linux/net/ipv4/tcp_cubic.c
<<
>>
Prefs
   1/*
   2 * TCP CUBIC: Binary Increase Congestion control for TCP v2.3
   3 * Home page:
   4 *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
   5 * This is from the implementation of CUBIC TCP in
   6 * Sangtae Ha, Injong Rhee and Lisong Xu,
   7 *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant"
   8 *  in ACM SIGOPS Operating System Review, July 2008.
   9 * Available from:
  10 *  http://netsrv.csc.ncsu.edu/export/cubic_a_new_tcp_2008.pdf
  11 *
  12 * CUBIC integrates a new slow start algorithm, called HyStart.
  13 * The details of HyStart are presented in
  14 *  Sangtae Ha and Injong Rhee,
  15 *  "Taming the Elephants: New TCP Slow Start", NCSU TechReport 2008.
  16 * Available from:
  17 *  http://netsrv.csc.ncsu.edu/export/hystart_techreport_2008.pdf
  18 *
  19 * All testing results are available from:
  20 * http://netsrv.csc.ncsu.edu/wiki/index.php/TCP_Testing
  21 *
  22 * Unless CUBIC is enabled and congestion window is large
  23 * this behaves the same as the original Reno.
  24 */
  25
  26#include <linux/mm.h>
  27#include <linux/module.h>
  28#include <linux/math64.h>
  29#include <net/tcp.h>
  30
  31#define BICTCP_BETA_SCALE    1024       /* Scale factor beta calculation
  32                                         * max_cwnd = snd_cwnd * beta
  33                                         */
  34#define BICTCP_HZ               10      /* BIC HZ 2^10 = 1024 */
  35
  36/* Two methods of hybrid slow start */
  37#define HYSTART_ACK_TRAIN       0x1
  38#define HYSTART_DELAY           0x2
  39
  40/* Number of delay samples for detecting the increase of delay */
  41#define HYSTART_MIN_SAMPLES     8
  42#define HYSTART_DELAY_MIN       (4U<<3)
  43#define HYSTART_DELAY_MAX       (16U<<3)
  44#define HYSTART_DELAY_THRESH(x) clamp(x, HYSTART_DELAY_MIN, HYSTART_DELAY_MAX)
  45
  46static int fast_convergence __read_mostly = 1;
  47static int beta __read_mostly = 717;    /* = 717/1024 (BICTCP_BETA_SCALE) */
  48static int initial_ssthresh __read_mostly;
  49static int bic_scale __read_mostly = 41;
  50static int tcp_friendliness __read_mostly = 1;
  51
  52static int hystart __read_mostly = 1;
  53static int hystart_detect __read_mostly = HYSTART_ACK_TRAIN | HYSTART_DELAY;
  54static int hystart_low_window __read_mostly = 16;
  55static int hystart_ack_delta __read_mostly = 2;
  56
  57static u32 cube_rtt_scale __read_mostly;
  58static u32 beta_scale __read_mostly;
  59static u64 cube_factor __read_mostly;
  60
  61/* Note parameters that are used for precomputing scale factors are read-only */
  62module_param(fast_convergence, int, 0644);
  63MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  64module_param(beta, int, 0644);
  65MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  66module_param(initial_ssthresh, int, 0644);
  67MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  68module_param(bic_scale, int, 0444);
  69MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
  70module_param(tcp_friendliness, int, 0644);
  71MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
  72module_param(hystart, int, 0644);
  73MODULE_PARM_DESC(hystart, "turn on/off hybrid slow start algorithm");
  74module_param(hystart_detect, int, 0644);
  75MODULE_PARM_DESC(hystart_detect, "hyrbrid slow start detection mechanisms"
  76                 " 1: packet-train 2: delay 3: both packet-train and delay");
