net/ipv4/tcp_illinois.c

   1 /*
   2  * TCP Illinois congestion control.
   3  * Home page:
   4  *      http://www.ews.uiuc.edu/~shaoliu/tcpillinois/index.html
   5  *
   6  * The algorithm is described in:
   7  * "TCP-Illinois: A Loss and Delay-Based Congestion Control Algorithm
   8  *  for High-Speed Networks"
   9  * http://www.ews.uiuc.edu/~shaoliu/papersandslides/liubassri06perf.pdf
  10  *
  11  * Implemented from description in paper and ns-2 simulation.
  12  * Copyright (C) 2007 Stephen Hemminger <shemminger@linux-foundation.org>
  13  */
  14
  15 #include <linux/module.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/inet_diag.h>
  18 #include <asm/div64.h>
  19 #include <net/tcp.h>
  20
  21 #define ALPHA_SHIFT     7
  22 #define ALPHA_SCALE     (1u<<ALPHA_SHIFT)
  23 #define ALPHA_MIN       ((3*ALPHA_SCALE)/10)    /* ~0.3 */
  24 #define ALPHA_MAX       (10*ALPHA_SCALE)        /* 10.0 */
  25 #define ALPHA_BASE      ALPHA_SCALE             /* 1.0 */
  26 #define U32_MAX         ((u32)~0U)
  27 #define RTT_MAX         (U32_MAX / ALPHA_MAX)   /* 3.3 secs */
  28
  29 #define BETA_SHIFT      6
  30 #define BETA_SCALE      (1u<<BETA_SHIFT)
  31 #define BETA_MIN        (BETA_SCALE/8)          /* 0.125 */
  32 #define BETA_MAX        (BETA_SCALE/2)          /* 0.5 */
  33 #define BETA_BASE       BETA_MAX
  34
  35 static int win_thresh __read_mostly = 15;
  36 module_param(win_thresh, int, 0);
  37 MODULE_PARM_DESC(win_thresh, "Window threshold for starting adaptive sizing");
  38
  39 static int theta __read_mostly = 5;
  40 module_param(theta, int, 0);
  41 MODULE_PARM_DESC(theta, "# of fast RTT's before full growth");
  42
  43 /* TCP Illinois Parameters */
  44 struct illinois {
  45         u64     sum_rtt;        /* sum of rtt's measured within last rtt */
  46         u16     cnt_rtt;        /* # of rtts measured within last rtt */
  47         u32     base_rtt;       /* min of all rtt in usec */
  48         u32     max_rtt;        /* max of all rtt in usec */
  49         u32     end_seq;        /* right edge of current RTT */
  50         u32     alpha;          /* Additive increase */
  51         u32     beta;           /* Muliplicative decrease */
  52         u16     acked;          /* # packets acked by current ACK */
  53         u8      rtt_above;      /* average rtt has gone above threshold */
  54         u8      rtt_low;        /* # of rtts measurements below threshold */
  55 };
  56
  57 static void rtt_reset(struct sock *sk)
  58 {
  59         struct tcp_sock *tp = tcp_sk(sk);
  60         struct illinois *ca = inet_csk_ca(sk);
  61
  62         ca->end_seq = tp->snd_nxt;
  63         ca->cnt_rtt = 0;
  64         ca->sum_rtt = 0;
  65
  66         /* TODO: age max_rtt? */
  67 }
  68
  69 static void tcp_illinois_init(struct sock *sk)
  70 {
  71         struct illinois *ca = inet_csk_ca(sk);
  72
  73         ca->alpha = ALPHA_MAX;
  74         ca->beta = BETA_BASE;
  75         ca->base_rtt = 0x7fffffff;
  76         ca->max_rtt = 0;
  77
  78         ca->acked = 0;
  79         ca->rtt_low = 0;
  80         ca->rtt_above = 0;
  81
  82         rtt_reset(sk);
  83 }
  84
  85 /* Measure RTT for each ack. */
  86 static void tcp_illinois_acked(struct sock *sk, u32 pkts_acked, ktime_t last)
  87 {
  88         struct illinois *ca = inet_csk_ca(sk);
  89         u32 rtt;
  90
  91         ca->acked = pkts_acked;
  92
  93         if (ktime_equal(last, net_invalid_timestamp()))
  94                 return;
  95
  96         rtt = ktime_to_us(net_timedelta(last));
  97
  98         /* ignore bogus values, this prevents wraparound in alpha math */
  99         if (rtt > RTT_MAX)
 100                 rtt = RTT_MAX;
 101
 102         /* keep track of minimum RTT seen so far */
 103         if (ca->base_rtt > rtt)
 104                 ca->base_rtt = rtt;
 105
 106         /* and max */
 107         if (ca->max_rtt < rtt)
 108                 ca->max_rtt = rtt;
 109
 110         ++ca->cnt_rtt;
 111         ca->sum_rtt += rtt;
 112 }
 113
 114 /* Maximum queuing delay */
 115 static inline u32 max_delay(const struct illinois *ca)
 116 {
 117         return ca->max_rtt - ca->base_rtt;
 118 }
 119
 120 /* Average queuing delay */
 121 static inline u32 avg_delay(const struct illinois *ca)
 122 {
 123         u64 t = ca->sum_rtt;
 124
 125         do_div(t, ca->cnt_rtt);
 126         return t - ca->base_rtt;
 127 }
 128
 129 /*
 130  * Compute value of alpha used for additive increase.
