net/ipv4/tcp_cubic.c

   1 /*
   2  * TCP CUBIC: Binary Increase Congestion control for TCP v2.0
   3  *
   4  * This is from the implementation of CUBIC TCP in
   5  * Injong Rhee, Lisong Xu.
   6  *  "CUBIC: A New TCP-Friendly High-Speed TCP Variant
   7  *  in PFLDnet 2005
   8  * Available from:
   9  *  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp/cubic-paper.pdf
  10  *
  11  * Unless CUBIC is enabled and congestion window is large
  12  * this behaves the same as the original Reno.
  13  */
  14
  15 #include <linux/config.h>
  16 #include <linux/mm.h>
  17 #include <linux/module.h>
  18 #include <net/tcp.h>
  19 #include <asm/div64.h>
  20
  21 #define BICTCP_BETA_SCALE    1024       /* Scale factor beta calculation
  22                                          * max_cwnd = snd_cwnd * beta
  23                                          */
  24 #define BICTCP_B                4        /*
  25                                           * In binary search,
  26                                           * go to point (max+min)/N
  27                                           */
  28 #define BICTCP_HZ               10      /* BIC HZ 2^10 = 1024 */
  29
  30 static int fast_convergence = 1;
  31 static int max_increment = 16;
  32 static int beta = 819;          /* = 819/1024 (BICTCP_BETA_SCALE) */
  33 static int initial_ssthresh = 100;
  34 static int bic_scale = 41;
  35 static int tcp_friendliness = 1;
  36
  37 static u32 cube_rtt_scale;
  38 static u32 beta_scale;
  39 static u64 cube_factor;
  40
  41 /* Note parameters that are used for precomputing scale factors are read-only */
  42 module_param(fast_convergence, int, 0644);
  43 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  44 module_param(max_increment, int, 0644);
  45 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
  46 module_param(beta, int, 0444);
  47 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  48 module_param(initial_ssthresh, int, 0644);
  49 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  50 module_param(bic_scale, int, 0444);
  51 MODULE_PARM_DESC(bic_scale, "scale (scaled by 1024) value for bic function (bic_scale/1024)");
  52 module_param(tcp_friendliness, int, 0644);
  53 MODULE_PARM_DESC(tcp_friendliness, "turn on/off tcp friendliness");
  54
  55 #include <asm/div64.h>
  56
  57 /* BIC TCP Parameters */
  58 struct bictcp {
  59         u32     cnt;            /* increase cwnd by 1 after ACKs */
  60         u32     last_max_cwnd;  /* last maximum snd_cwnd */
  61         u32     loss_cwnd;      /* congestion window at last loss */
  62         u32     last_cwnd;      /* the last snd_cwnd */
  63         u32     last_time;      /* time when updated last_cwnd */
  64         u32     bic_origin_point;/* origin point of bic function */
  65         u32     bic_K;          /* time to origin point from the beginning of the current epoch */
  66         u32     delay_min;      /* min delay */
  67         u32     epoch_start;    /* beginning of an epoch */
  68         u32     ack_cnt;        /* number of acks */
  69         u32     tcp_cwnd;       /* estimated tcp cwnd */
  70 #define ACK_RATIO_SHIFT 4
  71         u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
  72 };
  73
  74 static inline void bictcp_reset(struct bictcp *ca)
  75 {
  76         ca->cnt = 0;
  77         ca->last_max_cwnd = 0;
  78         ca->loss_cwnd = 0;
  79         ca->last_cwnd = 0;
  80         ca->last_time = 0;
  81         ca->bic_origin_point = 0;
  82         ca->bic_K = 0;
  83         ca->delay_min = 0;
  84         ca->epoch_start = 0;
  85         ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
  86         ca->ack_cnt = 0;
