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 /* Default values of the Vegas variables, in fixed-point representation
  44  * with V_PARAM_SHIFT bits to the right of the binary point.
  45  */
  46 #define V_PARAM_SHIFT 1
  47 static int alpha = 2<<V_PARAM_SHIFT;
  48 static int beta  = 4<<V_PARAM_SHIFT;
  49 static int gamma = 1<<V_PARAM_SHIFT;
  50
  51 module_param(alpha, int, 0644);
  52 MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)");
  53 module_param(beta, int, 0644);
  54 MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)");
  55 module_param(gamma, int, 0644);
  56 MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
  57
  58
  59 /* There are several situations when we must "re-start" Vegas:
  60  *
  61  *  o when a connection is established
  62  *  o after an RTO
  63  *  o after fast recovery
  64  *  o when we send a packet and there is no outstanding
  65  *    unacknowledged data (restarting an idle connection)
  66  *
  67  * In these circumstances we cannot do a Vegas calculation at the
  68  * end of the first RTT, because any calculation we do is using
  69  * stale info -- both the saved cwnd and congestion feedback are
  70  * stale.
  71  *
  72  * Instead we must wait until the completion of an RTT during
  73  * which we actually receive ACKs.
  74  */
  75 static void vegas_enable(struct sock *sk)
  76 {
  77         const struct tcp_sock *tp = tcp_sk(sk);
  78         struct vegas *vegas = inet_csk_ca(sk);
  79
  80         /* Begin taking Vegas samples next time we send something. */
  81         vegas->doing_vegas_now = 1;
  82
  83         /* Set the beginning of the next send window. */
  84         vegas->beg_snd_nxt = tp->snd_nxt;
  85
  86         vegas->cntRTT = 0;
  87         vegas->minRTT = 0x7fffffff;
  88 }
  89
  90 /* Stop taking Vegas samples for now. */
  91 static inline void vegas_disable(struct sock *sk)
  92 {
  93         struct vegas *vegas = inet_csk_ca(sk);
  94
  95         vegas->doing_vegas_now = 0;
  96 }
  97
  98 void tcp_vegas_init(struct sock *sk)
  99 {
 100         struct vegas *vegas = inet_csk_ca(sk);
 101
 102         vegas->baseRTT = 0x7fffffff;
 103         vegas_enable(sk);
 104 }
 105 EXPORT_SYMBOL_GPL(tcp_vegas_init);
 106
 107 /* Do RTT sampling needed for Vegas.
 108  * Basically we:
 109  *   o min-filter RTT samples from within an RTT to get the current
 110  *     propagation delay + queuing delay (we are min-filtering to try to
 111  *     avoid the effects of delayed ACKs)
 112  *   o min-filter RTT samples from a much longer window (forever for now)
 113  *     to find the propagation delay (baseRTT)
 114  */
 115 void tcp_vegas_pkts_acked(struct sock *sk, u32 cnt, s32 rtt_us)
 116 {
 117         struct vegas *vegas = inet_csk_ca(sk);
 118         u32 vrtt;
 119
 120         if (rtt_us < 0)
 121                 return;
 122
 123         /* Never allow zero rtt or baseRTT */
 124         vrtt = rtt_us + 1;
 125
 126         /* Filter to find propagation delay: */
 127         if (vrtt < vegas->baseRTT)
 128                 vegas->baseRTT = vrtt;
 129
 130         /* Find the min RTT during the last RTT to find
 131          * the current prop. delay + queuing delay:
 132          */
 133         vegas->minRTT = min(vegas->minRTT, vrtt);
 134         vegas->cntRTT++;
 135 }
 136 EXPORT_SYMBOL_GPL(tcp_vegas_pkts_acked);
 137
 138 void tcp_vegas_state(struct sock *sk, u8 ca_state)
 139 {
 140
 141         if (ca_state == TCP_CA_Open)
 142                 vegas_enable(sk);
 143         else
 144                 vegas_disable(sk);
 145 }
 146 EXPORT_SYMBOL_GPL(tcp_vegas_state);
 147
 148 /*
 149  * If the connection is idle and we are restarting,
 150  * then we don't want to do any Vegas calculations
 151  * until we get fresh RTT samples.  So when we
 152  * restart, we reset our Vegas state to a clean
 153  * slate. After we get acks for this flight of
 154  * packets, _then_ we can make Vegas calculations
 155  * again.