  77module_param(hystart_low_window, int, 0644);
  78MODULE_PARM_DESC(hystart_low_window, "lower bound cwnd for hybrid slow start");
  79module_param(hystart_ack_delta, int, 0644);
  80MODULE_PARM_DESC(hystart_ack_delta, "spacing between ack's indicating train (msecs)");
  81
  82/* BIC TCP Parameters */
  83struct bictcp {
  84        u32     cnt;            /* increase cwnd by 1 after ACKs */
  85        u32     last_max_cwnd;  /* last maximum snd_cwnd */
  86        u32     loss_cwnd;      /* congestion window at last loss */
  87        u32     last_cwnd;      /* the last snd_cwnd */
  88        u32     last_time;      /* time when updated last_cwnd */
  89        u32     bic_origin_point;/* origin point of bic function */
  90        u32     bic_K;          /* time to origin point from the beginning of the current epoch */
  91        u32     delay_min;      /* min delay (msec << 3) */
  92        u32     epoch_start;    /* beginning of an epoch */
  93        u32     ack_cnt;        /* number of acks */
  94        u32     tcp_cwnd;       /* estimated tcp cwnd */
  95#define ACK_RATIO_SHIFT 4
  96#define ACK_RATIO_LIMIT (32u << ACK_RATIO_SHIFT)
  97        u16     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
  98        u8      sample_cnt;     /* number of samples to decide curr_rtt */
  99        u8      found;          /* the exit point is found? */
 100        u32     round_start;    /* beginning of each round */
 101        u32     end_seq;        /* end_seq of the round */
 102        u32     last_ack;       /* last time when the ACK spacing is close */
 103        u32     curr_rtt;       /* the minimum rtt of current round */
 104};
 105
 106static inline void bictcp_reset(struct bictcp *ca)
 107{
 108        ca->cnt = 0;
 109        ca->last_max_cwnd = 0;
 110        ca->last_cwnd = 0;
 111        ca->last_time = 0;
 112        ca->bic_origin_point = 0;
 113        ca->bic_K = 0;
 114        ca->delay_min = 0;
 115        ca->epoch_start = 0;
 116        ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
 117        ca->ack_cnt = 0;
 118        ca->tcp_cwnd = 0;
 119        ca->found = 0;
 120}
 121
 122static inline u32 bictcp_clock(void)
 123{
 124#if HZ < 1000
 125        return ktime_to_ms(ktime_get_real());
 126#else
 127        return jiffies_to_msecs(jiffies);
 128#endif
 129}
 130
 131static inline void bictcp_hystart_reset(struct sock *sk)
 132{
 133        struct tcp_sock *tp = tcp_sk(sk);
 134        struct bictcp *ca = inet_csk_ca(sk);
 135
 136        ca->round_start = ca->last_ack = bictcp_clock();
 137        ca->end_seq = tp->snd_nxt;
 138        ca->curr_rtt = 0;
 139        ca->sample_cnt = 0;
 140}
 141
 142static void bictcp_init(struct sock *sk)
 143{
 144        struct bictcp *ca = inet_csk_ca(sk);
 145
 146        bictcp_reset(ca);
 147        ca->loss_cwnd = 0;
 148
 149        if (hystart)
 150                bictcp_hystart_reset(sk);
 151
 152        if (!hystart && initial_ssthresh)
 153                tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
 154}
 155
 156/* calculate the cubic root of x using a table lookup followed by one
 157 * Newton-Raphson iteration.
 158 * Avg err ~= 0.195%
 159 */
 160static u32 cubic_root(u64 a)
 161{
 162        u32 x, b, shift;
 163        /*
 164         * cbrt(x) MSB values for x MSB values in [0..63].
