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