  87         ca->tcp_cwnd = 0;
  88 }
  89
  90 static void bictcp_init(struct sock *sk)
  91 {
  92         bictcp_reset(inet_csk_ca(sk));
  93         if (initial_ssthresh)
  94                 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
  95 }
  96
  97 /* 64bit divisor, dividend and result. dynamic precision */
  98 static inline u_int64_t div64_64(u_int64_t dividend, u_int64_t divisor)
  99 {
 100         u_int32_t d = divisor;
 101
 102         if (divisor > 0xffffffffULL) {
 103                 unsigned int shift = fls(divisor >> 32);
 104
 105                 d = divisor >> shift;
 106                 dividend >>= shift;
 107         }
 108
 109         /* avoid 64 bit division if possible */
 110         if (dividend >> 32)
 111                 do_div(dividend, d);
 112         else
 113                 dividend = (uint32_t) dividend / d;
 114
 115         return dividend;
 116 }
 117
 118 /*
 119  * calculate the cubic root of x using Newton-Raphson
 120  */
 121 static u32 cubic_root(u64 a)
 122 {
 123         u32 x, x1;
 124
 125         /* Initial estimate is based on:
 126          * cbrt(x) = exp(log(x) / 3)
 127          */
 128         x = 1u << (fls64(a)/3);
 129
 130         /*
 131          * Iteration based on:
 132          *                         2
 133          * x    = ( 2 * x  +  a / x  ) / 3
 134          *  k+1          k         k
 135          */
 136         do {
 137                 x1 = x;
 138                 x = (2 * x + (uint32_t) div64_64(a, x*x)) / 3;
 139         } while (abs(x1 - x) > 1);
 140
 141         return x;
 142 }
 143
 144 /*
 145  * Compute congestion window to use.
 146  */
 147 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
 148 {
 149         u64 offs;
 150         u32 delta, t, bic_target, min_cnt, max_cnt;
 151
 152         ca->ack_cnt++;  /* count the number of ACKs */
 153
 154         if (ca->last_cwnd == cwnd &&
 155             (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
 156                 return;
 157
 158         ca->last_cwnd = cwnd;
 159         ca->last_time = tcp_time_stamp;
 160
 161         if (ca->epoch_start == 0) {
 162                 ca->epoch_start = tcp_time_stamp;       /* record the beginning of an epoch */
 163                 ca->ack_cnt = 1;                        /* start counting */
 164                 ca->tcp_cwnd = cwnd;                    /* syn with cubic */
 165
 166                 if (ca->last_max_cwnd <= cwnd) {
 167                         ca->bic_K = 0;
 168                         ca->bic_origin_point = cwnd;
 169                 } else {
 170                         /* Compute new K based on
 171                          * (wmax-cwnd) * (srtt>>3 / HZ) / c * 2^(3*bictcp_HZ)
 172                          */
 173                         ca->bic_K = cubic_root(cube_factor
 174                                                * (ca->last_max_cwnd - cwnd));
 175                         ca->bic_origin_point = ca->last_max_cwnd;
 176                 }
 177         }
 178
 179         /* cubic function - calc*/
 180         /* calculate c * time^3 / rtt,
 181          *  while considering overflow in calculation of time^3
 182          * (so time^3 is done by using 64 bit)
 183          * and without the support of division of 64bit numbers
 184          * (so all divisions are done by using 32 bit)
 185          *  also NOTE the unit of those veriables
 186          *        time  = (t - K) / 2^bictcp_HZ
 187          *        c = bic_scale >> 10
 188          * rtt  = (srtt >> 3) / HZ
 189          * !!! The following code does not have overflow problems,
 190          * if the cwnd < 1 million packets !!!
