linux/net/ipv4/tcp_bic.c
<<
>>
Prefs
   1/*
   2 * Binary Increase Congestion control for TCP
   3 * Home page:
   4 *      http://netsrv.csc.ncsu.edu/twiki/bin/view/Main/BIC
   5 * This is from the implementation of BICTCP in
   6 * Lison-Xu, Kahaled Harfoush, and Injong Rhee.
   7 *  "Binary Increase Congestion Control for Fast, Long Distance
   8 *  Networks" in InfoComm 2004
   9 * Available from:
  10 *  http://netsrv.csc.ncsu.edu/export/bitcp.pdf
  11 *
  12 * Unless BIC 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 <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
  29static int fast_convergence = 1;
  30static int max_increment = 16;
  31static int low_window = 14;
  32static int beta = 819;          /* = 819/1024 (BICTCP_BETA_SCALE) */
  33static int initial_ssthresh;
  34static int smooth_part = 20;
  35
  36module_param(fast_convergence, int, 0644);
  37MODULE_PARM_DESC(fast_convergence, "turn on/off fast convergence");
  38module_param(max_increment, int, 0644);
  39MODULE_PARM_DESC(max_increment, "Limit on increment allowed during binary search");
  40module_param(low_window, int, 0644);
  41MODULE_PARM_DESC(low_window, "lower bound on congestion window (for TCP friendliness)");
  42module_param(beta, int, 0644);
  43MODULE_PARM_DESC(beta, "beta for multiplicative increase");
  44module_param(initial_ssthresh, int, 0644);
  45MODULE_PARM_DESC(initial_ssthresh, "initial value of slow start threshold");
  46module_param(smooth_part, int, 0644);
  47MODULE_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 */
  51struct 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
  62static inline void bictcp_reset(struct bictcp *ca)
  63{
  64        ca->cnt = 0;
  65        ca->last_max_cwnd = 0;
  66        ca->last_cwnd = 0;
  67        ca->last_time = 0;
  68        ca->epoch_start = 0;
  69        ca->delayed_ack = 2 << ACK_RATIO_SHIFT;
  70}
  71
  72static void bictcp_init(struct sock *sk)
  73{
  74        struct bictcp *ca = inet_csk_ca(sk);
  75
  76        bictcp_reset(ca);
  77        ca->loss_cwnd = 0;
  78
  79        if (initial_ssthresh)
  80                tcp_sk(sk)->snd_ssthresh = initial_ssthresh;
  81}
  82
  83/*
  84 * Compute congestion window to use.
  85 */
  86static inline void bictcp_update(struct bictcp *ca, u32 cwnd)
  87{
  88        if (ca->last_cwnd == cwnd &&
  89            (s32)(tcp_time_stamp - ca->last_time) <= HZ / 32)
  90                return;
  91
  92        ca->last_cwnd = cwnd;
  93        ca->last_time = tcp_time_stamp;
  94
  95        if (ca->epoch_start == 0) /* record the beginning of an epoch */
  96                ca->epoch_start = tcp_time_stamp;
  97
  98        /* start off normal */
  99        if (cwnd <= low_window) {
 100                ca->cnt = cwnd;
 101                return;
 102        }
 103
 104        /* binary increase */
 105        if (cwnd < ca->last_max_cwnd) {
 106                __u32   dist = (ca->last_max_cwnd - cwnd)
 107                        / BICTCP_B;
 108
 109                if (dist > max_increment)
 110                        /* linear increase */
 111                        ca->cnt = cwnd / max_increment;
 112                else if (dist <= 1U)
 113                        /* binary search increase */
 114                        ca->cnt = (cwnd * smooth_part) / BICTCP_B;
 115                else
 116                        /* binary search increase */
 117                        ca->cnt = cwnd / dist;
 118        } else {
 119                /* slow start AMD linear increase */
 120                if (cwnd < ca->last_max_cwnd + BICTCP_B)
 121                        /* slow start */
 122                        ca->cnt = (cwnd * smooth_part) / BICTCP_B;
 123                else if (cwnd < ca->last_max_cwnd + max_increment*(BICTCP_B-1))
 124                        /* slow start */
 125                        ca->cnt = (cwnd * (BICTCP_B-1))
