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 /* Default values of the Vegas variables, in fixed-point representation
  42  * with V_PARAM_SHIFT bits to the right of the binary point.
  43  */
  44 #define V_PARAM_SHIFT 1
  45 static int alpha = 2<<V_PARAM_SHIFT;
  46 static int beta  = 4<<V_PARAM_SHIFT;
  47 static int gamma = 1<<V_PARAM_SHIFT;
  48
  49 module_param(alpha, int, 0644);
  50 MODULE_PARM_DESC(alpha, "lower bound of packets in network (scale by 2)");
  51 module_param(beta, int, 0644);
  52 MODULE_PARM_DESC(beta, "upper bound of packets in network (scale by 2)");
  53 module_param(gamma, int, 0644);
  54 MODULE_PARM_DESC(gamma, "limit on increase (scale by 2)");
  55
  56
  57 /* Vegas variables */
  58 struct vegas {
  59         u32     beg_snd_nxt;    /* right edge during last RTT */
  60         u32     beg_snd_una;    /* left edge  during last RTT */
  61         u32     beg_snd_cwnd;   /* saves the size of the cwnd */
  62         u8      doing_vegas_now;/* if true, do vegas for this RTT */
  63         u16     cntRTT;         /* # of RTTs measured within last RTT */
  64         u32     minRTT;         /* min of RTTs measured within last RTT (in usec) */
  65         u32     baseRTT;        /* the min of all Vegas RTT measurements seen (in usec) */
  66 };
  67
  68 /* There are several situations when we must "re-start" Vegas:
  69  *
  70  *  o when a connection is established
  71  *  o after an RTO
  72  *  o after fast recovery
  73  *  o when we send a packet and there is no outstanding
  74  *    unacknowledged data (restarting an idle connection)
  75  *
  76  * In these circumstances we cannot do a Vegas calculation at the
  77  * end of the first RTT, because any calculation we do is using
  78  * stale info -- both the saved cwnd and congestion feedback are
  79  * stale.
  80  *
  81  * Instead we must wait until the completion of an RTT during
  82  * which we actually receive ACKs.
  83  */
  84 static inline void vegas_enable(struct sock *sk)
  85 {
  86         const struct tcp_sock *tp = tcp_sk(sk);
  87         struct vegas *vegas = inet_csk_ca(sk);
  88
  89         /* Begin taking Vegas samples next time we send something. */
  90         vegas->doing_vegas_now = 1;
  91
  92         /* Set the beginning of the next send window. */
  93         vegas->beg_snd_nxt = tp->snd_nxt;
  94
  95         vegas->cntRTT = 0;
  96         vegas->minRTT = 0x7fffffff;
  97 }
  98
  99 /* Stop taking Vegas samples for now. */
 100 static inline void vegas_disable(struct sock *sk)
 101 {
 102         struct vegas *vegas = inet_csk_ca(sk);
 103
 104         vegas->doing_vegas_now = 0;
 105 }
 106
 107 static void tcp_vegas_init(struct sock *sk)
 108 {
 109         struct vegas *vegas = inet_csk_ca(sk);
 110
 111         vegas->baseRTT = 0x7fffffff;
 112         vegas_enable(sk);
 113 }
 114
 115 /* Do RTT sampling needed for Vegas.
 116  * Basically we:
 117  *   o min-filter RTT samples from within an RTT to get the current
 118  *     propagation delay + queuing delay (we are min-filtering to try to
 119  *     avoid the effects of delayed ACKs)
 120  *   o min-filter RTT samples from a much longer window (forever for now)
 121  *     to find the propagation delay (baseRTT)
 122  */
 123 static void tcp_vegas_pkts_acked(struct sock *sk, u32 cnt, ktime_t last)
 124 {
 125         struct vegas *vegas = inet_csk_ca(sk);
 126         u32 vrtt;
 127
 128         /* Never allow zero rtt or baseRTT */
 129         vrtt = (ktime_to_ns(net_timedelta(last)) / NSEC_PER_USEC) + 1;
 130
 131         /* Filter to find propagation delay: */
 132         if (vrtt < vegas->baseRTT)
 133                 vegas->baseRTT = vrtt;
 134
 135         /* Find the min RTT during the last RTT to find
 136          * the current prop. delay + queuing delay:
 137          */
 138         vegas->minRTT = min(vegas->minRTT, vrtt);
 139         vegas->cntRTT++;
 140 }
 141
 142 static void tcp_vegas_state(struct sock *sk, u8 ca_state)
 143 {
 144
 145         if (ca_state == TCP_CA_Open)
 146                 vegas_enable(sk);
 147         else
 148                 vegas_disable(sk);
 149 }
 150
 151 /*
 152  * If the connection is idle and we are restarting,
 153  * then we don't want to do any Vegas calculations
 154  * until we get fresh RTT samples.  So when we
 155  * restart, we reset our Vegas state to a clean
 156  * slate. After we get acks for this flight of
 157  * packets, _then_ we can make Vegas calculations
 158  * again.
