net/ipv4/tcp_cubic.c

   1 /*
   2  * TCP CUBIC: Binary Increase Congestion control for TCP v2.2
   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  * Injong Rhee, Lisong Xu.
   7  *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant
   8  *  in PFLDnet 2005
   9  * Available from:
  10  *  http://netsrv.csc.ncsu.edu/export/cubic-paper.pdf
  11  *
  12  * Unless CUBIC is enabled and congestion window is large
  13  * this behaves the same as the original Reno.
  14  */
  15
  16 #include <linux/mm.h>
  17 #include <linux/module.h>
  18 #include <linux/math64.h>
  19 #include <net/tcp.h>
  20
  21 #define BICTCP_BETA_SCALE    1024       /* Scale factor beta calculation
  22                                          * max_cwnd = snd_cwnd * beta
  23                                          */
  24 #define BICTCP_HZ               10      /* BIC HZ 2^10 = 1024 */
  25
  26 static int fast_convergence __read_mostly = 1;
  27 static int beta __read_mostly = 717;    /* = 717/1024 (BICTCP_BETA_SCALE) */
  28 static int initial_ssthresh __read_mostly;
  29 static int bic_scale __read_mostly = 41;
  30 static int tcp_friendliness __read_mostly = 1;
  31
  32 static u32 cube_rtt_scale __read_mostly;
  33 static u32 beta_scale __read_mostly;
  34 static u64 cube_factor __read_mostly;
  35
  36 /* Note parameters that are used for precomputing scale factors are read-only */
  37 module_param(fast_convergence, int, 0644);
  38 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  39 module_param(beta, int, 0644);
  40 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  41 module_param(initial_ssthresh, int, 0644);
  42 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  43 module_param(bic_scale, int, 0444);
  44 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
  45 module_param(tcp_friendliness, int, 0644);
  46 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
  47
  48 /* BIC TCP Parameters */
  49 struct bictcp {
  50         u32     cnt;            /* increase cwnd by 1 after ACKs */
  51         u32     last_max_cwnd;  /* last maximum snd_cwnd */
  52         u32     loss_cwnd;      /* congestion window at last loss */
  53         u32     last_cwnd;      /* the last snd_cwnd */
  54         u32     last_time;      /* time when updated last_cwnd */
  55         u32     bic_origin_point;/* origin point of bic function */
  56         u32     bic_K;          /* time to origin point from the beginning of the current epoch */
  57         u32     delay_min;      /* min delay */
  58         u32     epoch_start;    /* beginning of an epoch */
  59         u32     ack_cnt;        /* number of acks */
  60         u32     tcp_cwnd;       /* estimated tcp cwnd */
  61 #define ACK_RATIO_SHIFT 4
  62         u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
  63 };
  64
  65 static inline void bictcp_reset(struct bictcp *ca)
  66 {
  67         ca->cnt = 0;
  68         ca->last_max_cwnd = 0;
  69         ca->loss_cwnd = 0;
  70         ca->last_cwnd = 0;
  71         ca->last_time = 0;
  72         ca->bic_origin_point = 0;
  73         ca->bic_K = 0;
  74         ca->delay_min = 0;
  75         ca->epoch_start = 0;
  76         ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
  77         ca->ack_cnt = 0;
  78         ca->tcp_cwnd = 0;
  79 }
  80
  81 static void bictcp_init(struct sock *sk)
  82 {
  83         bictcp_reset(inet_csk_ca(sk));
  84         if (initial_ssthresh)
  85                 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
  86 }
  87
  88 /* calculate the cubic root of x using a table lookup followed by one
  89  * Newton-Raphson iteration.
  90  * Avg err ~= 0.195%
  91  */
  92 static u32 cubic_root(u64 a)
  93 {
  94         u32 x, b, shift;
  95         /*
  96          * cbrt(x) MSB values for x MSB values in [0..63].
  97          * Precomputed then refined by hand - Willy Tarreau
  98          *
  99          * For x in [0..63],
 100          *   v = cbrt(x << 18) - 1
 101          *   cbrt(x) = (v[x] + 10) >> 6
 102          */
 103         static const u8 v[] = {
 104                 /* 0x00 */    0,   54,   54,   54,  118,  118,  118,  118,
 105                 /* 0x08 */  123,  129,  134,  138,  143,  147,  151,  156,
 106                 /* 0x10 */  157,  161,  164,  168,  170,  173,  176,  179,
 107                 /* 0x18 */  181,  185,  187,  190,  192,  194,  197,  199,
 108                 /* 0x20 */  200,  202,  204,  206,  209,  211,  213,  215,
 109                 /* 0x28 */  217,  219,  221,  222,  224,  225,  227,  229,
 110                 /* 0x30 */  231,  232,  234,  236,  237,  239,  240,  242,
 111                 /* 0x38 */  244,  245,  246,  248,  250,  251,  252,  254,
 112         };
 113
 114         b = fls64(a);
 115         if (b < 7) {
 116                 /* a in [0..63] */
 117                 return ((u32)v[(u32)a] + 35) >> 6;
 118         }
 119
 120         b = ((b * 84) >> 8) - 1;
 121         shift = (a >> (b * 3));
 122
 123         x = ((u32)(((u32)v[shift] + 10) << b)) >> 6;
 124
 125         /*
 126          * Newton-Raphson iteration
 127          *                         2
 128          * x    = ( 2 * x  +  a / x  ) / 3
 129          *  k+1          k         k
 130          */
 131         x = (2 * x + (u32)div64_u64(a, (u64)x * (u64)(x - 1)));
 132         x = ((x * 341) >> 10);
 133         return x;
 134 }
 135
 136 /*
 137  * Compute congestion window to use.