 165         * Precomputed then refined by hand - Willy Tarreau
 166         *
 167         * For x in [0..63],
 168         *   v = cbrt(x << 18) - 1
 169         *   cbrt(x) = (v[x] + 10) >> 6
 170         */
 171        static const u8 v[] = {
 172                /* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
 173                /* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
 174                /* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
 175                /* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
 176                /* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
 177                /* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
 178                /* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
 179                /* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
 180        };
 181
 182        b = fls64(a);
 183        if (b < 7) {
 184                /* a in [0..63] */
 185                return ((u32)v[(u32)a] + 35) >> 6;
 186        }
 187
 188        b = ((b * 84) >> 8) - 1;
 189        shift = (a >> (b * 3));
 190
 191        x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
 192
 193        /*
 194         * Newton-Raphson iteration
 195         *                         2
 196         * x    = ( 2 * x  +  a / x  ) / 3
 197         *  k+1          k         k
 198         */
 199        x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
 200        x = ((x * 341) >> 10);
 201        return x;
 202}
 203
 204/*
 205 * Compute congestion window to use.
 206 */
 207static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 208{
 209        u64 offs;
 210        u32 delta, t, bic_target, max_cnt;
 211
 212        ca->ack_cnt++;  /* count the number of ACKs */
 213
 214        if (ca->last_cwnd == cwnd &&
 215            (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
 216                return;
 217
 218        ca->last_cwnd = cwnd;
 219        ca->last_time = tcp_time_stamp;
 220
 221        if (ca->epoch_start == 0) {
 222                ca->epoch_start = tcp_time_stamp;       /* record the beginning of an epoch */
 223                ca->ack_cnt = 1;                        /* start counting */
 224                ca->tcp_cwnd = cwnd;                    /* syn with cubic */
 225
 226                if (ca->last_max_cwnd <= cwnd) {
 227                        ca->bic_K = 0;
 228                        ca->bic_origin_point = cwnd;
 229                } else {
 230                        /* Compute new K based on
 231                         * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
 232                         */
 233                        ca->bic_K = cubic_root(cube_factor
 234                                               * (ca->last_max_cwnd - cwnd));
 235                        ca->bic_origin_point = ca->last_max_cwnd;
 236                }
 237        }
 238
 239        /* cubic function - calc*/
 240        /* calculate c * time^3 / rtt,
 241         *  while considering overflow in calculation of time^3
 242         * (so time^3 is done by using 64 bit)
 243         * and without the support of division of 64bit numbers
 244         * (so all divisions are done by using 32 bit)
 245         *  also NOTE the unit of those veriables
 246         *        time  = (t - K) / 2^bictcp_HZ
 247         *        c = bic_scale >> 10
 248         * rtt  = (srtt >> 3) / HZ
 249         * !!! The following code does not have overflow problems,
 250         * if the cwnd < 1 million packets !!!
 251         */
 252
 253        /* change the unit from HZ to bictcp_HZ */
 254        t = ((tcp_time_stamp + msecs_to_jiffies(ca->delay_min>>3)
 255              - ca->epoch_start) << BICTCP_HZ) / HZ;
 256
 257        if (t < ca->bic_K)              /* t - K */
 258                offs = ca->bic_K - t;
 259        else
 260                offs = t - ca->bic_K;
 261
 262        /* c/rtt * (t-K)^3 */
 263        delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
 264        if (t < ca->bic_K)                                      /* below origin*/
 265                bic_target = ca->bic_origin_point - delta;
 266        else                                                    /* above origin*/
 267                bic_target = ca->bic_origin_point + delta;
 268
 269        /* cubic function - calc bictcp_cnt*/
 270        if (bic_target > cwnd) {
 271                ca->cnt = cwnd / (bic_target - cwnd);
 272        } else {
 273                ca->cnt = 100 * cwnd;              /* very small increment*/
 274        }
 275
 276        /*
 277         * The initial growth of cubic function may be too conservative
 278         * when the available bandwidth is still unknown.