 191          */
 192
 193         /* change the unit from HZ to bictcp_HZ */
 194         t = ((tcp_time_stamp + ca->delay_min - ca->epoch_start)
 195              << BICTCP_HZ) / HZ;
 196
 197         if (t < ca->bic_K)              /* t - K */
 198                 offs = ca->bic_K - t;
 199         else
 200                 offs = t - ca->bic_K;
 201
 202         /* c/rtt * (t-K)^3 */
 203         delta = (cube_rtt_scale * offs * offs * offs) >> (10+3*BICTCP_HZ);
 204         if (t < ca->bic_K)                                      /* below origin*/
 205                 bic_target = ca->bic_origin_point - delta;
 206         else                                                    /* above origin*/
 207                 bic_target = ca->bic_origin_point + delta;
 208
 209         /* cubic function - calc bictcp_cnt*/
 210         if (bic_target > cwnd) {
 211                 ca->cnt = cwnd / (bic_target - cwnd);
 212         } else {
 213                 ca->cnt = 100 * cwnd;              /* very small increment*/
 214         }
 215
 216         if (ca->delay_min > 0) {
 217                 /* max increment = Smax * rtt / 0.1  */
 218                 min_cnt = (cwnd * HZ * 8)/(10 * max_increment * ca->delay_min);
 219                 if (ca->cnt < min_cnt)
 220                         ca->cnt = min_cnt;
 221         }
 222
 223         /* slow start and low utilization  */
 224         if (ca->loss_cwnd == 0)         /* could be aggressive in slow start */
 225                 ca->cnt = 50;
 226
 227         /* TCP Friendly */
 228         if (tcp_friendliness) {
 229                 u32 scale = beta_scale;
 230                 delta = (cwnd * scale) >> 3;
 231                 while (ca->ack_cnt > delta) {           /* update tcp cwnd */
 232                         ca->ack_cnt -= delta;
 233                         ca->tcp_cwnd++;
 234                 }
 235
 236                 if (ca->tcp_cwnd > cwnd){       /* if bic is slower than tcp */
 237                         delta = ca->tcp_cwnd - cwnd;
 238                         max_cnt = cwnd / delta;
 239                         if (ca->cnt > max_cnt)
 240                                 ca->cnt = max_cnt;
 241                 }
 242         }
 243
 244         ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
 245         if (ca->cnt == 0)                       /* cannot be zero */
 246                 ca->cnt = 1;
 247 }
 248
 249
 250 /* Keep track of minimum rtt */
 251 static inline void measure_delay(struct sock *sk)
 252 {
 253         const struct tcp_sock *tp = tcp_sk(sk);
 254         struct bictcp *ca = inet_csk_ca(sk);
 255         u32 delay;
 256
 257         /* No time stamp */
 258         if (!(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) ||
 259              /* Discard delay samples right after fast recovery */
 260             (s32)(tcp_time_stamp - ca->epoch_start) < HZ)
 261                 return;
 262
 263         delay = tcp_time_stamp - tp->rx_opt.rcv_tsecr;
 264         if (delay == 0)
 265                 delay = 1;
 266
 267         /* first time call or link delay decreases */
 268         if (ca->delay_min == 0 || ca->delay_min > delay)
 269                 ca->delay_min = delay;
 270 }
 271
 272 static void bictcp_cong_avoid(struct sock *sk, u32 ack,
 273                               u32 seq_rtt, u32 in_flight, int data_acked)
 274 {
 275         struct tcp_sock *tp = tcp_sk(sk);
 276         struct bictcp *ca = inet_csk_ca(sk);
 277
 278         if (data_acked)
 279                 measure_delay(sk);
 280
 281         if (!tcp_is_cwnd_limited(sk, in_flight))
 282                 return;
 283
 284         if (tp->snd_cwnd <= tp->snd_ssthresh)
 285                 tcp_slow_start(tp);
 286         else {
 287                 bictcp_update(ca, tp->snd_cwnd);
 288
 289                 /* In dangerous area, increase slowly.