 159  */
 160 static void tcp_vegas_cwnd_event(struct sock *sk, enum tcp_ca_event event)
 161 {
 162         if (event == CA_EVENT_CWND_RESTART ||
 163             event == CA_EVENT_TX_START)
 164                 tcp_vegas_init(sk);
 165 }
 166
 167 static void tcp_vegas_cong_avoid(struct sock *sk, u32 ack,
 168                                  u32 seq_rtt, u32 in_flight, int flag)
 169 {
 170         struct tcp_sock *tp = tcp_sk(sk);
 171         struct vegas *vegas = inet_csk_ca(sk);
 172
 173         if (!vegas->doing_vegas_now)
 174                 return tcp_reno_cong_avoid(sk, ack, seq_rtt, in_flight, flag);
 175
 176         /* The key players are v_beg_snd_una and v_beg_snd_nxt.
 177          *
 178          * These are so named because they represent the approximate values
 179          * of snd_una and snd_nxt at the beginning of the current RTT. More
 180          * precisely, they represent the amount of data sent during the RTT.
 181          * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt,
 182          * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding
 183          * bytes of data have been ACKed during the course of the RTT, giving
 184          * an "actual" rate of:
 185          *
 186          *     (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration)
 187          *
 188          * Unfortunately, v_beg_snd_una is not exactly equal to snd_una,
 189          * because delayed ACKs can cover more than one segment, so they
 190          * don't line up nicely with the boundaries of RTTs.
 191          *
 192          * Another unfortunate fact of life is that delayed ACKs delay the
 193          * advance of the left edge of our send window, so that the number
 194          * of bytes we send in an RTT is often less than our cwnd will allow.
 195          * So we keep track of our cwnd separately, in v_beg_snd_cwnd.
 196          */
 197
 198         if (after(ack, vegas->beg_snd_nxt)) {
 199                 /* Do the Vegas once-per-RTT cwnd adjustment. */
 200                 u32 old_wnd, old_snd_cwnd;
 201
 202
 203                 /* Here old_wnd is essentially the window of data that was
 204                  * sent during the previous RTT, and has all
 205                  * been acknowledged in the course of the RTT that ended
 206                  * with the ACK we just received. Likewise, old_snd_cwnd
 207                  * is the cwnd during the previous RTT.
 208                  */
 209                 old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) /
 210                         tp->mss_cache;
 211                 old_snd_cwnd = vegas->beg_snd_cwnd;
 212
 213                 /* Save the extent of the current window so we can use this
 214                  * at the end of the next RTT.
 215                  */
 216                 vegas->beg_snd_una  = vegas->beg_snd_nxt;
 217                 vegas->beg_snd_nxt  = tp->snd_nxt;
 218                 vegas->beg_snd_cwnd = tp->snd_cwnd;
 219
 220                 /* We do the Vegas calculations only if we got enough RTT
 221                  * samples that we can be reasonably sure that we got
 222                  * at least one RTT sample that wasn't from a delayed ACK.
 223                  * If we only had 2 samples total,
 224                  * then that means we're getting only 1 ACK per RTT, which
 225                  * means they're almost certainly delayed ACKs.
 226                  * If  we have 3 samples, we should be OK.
 227                  */
 228
 229                 if (vegas->cntRTT <= 2) {
 230                         /* We don't have enough RTT samples to do the Vegas
 231                          * calculation, so we'll behave like Reno.
 232                          */
 233                         tcp_reno_cong_avoid(sk, ack, seq_rtt, in_flight, flag);
 234                 } else {
 235                         u32 rtt, target_cwnd, diff;
 236
 237                         /* We have enough RTT samples, so, using the Vegas
 238                          * algorithm, we determine if we should increase or
 239                          * decrease cwnd, and by how much.
 240                          */
 241
 242                         /* Pluck out the RTT we are using for the Vegas
 243                          * calculations. This is the min RTT seen during the
 244                          * last RTT. Taking the min filters out the effects
 245                          * of delayed ACKs, at the cost of noticing congestion
 246                          * a bit later.
 247                          */
 248                         rtt = vegas->minRTT;
 249
 250                         /* Calculate the cwnd we should have, if we weren't
 251                          * going too fast.
 252                          *
 253                          * This is:
 254                          *     (actual rate in segments) * baseRTT
 255                          * We keep it as a fixed point number with
 256                          * V_PARAM_SHIFT bits to the right of the binary point.
 257                          */
 258                         target_cwnd = ((old_wnd * vegas->baseRTT)
 259                                        << V_PARAM_SHIFT) / rtt;
 260
 261                         /* Calculate the difference between the window we had,
 262                          * and the window we would like to have. This quantity
 263                          * is the "Diff" from the Arizona Vegas papers.
 264                          *
 265                          * Again, this is a fixed point number with
 266                          * V_PARAM_SHIFT bits to the right of the binary
 267                          * point.