 138  */
 139 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 140 {
 141         u64 offs;
 142         u32 delta, t, bic_target, max_cnt;
 143
 144         ca->ack_cnt++;  /* count the number of ACKs */
 145
 146         if (ca->last_cwnd == cwnd &&
 147             (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
 148                 return;
 149
 150         ca->last_cwnd = cwnd;
 151         ca->last_time = tcp_time_stamp;
 152
 153         if (ca->epoch_start == 0) {
 154                 ca->epoch_start = tcp_time_stamp;       /* record the beginning of an epoch */
 155                 ca->ack_cnt = 1;                        /* start counting */
 156                 ca->tcp_cwnd = cwnd;                    /* syn with cubic */
 157
 158                 if (ca->last_max_cwnd <= cwnd) {
 159                         ca->bic_K = 0;
 160                         ca->bic_origin_point = cwnd;
 161                 } else {
 162                         /* Compute new K based on
 163                          * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
 164                          */
 165                         ca->bic_K = cubic_root(cube_factor
 166                                                * (ca->last_max_cwnd - cwnd));
 167                         ca->bic_origin_point = ca->last_max_cwnd;
 168                 }
 169         }
 170
 171         /* cubic function - calc*/
 172         /* calculate c * time^3 / rtt,
 173          *  while considering overflow in calculation of time^3
 174          * (so time^3 is done by using 64 bit)
 175          * and without the support of division of 64bit numbers
 176          * (so all divisions are done by using 32 bit)
 177          *  also NOTE the unit of those veriables
 178          *        time  = (t - K) / 2^bictcp_HZ
 179          *        c = bic_scale >> 10
 180          * rtt  = (srtt >> 3) / HZ
 181          * !!! The following code does not have overflow problems,
 182          * if the cwnd < 1 million packets !!!
 183          */
 184
 185         /* change the unit from HZ to bictcp_HZ */
 186         t = ((tcp_time_stamp + (ca->delay_min>>3) - ca->epoch_start)
 187              << BICTCP_HZ) / HZ;
 188
 189         if (t < ca->bic_K)              /* t - K */
 190                 offs = ca->bic_K - t;
 191         else
 192                 offs = t - ca->bic_K;
 193
 194         /* c/rtt * (t-K)^3 */
 195         delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
 196         if (t < ca->bic_K)                                      /* below origin*/
 197                 bic_target = ca->bic_origin_point - delta;
 198         else                                                    /* above origin*/
 199                 bic_target = ca->bic_origin_point + delta;
 200
 201         /* cubic function - calc bictcp_cnt*/
 202         if (bic_target > cwnd) {
 203                 ca->cnt = cwnd / (bic_target - cwnd);
 204         } else {
 205                 ca->cnt = 100 * cwnd;              /* very small increment*/
 206         }
 207
 208         /* TCP Friendly */
 209         if (tcp_friendliness) {
 210                 u32 scale = beta_scale;
 211                 delta = (cwnd * scale) >> 3;
 212                 while (ca->ack_cnt > delta) {           /* update tcp cwnd */
 213                         ca->ack_cnt -= delta;
 214                         ca->tcp_cwnd++;
 215                 }
 216
 217                 if (ca->tcp_cwnd > cwnd){       /* if bic is slower than tcp */
 218                         delta = ca->tcp_cwnd - cwnd;
 219                         max_cnt = cwnd / delta;
 220                         if (ca->cnt > max_cnt)
 221                                 ca->cnt = max_cnt;
 222                 }
 223         }
 224
 225         ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
 226         if (ca->cnt == 0)                       /* cannot be zero */
 227                 ca->cnt = 1;
 228 }
 229
 230 static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 231 {
 232         struct tcp_sock *tp = tcp_sk(sk);
 233         struct bictcp *ca = inet_csk_ca(sk);
 234
 235         if (!tcp_is_cwnd_limited(sk, in_flight))
 236                 return;
 237
 238         if (tp->snd_cwnd <= tp->snd_ssthresh)
 239                 tcp_slow_start(tp);
 240         else {
 241                 bictcp_update(ca, tp->snd_cwnd);
 242
 243                 /* In dangerous area, increase slowly.