 290                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
 291                  */
 292                 if (tp->snd_cwnd_cnt >= ca->cnt) {
 293                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 294                                 tp->snd_cwnd++;
 295                         tp->snd_cwnd_cnt = 0;
 296                 } else
 297                         tp->snd_cwnd_cnt++;
 298         }
 299
 300 }
 301
 302 static u32 bictcp_recalc_ssthresh(struct sock *sk)
 303 {
 304         const struct tcp_sock *tp = tcp_sk(sk);
 305         struct bictcp *ca = inet_csk_ca(sk);
 306
 307         ca->epoch_start = 0;    /* end of epoch */
 308
 309         /* Wmax and fast convergence */
 310         if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
 311                 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
 312                         / (2 * BICTCP_BETA_SCALE);
 313         else
 314                 ca->last_max_cwnd = tp->snd_cwnd;
 315
 316         ca->loss_cwnd = tp->snd_cwnd;
 317
 318         return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 319 }
 320
 321 static u32 bictcp_undo_cwnd(struct sock *sk)
 322 {
 323         struct bictcp *ca = inet_csk_ca(sk);
 324
 325         return max(tcp_sk(sk)->snd_cwnd, ca->last_max_cwnd);
 326 }
 327
 328 static void bictcp_state(struct sock *sk, u8 new_state)
 329 {
 330         if (new_state == TCP_CA_Loss)
 331                 bictcp_reset(inet_csk_ca(sk));
 332 }
 333
 334 /* Track delayed acknowledgment ratio using sliding window
 335  * ratio = (15*ratio + sample) / 16
 336  */
 337 static void bictcp_acked(struct sock *sk, u32 cnt)
 338 {
 339         const struct inet_connection_sock *icsk = inet_csk(sk);
 340
 341         if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
 342                 struct bictcp *ca = inet_csk_ca(sk);
 343                 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
 344                 ca->delayed_ack += cnt;
 345         }
 346 }
 347
 348
 349 static struct tcp_congestion_ops cubictcp = {
 350         .init           = bictcp_init,
 351         .ssthresh       = bictcp_recalc_ssthresh,
 352         .cong_avoid     = bictcp_cong_avoid,
 353         .set_state      = bictcp_state,
 354         .undo_cwnd      = bictcp_undo_cwnd,
 355         .pkts_acked     = bictcp_acked,
 356         .owner          = THIS_MODULE,
 357         .name           = "cubic",
 358 };
 359
 360 static int __init cubictcp_register(void)
 361 {
 362         BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
 363
 364         /* Precompute a bunch of the scaling factors that are used per-packet
 365          * based on SRTT of 100ms
 366          */
 367
 368         beta_scale = 8*(BICTCP_BETA_SCALE+beta)/ 3 / (BICTCP_BETA_SCALE - beta);
 369
 370         cube_rtt_scale = (bic_scale << 3) / 10; /* 1024*c/rtt */
 371
 372         /* calculate the "K" for (wmax-cwnd) = c/rtt * K^3
 373          *  so K = cubic_root( (wmax-cwnd)*rtt/c )
 374          * the unit of K is bictcp_HZ=2^10, not HZ
 375          *
 376          *  c = bic_scale >> 10
 377          *  rtt = 100ms
 378          *
 379          * the following code has been designed and tested for
 380          * cwnd < 1 million packets
 381          * RTT < 100 seconds
 382          * HZ < 1,000,00  (corresponding to 10 nano-second)
 383          */
 384
 385         /* 1/c * 2^2*bictcp_HZ * srtt */
 386         cube_factor = 1ull << (10+3*BICTCP_HZ); /* 2^40 */
 387
 388         /* divide by bic_scale and by constant Srtt (100ms) */
 389         do_div(cube_factor, bic_scale * 10);
 390
 391         return tcp_register_congestion_control(&cubictcp);
 392 }
 393
 394 static void __exit cubictcp_unregister(void)
 395 {
 396         tcp_unregister_congestion_control(&cubictcp);
 397 }
 398
 399 module_init(cubictcp_register);
 400 module_exit(cubictcp_unregister);
 401
 402 MODULE_AUTHOR("Sangtae Ha, Stephen Hemminger");
 403 MODULE_LICENSE("GPL");
 404 MODULE_DESCRIPTION("CUBIC TCP");
 405 MODULE_VERSION("2.0");