net/ipv4/tcp_vegas.c

   1 /*
   2  * TCP Vegas congestion control
   3  *
   4  * This is based on the congestion detection/avoidance scheme described in
   5  *    Lawrence S. Brakmo and Larry L. Peterson.
   6  *    "TCP Vegas: End to end congestion avoidance on a global internet."
   7  *    IEEE Journal on Selected Areas in Communication, 13(8):1465--1480,
   8  *    October 1995. Available from:
   9  *      ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps
  10  *
  11  * See http://www.cs.arizona.edu/xkernel/ for their implementation.
  12  * The main aspects that distinguish this implementation from the
  13  * Arizona Vegas implementation are:
  14  *   o We do not change the loss detection or recovery mechanisms of
  15  *     Linux in any way. Linux already recovers from losses quite well,
  16  *     using fine-grained timers, NewReno, and FACK.
  17  *   o To avoid the performance penalty imposed by increasing cwnd
  18  *     only every-other RTT during slow start, we increase during
  19  *     every RTT during slow start, just like Reno.
  20  *   o Largely to allow continuous cwnd growth during slow start,
  21  *     we use the rate at which ACKs come back as the "actual"
  22  *     rate, rather than the rate at which data is sent.
  23  *   o To speed convergence to the right rate, we set the cwnd
  24  *     to achieve the right ("actual") rate when we exit slow start.
  25  *   o To filter out the noise caused by delayed ACKs, we use the
  26  *     minimum RTT sample observed during the last RTT to calculate
  27  *     the actual rate.
  28  *   o When the sender re-starts from idle, it waits until it has
  29  *     received ACKs for an entire flight of new data before making
  30  *     a cwnd adjustment decision. The original Vegas implementation
  31  *     assumed senders never went idle.
  32  */
  33
  34 #include <linux/mm.h>
  35 #include <linux/module.h>
  36 #include <linux/skbuff.h>
  37 #include <linux/inet_diag.h>
  38
  39 #include <net/tcp.h>
  40
  41 #include "tcp_vegas.h"
  42
  43 static int alpha = 2;
  44 static int beta  = 4;
  45 static int gamma = 1;
  46
  47 module_param(alpha, int, 0644);
  48 MODULE_PARM_DESC(alpha, "lower bound of packets in network");
  49 module_param(beta, int, 0644);
  50 MODULE_PARM_DESC(beta, "upper bound of packets in network");
  51 module_param(gamma, int, 0644);
  52 MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
  53
  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, u32 cnt, s32 rtt_us)
 112 {
 113         struct vegas *vegas = inet_csk_ca(sk);
 114         u32 vrtt;
 115
 116         if (rtt_us < 0)
 117                 return;
 118
 119         /* Never allow zero rtt or baseRTT */
 120         vrtt = 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
 137         if (ca_state == TCP_CA_Open)
 138                 vegas_enable(sk);
 139         else
 140                 vegas_disable(sk);
 141 }
 142 EXPORT_SYMBOL_GPL(tcp_vegas_state);
 143
 144 /*
 145  * If the connection is idle and we are restarting,
 146  * then we don't want to do any Vegas calculations
 147  * until we get fresh RTT samples.  So when we
 148  * restart, we reset our Vegas state to a clean
 149  * slate. After we get acks for this flight of
 150  * packets, _then_ we can make Vegas calculations
 151  * again.
 152  */
 153 void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 154 {
 155         if (event == CA_EVENT_CWND_RESTART ||
 156             event == CA_EVENT_TX_START)
 157                 tcp_vegas_init(sk);
 158 }
 159 EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
 160
 161 static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 162 {
 163         struct tcp_sock *tp = tcp_sk(sk);
 164         struct vegas *vegas = inet_csk_ca(sk);
 165
 166         if (!vegas->doing_vegas_now) {
 167                 tcp_reno_cong_avoid(sk, ack, in_flight);
 168                 return;
 169         }
 170
 171         if (after(ack, vegas->beg_snd_nxt)) {
 172                 /* Do the Vegas once-per-RTT cwnd adjustment. */
 173
 174                 /* Save the extent of the current window so we can use this
 175                  * at the end of the next RTT.
 176                  */
 177                 vegas->beg_snd_nxt  = tp->snd_nxt;
 178
 179                 /* We do the Vegas calculations only if we got enough RTT
 180                  * samples that we can be reasonably sure that we got
 181                  * at least one RTT sample that wasn't from a delayed ACK.
 182                  * If we only had 2 samples total,
 183                  * then that means we're getting only 1 ACK per RTT, which
 184                  * means they're almost certainly delayed ACKs.
 185                  * If  we have 3 samples, we should be OK.
 186                  */
 187
 188                 if (vegas->cntRTT <= 2) {
 189                         /* We don't have enough RTT samples to do the Vegas
 190                          * calculation, so we'll behave like Reno.
 191                          */
 192                         tcp_reno_cong_avoid(sk, ack, in_flight);
 193                 } else {
 194                         u32 rtt, diff;
 195                         u64 target_cwnd;
 196
 197                         /* We have enough RTT samples, so, using the Vegas
 198                          * algorithm, we determine if we should increase or
 199                          * decrease cwnd, and by how much.