 126                                / (cwnd - ca->last_max_cwnd);
 127                else
 128                        /* linear increase */
 129                        ca->cnt = cwnd / max_increment;
 130        }
 131
 132        /* if in slow start or link utilization is very low */
 133        if (ca->last_max_cwnd == 0) {
 134                if (ca->cnt > 20) /* increase cwnd 5% per RTT */
 135                        ca->cnt = 20;
 136        }
 137
 138        ca->cnt = (ca->cnt << ACK_RATIO_SHIFT) / ca->delayed_ack;
 139        if (ca->cnt == 0)                       /* cannot be zero */
 140                ca->cnt = 1;
 141}
 142
 143static void bictcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 144{
 145        struct tcp_sock *tp = tcp_sk(sk);
 146        struct bictcp *ca = inet_csk_ca(sk);
 147
 148        if (!tcp_is_cwnd_limited(sk, in_flight))
 149                return;
 150
 151        if (tp->snd_cwnd <= tp->snd_ssthresh)
 152                tcp_slow_start(tp);
 153        else {
 154                bictcp_update(ca, tp->snd_cwnd);
 155                tcp_cong_avoid_ai(tp, ca->cnt);
 156        }
 157
 158}
 159
 160/*
 161 *      behave like Reno until low_window is reached,
 162 *      then increase congestion window slowly
 163 */
 164static u32 bictcp_recalc_ssthresh(struct sock *sk)
 165{
 166        const struct tcp_sock *tp = tcp_sk(sk);
 167        struct bictcp *ca = inet_csk_ca(sk);
 168
 169        ca->epoch_start = 0;    /* end of epoch */
 170
 171        /* Wmax and fast convergence */
 172        if (tp->snd_cwnd < ca->last_max_cwnd && fast_convergence)
 173                ca->last_max_cwnd = (tp->snd_cwnd * (BICTCP_BETA_SCALE + beta))
 174                        / (2 * BICTCP_BETA_SCALE);
 175        else
 176                ca->last_max_cwnd = tp->snd_cwnd;
 177
 178        ca->loss_cwnd = tp->snd_cwnd;
 179
 180
 181        if (tp->snd_cwnd <= low_window)
 182                return max(tp->snd_cwnd >> 1U, 2U);
 183        else
 184                return max((tp->snd_cwnd * beta) / BICTCP_BETA_SCALE, 2U);
 185}
 186
 187static u32 bictcp_undo_cwnd(struct sock *sk)
 188{
 189        const struct tcp_sock *tp = tcp_sk(sk);
 190        const struct bictcp *ca = inet_csk_ca(sk);
 191        return max(tp->snd_cwnd, ca->loss_cwnd);
 192}
 193
 194static void bictcp_state(struct sock *sk, u8 new_state)
 195{
 196        if (new_state == TCP_CA_Loss)
 197                bictcp_reset(inet_csk_ca(sk));
 198}
 199
 200/* Track delayed acknowledgment ratio using sliding window
 201 * ratio = (15*ratio + sample) / 16
 202 */
 203static void bictcp_acked(struct sock *sk, u32 cnt, s32 rtt)
 204{
 205        const struct inet_connection_sock *icsk = inet_csk(sk);
 206
 207        if (icsk->icsk_ca_state == TCP_CA_Open) {
 208                struct bictcp *ca = inet_csk_ca(sk);
 209                cnt -= ca->delayed_ack >> ACK_RATIO_SHIFT;
 210                ca->delayed_ack += cnt;
 211        }
 212}
 213
 214
 215static struct tcp_congestion_ops bictcp __read_mostly = {
 216        .init           = bictcp_init,
 217        .ssthresh       = bictcp_recalc_ssthresh,
 218        .cong_avoid     = bictcp_cong_avoid,
 219        .set_state      = bictcp_state,
 220        .undo_cwnd      = bictcp_undo_cwnd,
 221        .pkts_acked     = bictcp_acked,
 222        .owner          = THIS_MODULE,
 223        .name           = "bic",
 224};
 225
 226static int __init bictcp_register(void)
 227{
 228        BUILD_BUG_ON(sizeof(struct bictcp) > ICSK_CA_PRIV_SIZE);
 229        return tcp_register_congestion_control(&bictcp);
 230}
 231
 232static void __exit bictcp_unregister(void)
 233{
 234        tcp_unregister_congestion_control(&bictcp);
 235}
 236
 237module_init(bictcp_register);
 238module_exit(bictcp_unregister);
 239
 240MODULE_AUTHOR("Stephen Hemminger");
 241MODULE_LICENSE("GPL");
 242MODULE_DESCRIPTION("BIC TCP");
 243
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.