linux/net/ipv4/tcp_vegas.c
<<
>>
Prefs
   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
  43static int alpha = 2;
  44static int beta  = 4;
  45static int gamma = 1;
  46
  47module_param(alpha, int, 0644);
  48MODULE_PARM_DESC(alpha, "lower bound of packets in network");
  49module_param(beta, int, 0644);
  50MODULE_PARM_DESC(beta, "upper bound of packets in network");
  51module_param(gamma, int, 0644);
  52MODULE_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 */
  71static 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. */
  87static 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
  94void 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}
 101EXPORT_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 */
 111void 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}
 132EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked);
 133
 134void 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}
 142EXPORT_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 */
 153void 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}
 159EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
 160
 161static inline u32 tcp_vegas_ssthresh(struct tcp_sock *tp)
 162{
 163        return  min(tp->snd_ssthresh, tp->snd_cwnd-1);
 164}
 165
 166static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 167{
 168        struct tcp_sock *tp = tcp_sk(sk);
 169        struct vegas *vegas = inet_csk_ca(sk);
 170
 171        if (!vegas->doing_vegas_now) {
 172                tcp_reno_cong_avoid(sk, ack, in_flight);
 173                return;
 174        }
 175
 176        if (after(ack, vegas->beg_snd_nxt)) {
 177                /* Do the Vegas once-per-RTT cwnd adjustment. */
 178
 179                /* Save the extent of the current window so we can use this
 180                 * at the end of the next RTT.
 181                 */
 182                vegas->beg_snd_nxt  = tp->snd_nxt;
 183
 184                /* We do the Vegas calculations only if we got enough RTT
 185                 * samples that we can be reasonably sure that we got
 186                 * at least one RTT sample that wasn't from a delayed ACK.
 187                 * If we only had 2 samples total,
 188                 * then that means we're getting only 1 ACK per RTT, which
 189                 * means they're almost certainly delayed ACKs.
 190                 * If  we have 3 samples, we should be OK.
 191                 */
 192
 193                if (vegas->cntRTT <= 2) {
 194                        /* We don't have enough RTT samples to do the Vegas
 195                         * calculation, so we'll behave like Reno.
 196                         */
 197                        tcp_reno_cong_avoid(sk, ack, in_flight);
 198                } else {
 199                        u32 rtt, diff;
 200                        u64 target_cwnd;
 201
 202                        /* We have enough RTT samples, so, using the Vegas
 203                         * algorithm, we determine if we should increase or
 204                         * decrease cwnd, and by how much.
 205                         */
 206
 207                        /* Pluck out the RTT we are using for the Vegas
 208                         * calculations. This is the min RTT seen during the
 209                         * last RTT. Taking the min filters out the effects
 210                         * of delayed ACKs, at the cost of noticing congestion
 211                         * a bit later.
 212                         */
 213                        rtt = vegas->minRTT;
 214
 215                        /* Calculate the cwnd we should have, if we weren't
 216                         * going too fast.
 217                         *
 218                         * This is:
 219                         *     (actual rate in segments) * baseRTT
 220                         */
 221                        target_cwnd = tp->snd_cwnd * vegas->baseRTT / rtt;
 222
 223                        /* Calculate the difference between the window we had,
 224                         * and the window we would like to have. This quantity
 225                         * is the "Diff" from the Arizona Vegas papers.
 226                         */
 227                        diff = tp->snd_cwnd * (rtt-vegas->baseRTT) / vegas->baseRTT;
 228
 229                        if (diff > gamma && tp->snd_cwnd <= tp->snd_ssthresh) {
 230                                /* Going too fast. Time to slow down
 231                                 * and switch to congestion avoidance.
 232                                 */
 233
 234                                /* Set cwnd to match the actual rate
 235                                 * exactly:
 236                                 *   cwnd = (actual rate) * baseRTT
 237                                 * Then we add 1 because the integer
 238                                 * truncation robs us of full link
 239                                 * utilization.
 240                                 */
 241                                tp->snd_cwnd = min(tp->snd_cwnd, (u32)target_cwnd+1);
 242                                tp->snd_ssthresh = tcp_vegas_ssthresh(tp);
 243
 244                        } else if (tp->snd_cwnd <= tp->snd_ssthresh) {
 245                                /* Slow start.  */
 246                                tcp_slow_start(tp);
 247                        } else {
 248                                /* Congestion avoidance. */
 249
 250                                /* Figure out where we would like cwnd
 251                                 * to be.
 252                                 */
 253                                if (diff > beta) {
 254                                        /* The old window was too fast, so
 255                                         * we slow down.
 256                                         */
 257                                        tp->snd_cwnd--;
 258                                        tp->snd_ssthresh
 259                                                = tcp_vegas_ssthresh(tp);
 260                                } else if (diff < alpha) {
 261                                        /* We don't have enough extra packets
 262                                         * in the network, so speed up.
 263                                         */
 264                                        tp->snd_cwnd++;
 265                                } else {
 266                                        /* Sending just as fast as we
 267                                         * should be.
 268                                         */
 269                                }
 270                        }
 271
 272                        if (tp->snd_cwnd < 2)
 273                                tp->snd_cwnd = 2;
 274                        else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
 275                                tp->snd_cwnd = tp->snd_cwnd_clamp;
 276
 277                        tp->snd_ssthresh = tcp_current_ssthresh(sk);
 278                }
 279
 280                /* Wipe the slate clean for the next RTT. */
 281                vegas->cntRTT = 0;
 282                vegas->minRTT = 0x7fffffff;
 283        }
 284        /* Use normal slow start */
 285        else if (tp->snd_cwnd <= tp->snd_ssthresh)
 286                tcp_slow_start(tp);
 287
 288}
 289
 290/* Extract info for Tcp socket info provided via netlink. */
 291void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
 292{
 293        const struct vegas *ca = inet_csk_ca(sk);
 294        if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 295                struct tcpvegas_info info = {
 296                        .tcpv_enabled = ca->doing_vegas_now,
 297                        .tcpv_rttcnt = ca->cntRTT,
 298                        .tcpv_rtt = ca->baseRTT,
 299                        .tcpv_minrtt = ca->minRTT,
 300                };
 301
 302                nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 303        }
 304}
 305EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
 306
 307static struct tcp_congestion_ops tcp_vegas __read_mostly = {
 308        .flags          = TCP_CONG_RTT_STAMP,
 309        .init           = tcp_vegas_init,
 310        .ssthresh       = tcp_reno_ssthresh,
 311        .cong_avoid     = tcp_vegas_cong_avoid,
 312        .min_cwnd       = tcp_reno_min_cwnd,
 313        .pkts_acked     = tcp_vegas_pkts_acked,
 314        .set_state      = tcp_vegas_state,
 315        .cwnd_event     = tcp_vegas_cwnd_event,
 316        .get_info       = tcp_vegas_get_info,
 317
 318        .owner          = THIS_MODULE,
 319        .name           = "vegas",
 320};
 321
 322static int __init tcp_vegas_register(void)
 323{
 324        BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
 325        tcp_register_congestion_control(&tcp_vegas);
 326        return 0;
 327}
 328
 329static void __exit tcp_vegas_unregister(void)
 330{
 331        tcp_unregister_congestion_control(&tcp_vegas);
 332}
 333
 334module_init(tcp_vegas_register);
 335module_exit(tcp_vegas_unregister);
 336
 337MODULE_AUTHOR("Stephen Hemminger");
 338MODULE_LICENSE("GPL");
 339MODULE_DESCRIPTION("TCP Vegas");
 340
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.