 200                          */
 201
 202                         /* Pluck out the RTT we are using for the Vegas
 203                          * calculations. This is the min RTT seen during the
 204                          * last RTT. Taking the min filters out the effects
 205                          * of delayed ACKs, at the cost of noticing congestion
 206                          * a bit later.
 207                          */
 208                         rtt = vegas->minRTT;
 209
 210                         /* Calculate the cwnd we should have, if we weren't
 211                          * going too fast.
 212                          *
 213                          * This is:
 214                          *     (actual rate in segments) * baseRTT
 215                          */
 216                         target_cwnd = tp->snd_cwnd * vegas->baseRTT / rtt;
 217
 218                         /* Calculate the difference between the window we had,
 219                          * and the window we would like to have. This quantity
 220                          * is the "Diff" from the Arizona Vegas papers.
 221                          */
 222                         diff = tp->snd_cwnd * (rtt-vegas->baseRTT) / vegas->baseRTT;
 223
 224                         if (diff > gamma && tp->snd_ssthresh > 2 ) {
 225                                 /* Going too fast. Time to slow down
 226                                  * and switch to congestion avoidance.
 227                                  */
 228                                 tp->snd_ssthresh = 2;
 229
 230                                 /* Set cwnd to match the actual rate
 231                                  * exactly:
 232                                  *   cwnd = (actual rate) * baseRTT
 233                                  * Then we add 1 because the integer
 234                                  * truncation robs us of full link
 235                                  * utilization.
 236                                  */
 237                                 tp->snd_cwnd = min(tp->snd_cwnd, (u32)target_cwnd+1);
 238
 239                         } else if (tp->snd_cwnd <= tp->snd_ssthresh) {
 240                                 /* Slow start.  */
 241                                 tcp_slow_start(tp);
 242                         } else {
 243                                 /* Congestion avoidance. */
 244
 245                                 /* Figure out where we would like cwnd
 246                                  * to be.
 247                                  */
 248                                 if (diff > beta) {
 249                                         /* The old window was too fast, so
 250                                          * we slow down.
 251                                          */
 252                                         tp->snd_cwnd--;
 253                                 } else if (diff < alpha) {
 254                                         /* We don't have enough extra packets
 255                                          * in the network, so speed up.
 256                                          */
 257                                         tp->snd_cwnd++;
 258                                 } else {
 259                                         /* Sending just as fast as we
 260                                          * should be.
 261                                          */
 262                                 }
 263                         }
 264
 265                         if (tp->snd_cwnd < 2)
 266                                 tp->snd_cwnd = 2;
 267                         else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
 268                                 tp->snd_cwnd = tp->snd_cwnd_clamp;
 269
 270                         tp->snd_ssthresh = tcp_current_ssthresh(sk);
 271                 }
 272
 273                 /* Wipe the slate clean for the next RTT. */
 274                 vegas->cntRTT = 0;
 275                 vegas->minRTT = 0x7fffffff;
 276         }
 277         /* Use normal slow start */
 278         else if (tp->snd_cwnd <= tp->snd_ssthresh)
 279                 tcp_slow_start(tp);
 280
 281 }
 282
 283 /* Extract info for Tcp socket info provided via netlink. */
 284 void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
 285 {
 286         const struct vegas *ca = inet_csk_ca(sk);
 287         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 288                 struct tcpvegas_info info = {
 289                         .tcpv_enabled = ca->doing_vegas_now,
 290                         .tcpv_rttcnt = ca->cntRTT,
 291                         .tcpv_rtt = ca->baseRTT,
 292                         .tcpv_minrtt = ca->minRTT,
 293                 };
 294
 295                 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 296         }
 297 }
 298 EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
 299
 300 static struct tcp_congestion_ops tcp_vegas = {
 301         .flags          = TCP_CONG_RTT_STAMP,
 302         .init           = tcp_vegas_init,
 303         .ssthresh       = tcp_reno_ssthresh,
 304         .cong_avoid     = tcp_vegas_cong_avoid,
 305         .min_cwnd       = tcp_reno_min_cwnd,
 306         .pkts_acked     = tcp_vegas_pkts_acked,
 307         .set_state      = tcp_vegas_state,
 308         .cwnd_event     = tcp_vegas_cwnd_event,
 309         .get_info       = tcp_vegas_get_info,
 310
 311         .owner          = THIS_MODULE,
 312         .name           = "vegas",
 313 };
 314
 315 static int __init tcp_vegas_register(void)
 316 {
 317         BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
 318         tcp_register_congestion_control(&tcp_vegas);
 319         return 0;
 320 }
 321
 322 static void __exit tcp_vegas_unregister(void)
 323 {
 324         tcp_unregister_congestion_control(&tcp_vegas);
 325 }
 326
 327 module_init(tcp_vegas_register);
 328 module_exit(tcp_vegas_unregister);
 329
 330 MODULE_AUTHOR("Stephen Hemminger");
 331 MODULE_LICENSE("GPL");
 332 MODULE_DESCRIPTION("TCP Vegas");