 279         */
 280        if (ca->last_max_cwnd == 0 && ca->cnt > 20)
 281                ca->cnt = 20;   /* increase cwnd 5% per RTT */
 282
 283        /* TCP Friendly */
 284        if (tcp_friendliness) {
 285                u32 scale = beta_scale;
 286                delta = (cwnd * scale) >> 3;
 287                while (ca->ack_cnt > delta) {           /* update tcp cwnd */
 288                        ca->ack_cnt -= delta;
 289                        ca->tcp_cwnd++;
 290                }
 291
 292                if (ca->tcp_cwnd > cwnd){       /* if bic is slower than tcp */
 293                        delta = ca->tcp_cwnd - cwnd;
 294                        max_cnt = cwnd / delta;
 295                        if (ca->cnt > max_cnt)
 296                                ca->cnt = max_cnt;
 297                }
 298        }
 299
 300        ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
 301        if (ca->cnt == 0)                       /* cannot be zero */
 302                ca->cnt = 1;
 303}
 304
 305static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 306{
 307        struct tcp_sock *tp = tcp_sk(sk);
 308        struct bictcp *ca = inet_csk_ca(sk);
 309
 310        if (!tcp_is_cwnd_limited(sk, in_flight))
 311                return;
 312
 313        if (tp->snd_cwnd <= tp->snd_ssthresh) {
 314                if (hystart && after(ack, ca->end_seq))
 315                        bictcp_hystart_reset(sk);
 316                tcp_slow_start(tp);
 317        } else {
 318                bictcp_update(ca, tp->snd_cwnd);
 319                tcp_cong_avoid_ai(tp, ca->cnt);
 320        }
 321
 322}
 323
 324static u32 bictcp_recalc_ssthresh(struct sock *sk)
 325{
 326        const struct tcp_sock *tp = tcp_sk(sk);
 327        struct bictcp *ca = inet_csk_ca(sk);
 328
 329        ca->epoch_start = 0;    /* end of epoch */
 330
 331        /* Wmax and fast convergence */
 332        if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
 333                ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
 334                        / (2 * BICTCP_BETA_SCALE);
 335        else
 336                ca->last_max_cwnd = tp->snd_cwnd;
 337
 338        ca->loss_cwnd = tp->snd_cwnd;
 339
 340        return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 341}
 342
 343static u32 bictcp_undo_cwnd(struct sock *sk)
 344{
 345        struct bictcp *ca = inet_csk_ca(sk);
 346
 347        return max(tcp_sk(sk)->snd_cwnd, ca->loss_cwnd);
 348}
 349
 350static void bictcp_state(struct sock *sk, u8 new_state)
 351{
 352        if (new_state == TCP_CA_Loss) {
 353                bictcp_reset(inet_csk_ca(sk));
 354                bictcp_hystart_reset(sk);
 355        }
 356}
 357
 358static void hystart_update(struct sock *sk, u32 delay)
 359{
 360        struct tcp_sock *tp = tcp_sk(sk);
 361        struct bictcp *ca = inet_csk_ca(sk);
 362
 363        if (!(ca->found & hystart_detect)) {
 364                u32 now = bictcp_clock();
 365
 366                /* first detection parameter - ack-train detection */
 367                if ((s32)(now - ca->last_ack) <= hystart_ack_delta) {
 368                        ca->last_ack = now;
 369                        if ((s32)(now - ca->round_start) > ca->delay_min >> 4)
 370                                ca->found |= HYSTART_ACK_TRAIN;
 371                }
 372
 373                /* obtain the minimum delay of more than sampling packets */
 374                if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
 375                        if (ca->curr_rtt == 0 || ca->curr_rtt > delay)
 376                                ca->curr_rtt = delay;
 377
 378                        ca->sample_cnt++;
 379                } else {
 380                        if (ca->curr_rtt > ca->delay_min +
 381                            HYSTART_DELAY_THRESH(ca->delay_min>>4))
 382                                ca->found |= HYSTART_DELAY;
 383                }
 384                /*
 385                 * Either one of two conditions are met,
 386                 * we exit from slow start immediately.