 156  */
 157 void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 158 {
 159         if (event == CA_EVENT_CWND_RESTART ||
 160             event == CA_EVENT_TX_START)
 161                 tcp_vegas_init(sk);
 162 }
 163 EXPORT_SYMBOL_GPL(tcp_vegas_cwnd_event);
 164
 165 static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 166 {
 167         struct tcp_sock *tp = tcp_sk(sk);
 168         struct vegas *vegas = inet_csk_ca(sk);
 169
 170         if (!vegas->doing_vegas_now)
 171                 return tcp_reno_cong_avoid(sk, ack, in_flight);
 172
 173         /* The key players are v_beg_snd_una and v_beg_snd_nxt.
 174          *
 175          * These are so named because they represent the approximate values
 176          * of snd_una and snd_nxt at the beginning of the current RTT. More
 177          * precisely, they represent the amount of data sent during the RTT.
 178          * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
 179          * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
 180          * bytes of data have been ACKed during the course of the RTT, giving
 181          * an "actual" rate of:
 182          *
 183          *     (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
 184          *
 185          * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
 186          * because delayed ACKs can cover more than one segment, so they
 187          * don't line up nicely with the boundaries of RTTs.
 188          *
 189          * Another unfortunate fact of life is that delayed ACKs delay the
 190          * advance of the left edge of our send window, so that the number
 191          * of bytes we send in an RTT is often less than our cwnd will allow.
 192          * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
 193          */
 194
 195         if (after(ack, vegas->beg_snd_nxt)) {
 196                 /* Do the Vegas once-per-RTT cwnd adjustment. */
 197                 u32 old_wnd, old_snd_cwnd;
 198
 199
 200                 /* Here old_wnd is essentially the window of data that was
 201                  * sent during the previous RTT, and has all
 202                  * been acknowledged in the course of the RTT that ended
 203                  * with the ACK we just received. Likewise, old_snd_cwnd
 204                  * is the cwnd during the previous RTT.
 205                  */
 206                 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
 207                         tp->mss_cache;
 208                 old_snd_cwnd = vegas->beg_snd_cwnd;
 209
 210                 /* Save the extent of the current window so we can use this
 211                  * at the end of the next RTT.
 212                  */
 213                 vegas->beg_snd_una  = vegas->beg_snd_nxt;
 214                 vegas->beg_snd_nxt  = tp->snd_nxt;
 215                 vegas->beg_snd_cwnd = tp->snd_cwnd;
 216
 217                 /* We do the Vegas calculations only if we got enough RTT
 218                  * samples that we can be reasonably sure that we got
 219                  * at least one RTT sample that wasn't from a delayed ACK.
 220                  * If we only had 2 samples total,
 221                  * then that means we're getting only 1 ACK per RTT, which
 222                  * means they're almost certainly delayed ACKs.
 223                  * If  we have 3 samples, we should be OK.
 224                  */
 225
 226                 if (vegas->cntRTT <= 2) {
 227                         /* We don't have enough RTT samples to do the Vegas
 228                          * calculation, so we'll behave like Reno.
 229                          */
 230                         tcp_reno_cong_avoid(sk, ack, in_flight);
 231                 } else {
 232                         u32 rtt, target_cwnd, diff;
 233
 234                         /* We have enough RTT samples, so, using the Vegas
 235                          * algorithm, we determine if we should increase or
 236                          * decrease cwnd, and by how much.
 237                          */
 238
 239                         /* Pluck out the RTT we are using for the Vegas
 240                          * calculations. This is the min RTT seen during the
 241                          * last RTT. Taking the min filters out the effects
 242                          * of delayed ACKs, at the cost of noticing congestion
 243                          * a bit later.
 244                          */
 245                         rtt = vegas->minRTT;
 246
 247                         /* Calculate the cwnd we should have, if we weren't
 248                          * going too fast.
 249                          *
 250                          * This is:
 251                          *     (actual rate in segments) * baseRTT
 252                          * We keep it as a fixed point number with
 253                          * V_PARAM_SHIFT bits to the right of the binary point.
 254                          */
 255                         target_cwnd = ((old_wnd * vegas->baseRTT)
 256                                        << V_PARAM_SHIFT) / rtt;
 257
 258                         /* Calculate the difference between the window we had,
 259                          * and the window we would like to have. This quantity
 260                          * is the "Diff" from the Arizona Vegas papers.
 261                          *
 262                          * Again, this is a fixed point number with
 263                          * V_PARAM_SHIFT bits to the right of the binary
 264                          * point.
 265                          */
 266                         diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
 267
 268                         if (diff > gamma && tp->snd_ssthresh > 2 ) {
 269                                 /* Going too fast. Time to slow down
 270                                  * and switch to congestion avoidance.