 131  * If small window then use 1.0, equivalent to Reno.
 132  *
 133  * For larger windows, adjust based on average delay.
 134  * A. If average delay is at minimum (we are uncongested),
 135  *    then use large alpha (10.0) to increase faster.
 136  * B. If average delay is at maximum (getting congested)
 137  *    then use small alpha (0.3)
 138  *
 139  * The result is a convex window growth curve.
 140  */
 141 static u32 alpha(struct illinois *ca, u32 da, u32 dm)
 142 {
 143         u32 d1 = dm / 100;      /* Low threshold */
 144
 145         if (da <= d1) {
 146                 /* If never got out of low delay zone, then use max */
 147                 if (!ca->rtt_above)
 148                         return ALPHA_MAX;
 149
 150                 /* Wait for 5 good RTT's before allowing alpha to go alpha max.
 151                  * This prevents one good RTT from causing sudden window increase.
 152                  */
 153                 if (++ca->rtt_low < theta)
 154                         return ca->alpha;
 155
 156                 ca->rtt_low = 0;
 157                 ca->rtt_above = 0;
 158                 return ALPHA_MAX;
 159         }
 160
 161         ca->rtt_above = 1;
 162
 163         /*
 164          * Based on:
 165          *
 166          *      (dm - d1) amin amax
 167          * k1 = -------------------
 168          *         amax - amin
 169          *
 170          *       (dm - d1) amin
 171          * k2 = ----------------  - d1
 172          *        amax - amin
 173          *
 174          *             k1
 175          * alpha = ----------
 176          *          k2 + da
 177          */
 178
 179         dm -= d1;
 180         da -= d1;
 181         return (dm * ALPHA_MAX) /
 182                 (dm + (da  * (ALPHA_MAX - ALPHA_MIN)) / ALPHA_MIN);
 183 }
 184
 185 /*
 186  * Beta used for multiplicative decrease.
 187  * For small window sizes returns same value as Reno (0.5)
 188  *
 189  * If delay is small (10% of max) then beta = 1/8
 190  * If delay is up to 80% of max then beta = 1/2
 191  * In between is a linear function
 192  */
 193 static u32 beta(u32 da, u32 dm)
 194 {
 195         u32 d2, d3;
 196
 197         d2 = dm / 10;
 198         if (da <= d2)
 199                 return BETA_MIN;
 200
 201         d3 = (8 * dm) / 10;
 202         if (da >= d3 || d3 <= d2)
 203                 return BETA_MAX;
 204
 205         /*
 206          * Based on:
 207          *
 208          *       bmin d3 - bmax d2
 209          * k3 = -------------------
 210          *         d3 - d2
 211          *
 212          *       bmax - bmin
 213          * k4 = -------------
 214          *         d3 - d2
 215          *
 216          * b = k3 + k4 da
 217          */
 218         return (BETA_MIN * d3 - BETA_MAX * d2 + (BETA_MAX - BETA_MIN) * da)
 219                 / (d3 - d2);
 220 }
 221
 222 /* Update alpha and beta values once per RTT */
 223 static void update_params(struct sock *sk)
 224 {
 225         struct tcp_sock *tp = tcp_sk(sk);
 226         struct illinois *ca = inet_csk_ca(sk);
 227
 228         if (tp->snd_cwnd < win_thresh) {
 229                 ca->alpha = ALPHA_BASE;
 230                 ca->beta = BETA_BASE;
 231         } else if (ca->cnt_rtt > 0) {
 232                 u32 dm = max_delay(ca);
 233                 u32 da = avg_delay(ca);
 234
 235                 ca->alpha = alpha(ca, da, dm);
 236                 ca->beta = beta(da, dm);
 237         }
 238
 239         rtt_reset(sk);
 240 }
 241
 242 /*
 243  * In case of loss, reset to default values
 244  */
 245 static void tcp_illinois_state(struct sock *sk, u8 new_state)
 246 {
 247         struct illinois *ca = inet_csk_ca(sk);
 248
 249         if (new_state == TCP_CA_Loss) {
 250                 ca->alpha = ALPHA_BASE;
 251                 ca->beta = BETA_BASE;
 252                 ca->rtt_low = 0;
 253                 ca->rtt_above = 0;
 254                 rtt_reset(sk);
 255         }
 256 }
 257
 258 /*
 259  * Increase window in response to successful acknowledgment.
