT 242 * Then we add 1 because the integer 243 * truncation robs us of full link 244 * utilization. 245 */ 246 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), 247 (u32)target_cwnd + 1)); 248 WRITE_ONCE(tp->snd_ssthresh, 249 tcp_vegas_ssthresh(tp)); 250 251 } else if (tcp_in_slow_start(tp)) { 252 /* Slow start. */ 253 tcp_slow_start(tp, acked); 254 } else { 255 /* Congestion avoidance. */ 256 257 /* Figure out where we would like cwnd 258 * to be. 259 */ 260 if (diff > beta) { 261 /* The old window was too fast, so 262 * we slow down. 263 */ 264 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); 265 WRITE_ONCE(tp->snd_ssthresh, 266 tcp_vegas_ssthresh(tp)); 267 } else if (diff < alpha) { 268 /* We don't have enough extra packets 269 * in the network, so speed up. 270 */ 271 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1); 272 } else { 273 /* Sending just as fast as we 274 * should be. 275 */ 276 } 277 } 278 279 if (tcp_snd_cwnd(tp) < 2) 280 tcp_snd_cwnd_set(tp, 2); 281 else if (tcp_snd_cwnd(tp) > tp->snd_cwnd_clamp) 282 tcp_snd_cwnd_set(tp, tp->snd_cwnd_clamp); 283 284 WRITE_ONCE(tp->snd_ssthresh, tcp_current_ssthresh(sk)); 285 } 286 287 /* Wipe the slate clean for the next RTT. */ 288 vegas->cntRTT = 0; 289 vegas->minRTT = 0x7fffffff; 290 } 291 /* Use normal slow start */ 292 else if (tcp_in_slow_start(tp)) 293 tcp_slow_start(tp, acked); 294 } 295 296 /* Extract info for Tcp socket info provided via netlink. */ 297 size_t tcp_vegas_get_info(struct sock *sk, u32 ext, int *attr, 298 union tcp_cc_info *info) 299 { 300 const struct vegas *ca = inet_csk_ca(sk); 301 302 if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { 303 info->vegas.tcpv_enabled = ca->doing_vegas_now; 304 info->vegas.tcpv_rttcnt = ca->cntRTT; 305 info->vegas.tcpv_rtt = ca->baseRTT; 306 info->vegas.tcpv_minrtt = ca->minRTT; 307 308 *attr = INET_DIAG_VEGASINFO; 309 return sizeof(struct tcpvegas_info); 310 } 311 return 0; 312 } 313 EXPORT_SYMBOL_GPL(tcp_vegas_get_info); 314 315 static struct tcp_congestion_ops tcp_vegas __read_mostly = { 316 .init = tcp_vegas_init, 317 .ssthresh = tcp_reno_ssthresh, 318 .undo_cwnd = tcp_reno_undo_cwnd, 319 .cong_avoid = tcp_vegas_cong_avoid, 320 .pkts_acked = tcp_vegas_pkts_acked, 321 .set_state = tcp_vegas_state, 322 .cwnd_event = tcp_vegas_cwnd_event, 323 .cwnd_event_tx_start = tcp_vegas_cwnd_event_tx_start, 324 .get_info = tcp_vegas_get_info, 325 326 .owner = THIS_MODULE, 327 .name = "vegas", 328 }; 329 330 static int __init tcp_vegas_register(void) 331 { 332 BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE); 333 tcp_register_congestion_control(&tcp_vegas); 334 return 0; 335 } 336 337 static void __exit tcp_vegas_unregister(void) 338 { 339 tcp_unregister_congestion_control(&tcp_vegas); 340 } 341 342 module_init(tcp_vegas_register); 343 module_exit(tcp_vegas_unregister); 344 345 MODULE_AUTHOR("Stephen Hemminger"); 346 MODULE_LICENSE("GPL"); 347 MODULE_DESCRIPTION("TCP Vegas"); 348