 268                          */
 269                         diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd;
 270
 271                         if (tp->snd_cwnd <= tp->snd_ssthresh) {
 272                                 /* Slow start.  */
 273                                 if (diff > gamma) {
 274                                         /* Going too fast. Time to slow down
 275                                          * and switch to congestion avoidance.
 276                                          */
 277                                         tp->snd_ssthresh = 2;
 278
 279                                         /* Set cwnd to match the actual rate
 280                                          * exactly:
 281                                          *   cwnd = (actual rate) * baseRTT
 282                                          * Then we add 1 because the integer
 283                                          * truncation robs us of full link
 284                                          * utilization.
 285                                          */
 286                                         tp->snd_cwnd = min(tp->snd_cwnd,
 287                                                            (target_cwnd >>
 288                                                             V_PARAM_SHIFT)+1);
 289
 290                                 }
 291                                 tcp_slow_start(tp);
 292                         } else {
 293                                 /* Congestion avoidance. */
 294                                 u32 next_snd_cwnd;
 295
 296                                 /* Figure out where we would like cwnd
 297                                  * to be.
 298                                  */
 299                                 if (diff > beta) {
 300                                         /* The old window was too fast, so
 301                                          * we slow down.
 302                                          */
 303                                         next_snd_cwnd = old_snd_cwnd - 1;
 304                                 } else if (diff < alpha) {
 305                                         /* We don't have enough extra packets
 306                                          * in the network, so speed up.
 307                                          */
 308                                         next_snd_cwnd = old_snd_cwnd + 1;
 309                                 } else {
 310                                         /* Sending just as fast as we
 311                                          * should be.
 312                                          */
 313                                         next_snd_cwnd = old_snd_cwnd;
 314                                 }
 315
 316                                 /* Adjust cwnd upward or downward, toward the
 317                                  * desired value.
 318                                  */
 319                                 if (next_snd_cwnd > tp->snd_cwnd)
 320                                         tp->snd_cwnd++;
 321                                 else if (next_snd_cwnd < tp->snd_cwnd)
 322                                         tp->snd_cwnd--;
 323                         }
 324
 325                         if (tp->snd_cwnd < 2)
 326                                 tp->snd_cwnd = 2;
 327                         else if (tp->snd_cwnd > tp->snd_cwnd_clamp)
 328                                 tp->snd_cwnd = tp->snd_cwnd_clamp;
 329                 }
 330
 331                 /* Wipe the slate clean for the next RTT. */
 332                 vegas->cntRTT = 0;
 333                 vegas->minRTT = 0x7fffffff;
 334         }
 335         /* Use normal slow start */
 336         else if (tp->snd_cwnd <= tp->snd_ssthresh)
 337                 tcp_slow_start(tp);
 338
 339 }
 340
 341 /* Extract info for Tcp socket info provided via netlink. */
 342 static void tcp_vegas_get_info(struct sock *sk, u32 ext,
 343                                struct sk_buff *skb)
 344 {
 345         const struct vegas *ca = inet_csk_ca(sk);
 346         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 347                 struct tcpvegas_info info = {
 348                         .tcpv_enabled = ca->doing_vegas_now,
 349                         .tcpv_rttcnt = ca->cntRTT,
 350                         .tcpv_rtt = ca->baseRTT,
 351                         .tcpv_minrtt = ca->minRTT,
 352                 };
 353
 354                 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 355         }
 356 }
 357
 358 static struct tcp_congestion_ops tcp_vegas = {
 359         .flags          = TCP_CONG_RTT_STAMP,
 360         .init           = tcp_vegas_init,
 361         .ssthresh       = tcp_reno_ssthresh,
 362         .cong_avoid     = tcp_vegas_cong_avoid,
 363         .min_cwnd       = tcp_reno_min_cwnd,
 364         .pkts_acked     = tcp_vegas_pkts_acked,
 365         .set_state      = tcp_vegas_state,
 366         .cwnd_event     = tcp_vegas_cwnd_event,
 367         .get_info       = tcp_vegas_get_info,
 368
 369         .owner          = THIS_MODULE,
 370         .name           = "vegas",
 371 };
 372
 373 static int __init tcp_vegas_register(void)
 374 {
 375         BUILD_BUG_ON(sizeof(struct vegas) > ICSK_CA_PRIV_SIZE);
 376         tcp_register_congestion_control(&tcp_vegas);
 377         return 0;
 378 }
 379
 380 static void __exit tcp_vegas_unregister(void)
 381 {
 382         tcp_unregister_congestion_control(&tcp_vegas);
 383 }
 384
 385 module_init(tcp_vegas_register);
 386 module_exit(tcp_vegas_unregister);
 387
 388 MODULE_AUTHOR("Stephen Hemminger");
 389 MODULE_LICENSE("GPL");
 390 MODULE_DESCRIPTION("TCP Vegas");