 244                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
 245                  */
 246                 if (tp->snd_cwnd_cnt >= ca->cnt) {
 247                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 248                                 tp->snd_cwnd++;
 249                         tp->snd_cwnd_cnt = 0;
 250                 } else
 251                         tp->snd_cwnd_cnt++;
 252         }
 253
 254 }
 255
 256 static u32 bictcp_recalc_ssthresh(struct sock *sk)
 257 {
 258         const struct tcp_sock *tp = tcp_sk(sk);
 259         struct bictcp *ca = inet_csk_ca(sk);
 260
 261         ca->epoch_start = 0;    /* end of epoch */
 262
 263         /* Wmax and fast convergence */
 264         if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
 265                 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
 266                         / (2 * BICTCP_BETA_SCALE);
 267         else
 268                 ca->last_max_cwnd = tp->snd_cwnd;
 269
 270         ca->loss_cwnd = tp->snd_cwnd;
 271
 272         return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 273 }
 274
 275 static u32 bictcp_undo_cwnd(struct sock *sk)
 276 {
 277         struct bictcp *ca = inet_csk_ca(sk);
 278
 279         return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
 280 }
 281
 282 static void bictcp_state(struct sock *sk, u8 new_state)
 283 {
 284         if (new_state == TCP_CA_Loss)
 285                 bictcp_reset(inet_csk_ca(sk));
 286 }
 287
 288 /* Track delayed acknowledgment ratio using sliding window
 289  * ratio = (15*ratio + sample) / 16
 290  */
 291 static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt_us)
 292 {
 293         const struct inet_connection_sock *icsk = inet_csk(sk);
 294         struct bictcp *ca = inet_csk_ca(sk);
 295         u32 delay;
 296
 297         if (icsk->icsk_ca_state == TCP_CA_Open) {
 298                 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
 299                 ca->delayed_ack += cnt;
 300         }
 301
 302         /* Some calls are for duplicates without timetamps */
 303         if (rtt_us < 0)
 304                 return;
 305
 306         /* Discard delay samples right after fast recovery */
 307         if ((s32)(tcp_time_stamp - ca->epoch_start) < HZ)
 308                 return;
 309
 310         delay = usecs_to_jiffies(rtt_us) << 3;
 311         if (delay == 0)
 312                 delay = 1;
 313
 314         /* first time call or link delay decreases */
 315         if (ca->delay_min == 0 || ca->delay_min > delay)
 316                 ca->delay_min = delay;
 317 }
 318
 319 static struct tcp_congestion_ops cubictcp = {
 320         .init           = bictcp_init,
 321         .ssthresh       = bictcp_recalc_ssthresh,
 322         .cong_avoid     = bictcp_cong_avoid,
 323         .set_state      = bictcp_state,
 324         .undo_cwnd      = bictcp_undo_cwnd,
 325         .pkts_acked     = bictcp_acked,
 326         .owner          = THIS_MODULE,
 327         .name           = "cubic",
 328 };
 329
 330 static int __init cubictcp_register(void)
 331 {
 332         BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
 333
 334         /* Precompute a bunch of the scaling factors that are used per-packet
 335          * based on SRTT of 100ms
 336          */
 337
 338         beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
 339
 340         cube_rtt_scale = (bic_scale * 10);      /* 1024*c/rtt */
 341
 342         /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
 343          *  so K = cubic_root( (wmax-cwnd)*rtt/c )
 344          * the unit of K is bictcp_HZ=2^10, not HZ
 345          *
 346          *  c = bic_scale >> 10
 347          *  rtt = 100ms
 348          *
 349          * the following code has been designed and tested for
 350          * cwnd < 1 million packets
 351          * RTT < 100 seconds
 352          * HZ < 1,000,00  (corresponding to 10 nano-second)
 353          */
 354
 355         /* 1/c * 2^2*bictcp_HZ * srtt */
 356         cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
 357
 358         /* divide by bic_scale and by constant Srtt (100ms) */
 359         do_div(cube_factor, bic_scale * 10);
 360
 361         return tcp_register_congestion_control(&cubictcp);
 362 }
 363
 364 static void __exit cubictcp_unregister(void)
 365 {
 366         tcp_unregister_congestion_control(&cubictcp);
 367 }
 368
 369 module_init(cubictcp_register);
 370 module_exit(cubictcp_unregister);
 371
 372 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
 373 MODULE_LICENSE("GPL");
 374 MODULE_DESCRIPTION("CUBIC TCP");
 375 MODULE_VERSION("2.2");