 260  */
 261 static void tcp_illinois_cong_avoid(struct sock *sk, u32 ack, u32 rtt,
 262                                     u32 in_flight, int flag)
 263 {
 264         struct tcp_sock *tp = tcp_sk(sk);
 265         struct illinois *ca = inet_csk_ca(sk);
 266
 267         if (after(ack, ca->end_seq))
 268                 update_params(sk);
 269
 270         /* RFC2861 only increase cwnd if fully utilized */
 271         if (!tcp_is_cwnd_limited(sk, in_flight))
 272                 return;
 273
 274         /* In slow start */
 275         if (tp->snd_cwnd <= tp->snd_ssthresh)
 276                 tcp_slow_start(tp);
 277
 278         else {
 279                 u32 delta;
 280
 281                 /* snd_cwnd_cnt is # of packets since last cwnd increment */
 282                 tp->snd_cwnd_cnt += ca->acked;
 283                 ca->acked = 1;
 284
 285                 /* This is close approximation of:
 286                  * tp->snd_cwnd += alpha/tp->snd_cwnd
 287                 */
 288                 delta = (tp->snd_cwnd_cnt * ca->alpha) >> ALPHA_SHIFT;
 289                 if (delta >= tp->snd_cwnd) {
 290                         tp->snd_cwnd = min(tp->snd_cwnd + delta / tp->snd_cwnd,
 291                                            (u32) tp->snd_cwnd_clamp);
 292                         tp->snd_cwnd_cnt = 0;
 293                 }
 294         }
 295 }
 296
 297 static u32 tcp_illinois_ssthresh(struct sock *sk)
 298 {
 299         struct tcp_sock *tp = tcp_sk(sk);
 300         struct illinois *ca = inet_csk_ca(sk);
 301
 302         /* Multiplicative decrease */
 303         return max((tp->snd_cwnd * ca->beta) >> BETA_SHIFT, 2U);
 304 }
 305
 306
 307 /* Extract info for Tcp socket info provided via netlink. */
 308 static void tcp_illinois_info(struct sock *sk, u32 ext,
 309                               struct sk_buff *skb)
 310 {
 311         const struct illinois *ca = inet_csk_ca(sk);
 312
 313         if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
 314                 struct tcpvegas_info info = {
 315                         .tcpv_enabled = 1,
 316                         .tcpv_rttcnt = ca->cnt_rtt,
 317                         .tcpv_minrtt = ca->base_rtt,
 318                 };
 319                 u64 t = ca->sum_rtt;
 320
 321                 do_div(t, ca->cnt_rtt);
 322                 info.tcpv_rtt = t;
 323
 324                 nla_put(skb, INET_DIAG_VEGASINFO, sizeof(info), &info);
 325         }
 326 }
 327
 328 static struct tcp_congestion_ops tcp_illinois = {
 329         .flags          = TCP_CONG_RTT_STAMP,
 330         .init           = tcp_illinois_init,
 331         .ssthresh       = tcp_illinois_ssthresh,
 332         .min_cwnd       = tcp_reno_min_cwnd,
 333         .cong_avoid     = tcp_illinois_cong_avoid,
 334         .set_state      = tcp_illinois_state,
 335         .get_info       = tcp_illinois_info,
 336         .pkts_acked     = tcp_illinois_acked,
 337
 338         .owner          = THIS_MODULE,
 339         .name           = "illinois",
 340 };
 341
 342 static int __init tcp_illinois_register(void)
 343 {
 344         BUILD_BUG_ON(sizeof(struct illinois) > ICSK_CA_PRIV_SIZE);
 345         return tcp_register_congestion_control(&tcp_illinois);
 346 }
 347
 348 static void __exit tcp_illinois_unregister(void)
 349 {
 350         tcp_unregister_congestion_control(&tcp_illinois);
 351 }
 352
 353 module_init(tcp_illinois_register);
 354 module_exit(tcp_illinois_unregister);
 355
 356 MODULE_AUTHOR("Stephen Hemminger, Shao Liu");
 357 MODULE_LICENSE("GPL");
 358 MODULE_DESCRIPTION("TCP Illinois");
 359 MODULE_VERSION("1.0");