net/ipv4/tcp_bic.c

   1 /*
   2  * Binary Increase Congestion control for TCP
   3  *
   4  * This is from the implementation of BICTCP in
   5  * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
   6  *  "Binary Increase Congestion Control for Fast, Long Distance
   7  *  Networks" in InfoComm 2004
   8  * Available from:
   9  *  http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf
  10  *
  11  * Unless BIC 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
  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
  29 static int fast_convergence = 1;
  30 static int max_increment = 16;
  31 static int low_window = 14;
  32 static int beta = 819;          /* = 819/1024 (BICTCP_BETA_SCALE) */
  33 static int initial_ssthresh = 100;
  34 static int smooth_part = 20;
  35
  36 module_param(fast_convergence, int, 0644);
  37 MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  38 module_param(max_increment, int, 0644);
  39 MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
  40 module_param(low_window, int, 0644);
  41 MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
  42 module_param(beta, int, 0644);
  43 MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  44 module_param(initial_ssthresh, int, 0644);
  45 MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  46 module_param(smooth_part, int, 0644);
  47 MODULE_PARM_DESC(smooth_part, "log(B/(B*Smin))/log(B/(B-1))+B, # of RTT from Wmax-B to Wmax");
  48
  49
  50 /* BIC TCP Parameters */
  51 struct bictcp {
  52         u32     cnt;            /* increase cwnd by 1 after ACKs */
  53         u32     last_max_cwnd;  /* last maximum snd_cwnd */
  54         u32     loss_cwnd;      /* congestion window at last loss */
  55         u32     last_cwnd;      /* the last snd_cwnd */
  56         u32     last_time;      /* time when updated last_cwnd */
  57         u32     epoch_start;    /* beginning of an epoch */
  58 #define ACK_RATIO_SHIFT 4
  59         u32     delayed_ack;    /* estimate the ratio of Packets/ACKs << 4 */
  60 };
  61
  62 static inline void bictcp_reset(struct bictcp *ca)
  63 {
  64         ca->cnt = 0;
  65         ca->last_max_cwnd = 0;
  66         ca->loss_cwnd = 0;
  67         ca->last_cwnd = 0;
  68         ca->last_time = 0;
  69         ca->epoch_start = 0;
  70         ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
  71 }
  72
  73 static void bictcp_init(struct sock *sk)
  74 {
  75         bictcp_reset(inet_csk_ca(sk));
  76         if (initial_ssthresh)
  77                 tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
  78 }
  79
  80 /*
  81  * Compute congestion window to use.
  82  */
  83 static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
  84 {
  85         if (ca->last_cwnd == cwnd &&
  86             (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
  87                 return;
  88
  89         ca->last_cwnd = cwnd;
  90         ca->last_time = tcp_time_stamp;
  91
  92         if (ca->epoch_start == 0) /* record the beginning of an epoch */
  93                 ca->epoch_start = tcp_time_stamp;
  94
  95         /* start off normal */
  96         if (cwnd <= low_window) {
  97                 ca->cnt = cwnd;
  98                 return;
  99         }
 100
 101         /* binary increase */
 102         if (cwnd < ca->last_max_cwnd) {
 103                 __u32   dist = (ca->last_max_cwnd - cwnd)
 104                         / BICTCP_B;
 105
 106                 if (dist > max_increment)
 107                         /* linear increase */
 108                         ca->cnt = cwnd / max_increment;
 109                 else if (dist <= 1U)
 110                         /* binary search increase */
 111                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
 112                 else
 113                         /* binary search increase */
 114                         ca->cnt = cwnd / dist;
 115         } else {
 116                 /* slow start AMD linear increase */
 117                 if (cwnd < ca->last_max_cwnd + BICTCP_B)
 118                         /* slow start */
 119                         ca->cnt = (cwnd * smooth_part) / BICTCP_B;
 120                 else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
 121                         /* slow start */
 122                         ca->cnt = (cwnd * (BICTCP_B-1))
 123                                 / (cwnd - ca->last_max_cwnd);
 124                 else
 125                         /* linear increase */
 126                         ca->cnt = cwnd / max_increment;
 127         }
 128
 129         /* if in slow start or link utilization is very low */
 130         if (ca->loss_cwnd == 0) {
 131                 if (ca->cnt > 20) /* increase cwnd 5% per RTT */
 132                         ca->cnt = 20;
 133         }
 134
 135         ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
 136         if (ca->cnt == 0)                       /* cannot be zero */
 137                 ca->cnt = 1;
 138 }
 139
 140 static void bictcp_cong_avoid(struct sock *sk, u32 ack,
 141                               u32 seq_rtt, u32 in_flight, int data_acked)
 142 {
 143         struct tcp_sock *tp = tcp_sk(sk);
 144         struct bictcp *ca = inet_csk_ca(sk);
 145
 146         if (!tcp_is_cwnd_limited(sk, in_flight))
 147                 return;
 148
 149         if (tp->snd_cwnd <= tp->snd_ssthresh)
 150                 tcp_slow_start(tp);
 151         else {
 152                 bictcp_update(ca, tp->snd_cwnd);
 153
 154                 /* In dangerous area, increase slowly.