 271                                  */
 272                                 tp->snd_ssthresh = 2;
 273
 274                                 /* Set cwnd to match the actual rate
 275                                  * exactly:
 276                                  *   cwnd = (actual rate) * baseRTT
 277                                  * Then we add 1 because the integer
 278                                  * truncation robs us of full link
 279                                  * utilization.
 280                                  */
 281                                 tp->snd_cwnd = min(tp->snd_cwnd,
 282                                                    (target_cwnd >>
 283                                                     V_PARAM_SHIFT)+1);
 284
 285                         } else if (tp->snd_cwnd <= tp->snd_ssthresh) {
 286                                 /* Slow start.  */
 287                                 tcp_slow_start(tp);
 288                         } else {
 289                                 /* Congestion avoidance. */
 290                                 u32 next_snd_cwnd;
 291
 292                                 /* Figure out where we would like cwnd
 293                                  * to be.
 294                                  */
 295                                 if (diff > beta) {
 296                                         /* The old window was too fast, so
 297                                          * we slow down.
 298                                          */
 299                                         next_snd_cwnd = old_snd_cwnd - 1;
 300                                 } else if (diff < alpha) {
 301                                         /* We don't have enough extra packets
 302                                          * in the network, so speed up.
 303                                          */
 304                                         next_snd_cwnd = old_snd_cwnd + 1;
 305                                 } else {
 306                                         /* Sending just as fast as we
 307                                          * should be.
 308                                          */
 309                                         next_snd_cwnd = old_snd_cwnd;
 310                                 }
 311
 312                                 /* Adjust cwnd upward or downward, toward the
 313                                  * desired value.
 314                                  */
 315                                 if (next_snd_cwnd > tp->snd_cwnd)
 316                                         tp->snd_cwnd++;
 317                                 else if (next_snd_cwnd < tp->snd_cwnd)
 318                                         tp->snd_cwnd--;
 319                         }
 320
 321                         if (tp->snd_cwnd < 2)
 322                                 tp->snd_cwnd = 2;
 323                         else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
 324                                 tp->snd_cwnd = tp->snd_cwnd_clamp;
 325                 }
 326
 327                 /* Wipe the slate clean for the next RTT. */
 328                 vegas->cntRTT = 0;
 329                 vegas->minRTT = 0x7fffffff;
 330         }
 331         /* Use normal slow start */
 332         else if (tp->snd_cwnd <= tp->snd_ssthresh)
 333                 tcp_slow_start(tp);
 334
 335 }
 336
 337 /* Extract info for Tcp socket info provided via netlink. */
 338 void tcp_vegas_get_info(struct sock *sk, u32 ext, struct sk_buff *skb)
 339 {
 340         const struct vegas *ca = inet_csk_ca(sk);
 341         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 342                 struct tcpvegas_info info = {
 343                         .tcpv_enabled = ca->doing_vegas_now,
 344                         .tcpv_rttcnt = ca->cntRTT,
 345                         .tcpv_rtt = ca->baseRTT,
 346                         .tcpv_minrtt = ca->minRTT,
 347                 };
 348
 349                 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 350         }
 351 }
 352 EXPORT_SYMBOL_GPL(tcp_vegas_get_info);
 353
 354 static struct tcp_congestion_ops tcp_vegas = {
 355         .flags          = TCP_CONG_RTT_STAMP,
 356         .init           = tcp_vegas_init,
 357         .ssthresh       = tcp_reno_ssthresh,
 358         .cong_avoid     = tcp_vegas_cong_avoid,
 359         .min_cwnd       = tcp_reno_min_cwnd,
 360         .pkts_acked     = tcp_vegas_pkts_acked,
 361         .set_state      = tcp_vegas_state,
 362         .cwnd_event     = tcp_vegas_cwnd_event,
 363         .get_info       = tcp_vegas_get_info,
 364
 365         .owner          = THIS_MODULE,
 366         .name           = "vegas",
 367 };
 368
 369 static int __init tcp_vegas_register(void)
 370 {
 371         BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
 372         tcp_register_congestion_control(&tcp_vegas);
 373         return 0;
 374 }
 375
 376 static void __exit tcp_vegas_unregister(void)
 377 {
 378         tcp_unregister_congestion_control(&tcp_vegas);
 379 }
 380
 381 module_init(tcp_vegas_register);
 382 module_exit(tcp_vegas_unregister);
 383
 384 MODULE_AUTHOR("Stephen Hemminger");
 385 MODULE_LICENSE("GPL");
 386 MODULE_DESCRIPTION("TCP Vegas");