 387                 */
 388                if (ca->found & hystart_detect)
 389                        tp->snd_ssthresh = tp->snd_cwnd;
 390        }
 391}
 392
 393/* Track delayed acknowledgment ratio using sliding window
 394 * ratio = (15*ratio + sample) / 16
 395 */
 396static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
 397{
 398        const struct inet_connection_sock *icsk = inet_csk(sk);
 399        const struct tcp_sock *tp = tcp_sk(sk);
 400        struct bictcp *ca = inet_csk_ca(sk);
 401        u32 delay;
 402
 403        if (icsk->icsk_ca_state == TCP_CA_Open) {
 404                u32 ratio = ca->delayed_ack;
 405
 406                ratio -= ca->delayed_ack >> ACK_RATIO_SHIFT;
 407                ratio += cnt;
 408
 409                ca->delayed_ack = min(ratio, ACK_RATIO_LIMIT);
 410        }
 411
 412        /* Some calls are for duplicates without timetamps */
 413        if (rtt_us < 0)
 414                return;
 415
 416        /* Discard delay samples right after fast recovery */
 417        if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
 418                return;
 419
 420        delay = (rtt_us << 3) / USEC_PER_MSEC;
 421        if (delay == 0)
 422                delay = 1;
 423
 424        /* first time call or link delay decreases */
 425        if (ca->delay_min == 0 || ca->delay_min > delay)
 426                ca->delay_min = delay;
 427
 428        /* hystart triggers when cwnd is larger than some threshold */
 429        if (hystart && tp->snd_cwnd <= tp->snd_ssthresh &&
 430            tp->snd_cwnd >= hystart_low_window)
 431                hystart_update(sk, delay);
 432}
 433
 434static struct tcp_congestion_ops cubictcp __read_mostly = {
 435        .init           = bictcp_init,
 436        .ssthresh       = bictcp_recalc_ssthresh,
 437        .cong_avoid     = bictcp_cong_avoid,
 438        .set_state      = bictcp_state,
 439        .undo_cwnd      = bictcp_undo_cwnd,
 440        .pkts_acked     = bictcp_acked,
 441        .owner          = THIS_MODULE,
 442        .name           = "cubic",
 443};
 444
 445static int __init cubictcp_register(void)
 446{
 447        BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
 448
 449        /* Precompute a bunch of the scaling factors that are used per-packet
 450         * based on SRTT of 100ms
 451         */
 452
 453        beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
 454
 455        cube_rtt_scale = (bic_scale * 10);      /* 1024*c/rtt */
 456
 457        /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
 458         *  so K = cubic_root( (wmax-cwnd)*rtt/c )
 459         * the unit of K is bictcp_HZ=2^10, not HZ
 460         *
 461         *  c = bic_scale >> 10
 462         *  rtt = 100ms
 463         *
 464         * the following code has been designed and tested for
 465         * cwnd < 1 million packets
 466         * RTT < 100 seconds
 467         * HZ < 1,000,00  (corresponding to 10 nano-second)
 468         */
 469
 470        /* 1/c * 2^2*bictcp_HZ * srtt */
 471        cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
 472
 473        /* divide by bic_scale and by constant Srtt (100ms) */
 474        do_div(cube_factor, bic_scale * 10);
 475
 476        /* hystart needs ms clock resolution */
 477        if (hystart && HZ < 1000)
 478                cubictcp.flags |= TCP_CONG_RTT_STAMP;
 479
 480        return tcp_register_congestion_control(&cubictcp);
 481}
 482
 483static void __exit cubictcp_unregister(void)
 484{
 485        tcp_unregister_congestion_control(&cubictcp);
 486}
 487
 488module_init(cubictcp_register);
 489module_exit(cubictcp_unregister);
 490
 491MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
 492MODULE_LICENSE("GPL");
 493MODULE_DESCRIPTION("CUBIC TCP");
 494MODULE_VERSION("2.3");
 495
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.