 155                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
 156                  */
 157                 if (tp->snd_cwnd_cnt >= ca->cnt) {
 158                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 159                                 tp->snd_cwnd++;
 160                         tp->snd_cwnd_cnt = 0;
 161                 } else
 162                         tp->snd_cwnd_cnt++;
 163         }
 164
 165 }
 166
 167 /*
 168  *      behave like Reno until low_window is reached,
 169  *      then increase congestion window slowly
 170  */
 171 static u32 bictcp_recalc_ssthresh(struct sock *sk)
 172 {
 173         const struct tcp_sock *tp = tcp_sk(sk);
 174         struct bictcp *ca = inet_csk_ca(sk);
 175
 176         ca->epoch_start = 0;    /* end of epoch */
 177
 178         /* Wmax and fast convergence */
 179         if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
 180                 ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
 181                         / (2 * BICTCP_BETA_SCALE);
 182         else
 183                 ca->last_max_cwnd = tp->snd_cwnd;
 184
 185         ca->loss_cwnd = tp->snd_cwnd;
 186
 187
 188         if (tp->snd_cwnd <= low_window)
 189                 return max(tp->snd_cwnd >> 1U, 2U);
 190         else
 191                 return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 192 }
 193
 194 static u32 bictcp_undo_cwnd(struct sock *sk)
 195 {
 196         const struct tcp_sock *tp = tcp_sk(sk);
 197         const struct bictcp *ca = inet_csk_ca(sk);
 198         return max(tp->snd_cwnd, ca->last_max_cwnd);
 199 }
 200
 201 static u32 bictcp_min_cwnd(struct sock *sk)
 202 {
 203         const struct tcp_sock *tp = tcp_sk(sk);
 204         return tp->snd_ssthresh;
 205 }
 206
 207 static void bictcp_state(struct sock *sk, u8 new_state)
 208 {
 209         if (new_state == TCP_CA_Loss)
 210                 bictcp_reset(inet_csk_ca(sk));
 211 }
 212
 213 /* Track delayed acknowledgment ratio using sliding window
 214  * ratio = (15*ratio + sample) / 16
 215  */
 216 static void bictcp_acked(struct sock *sk, u32 cnt)
 217 {
 218         const struct inet_connection_sock *icsk = inet_csk(sk);
 219
 220         if (cnt > 0 && icsk->icsk_ca_state == TCP_CA_Open) {
 221                 struct bictcp *ca = inet_csk_ca(sk);
 222                 cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
 223                 ca->delayed_ack += cnt;
 224         }
 225 }
 226
 227
 228 static struct tcp_congestion_ops bictcp = {
 229         .init           = bictcp_init,
 230         .ssthresh       = bictcp_recalc_ssthresh,
 231         .cong_avoid     = bictcp_cong_avoid,
 232         .set_state      = bictcp_state,
 233         .undo_cwnd      = bictcp_undo_cwnd,
 234         .min_cwnd       = bictcp_min_cwnd,
 235         .pkts_acked     = bictcp_acked,
 236         .owner          = THIS_MODULE,
 237         .name           = "bic",
 238 };
 239
 240 static int __init bictcp_register(void)
 241 {
 242         BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
 243         return tcp_register_congestion_control(&bictcp);
 244 }
 245
 246 static void __exit bictcp_unregister(void)
 247 {
 248         tcp_unregister_congestion_control(&bictcp);
 249 }
 250
 251 module_init(bictcp_register);
 252 module_exit(bictcp_unregister);
 253
 254 MODULE_AUTHOR("Stephen Hemminger");
 255 MODULE_LICENSE("GPL");
 256 MODULE_DESCRIPTION("BIC TCP");