linux/net/ipv4/tcp_htcp.c
<<
>>
Prefs
   1/*
   2 * H-TCP congestion control. The algorithm is detailed in:
   3 * R.N.Shorten, D.J.Leith:
   4 *   "H-TCP: TCP for high-speed and long-distance networks"
   5 *   Proc. PFLDnet, Argonne, 2004.
   6 * http://www.hamilton.ie/net/htcp3.pdf
   7 */
   8
   9#include <linux/mm.h>
  10#include <linux/module.h>
  11#include <net/tcp.h>
  12
  13#define ALPHA_BASE      (1<<7)  /* 1.0 with shift << 7 */
  14#define BETA_MIN        (1<<6)  /* 0.5 with shift << 7 */
  15#define BETA_MAX        102     /* 0.8 with shift << 7 */
  16
  17static int use_rtt_scaling __read_mostly = 1;
  18module_param(use_rtt_scaling, int, 0644);
  19MODULE_PARM_DESC(use_rtt_scaling, "turn on/off RTT scaling");
  20
  21static int use_bandwidth_switch __read_mostly = 1;
  22module_param(use_bandwidth_switch, int, 0644);
  23MODULE_PARM_DESC(use_bandwidth_switch, "turn on/off bandwidth switcher");
  24
  25struct htcp {
  26        u32     alpha;          /* Fixed point arith, << 7 */
  27        u8      beta;           /* Fixed point arith, << 7 */
  28        u8      modeswitch;     /* Delay modeswitch
  29                                   until we had at least one congestion event */
  30        u16     pkts_acked;
  31        u32     packetcount;
  32        u32     minRTT;
  33        u32     maxRTT;
  34        u32     last_cong;      /* Time since last congestion event end */
  35        u32     undo_last_cong;
  36
  37        u32     undo_maxRTT;
  38        u32     undo_old_maxB;
  39
  40        /* Bandwidth estimation */
  41        u32     minB;
  42        u32     maxB;
  43        u32     old_maxB;
  44        u32     Bi;
  45        u32     lasttime;
  46};
  47
  48static inline u32 htcp_cong_time(const struct htcp *ca)
  49{
  50        return jiffies - ca->last_cong;
  51}
  52
  53static inline u32 htcp_ccount(const struct htcp *ca)
  54{
  55        return htcp_cong_time(ca) / ca->minRTT;
  56}
  57
  58static inline void htcp_reset(struct htcp *ca)
  59{
  60        ca->undo_last_cong = ca->last_cong;
  61        ca->undo_maxRTT = ca->maxRTT;
  62        ca->undo_old_maxB = ca->old_maxB;
  63
  64        ca->last_cong = jiffies;
  65}
  66
  67static u32 htcp_cwnd_undo(struct sock *sk)
  68{
  69        const struct tcp_sock *tp = tcp_sk(sk);
  70        struct htcp *ca = inet_csk_ca(sk);
  71
  72        if (ca->undo_last_cong) {
  73                ca->last_cong = ca->undo_last_cong;
  74                ca->maxRTT = ca->undo_maxRTT;
  75                ca->old_maxB = ca->undo_old_maxB;
  76                ca->undo_last_cong = 0;
  77        }
  78
  79        return max(tp->snd_cwnd, (tp->snd_ssthresh << 7) / ca->beta);
  80}
  81
  82static inline void measure_rtt(struct sock *sk, u32 srtt)
  83{
  84        const struct inet_connection_sock *icsk = inet_csk(sk);
  85        struct htcp *ca = inet_csk_ca(sk);
  86
  87        /* keep track of minimum RTT seen so far, minRTT is zero at first */
  88        if (ca->minRTT > srtt || !ca->minRTT)
  89                ca->minRTT = srtt;
  90
  91        /* max RTT */
  92        if (icsk->icsk_ca_state == TCP_CA_Open) {
  93                if (ca->maxRTT < ca->minRTT)
  94                        ca->maxRTT = ca->minRTT;
  95                if (ca->maxRTT < srtt &&
  96                    srtt <= ca->maxRTT + msecs_to_jiffies(20))
  97                        ca->maxRTT = srtt;
  98        }
  99}
 100
 101static void measure_achieved_throughput(struct sock *sk, u32 pkts_acked, s32 rtt)
 102{
 103        const struct inet_connection_sock *icsk = inet_csk(sk);
 104        const struct tcp_sock *tp = tcp_sk(sk);
 105        struct htcp *ca = inet_csk_ca(sk);
 106        u32 now = tcp_time_stamp;
 107
 108        if (icsk->icsk_ca_state == TCP_CA_Open)
 109                ca->pkts_acked = pkts_acked;
 110
 111        if (rtt > 0)
 112                measure_rtt(sk, usecs_to_jiffies(rtt));
 113
 114        if (!use_bandwidth_switch)
 115                return;
 116
 117        /* achieved throughput calculations */
 118        if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_Disorder))) {
 119                ca->packetcount = 0;
 120                ca->lasttime = now;
 121                return;
 122        }
 123
 124        ca->packetcount += pkts_acked;
 125
 126        if (ca->packetcount >= tp->snd_cwnd - (ca->alpha >> 7 ? : 1) &&
 127            now - ca->lasttime >= ca->minRTT &&
 128            ca->minRTT > 0) {
 129                __u32 cur_Bi = ca->packetcount * HZ / (now - ca->lasttime);
 130
 131                if (htcp_ccount(ca) <= 3) {
 132                        /* just after backoff */
 133                        ca->minB = ca->maxB = ca->Bi = cur_Bi;
 134                } else {
 135                        ca->Bi = (3 * ca->Bi + cur_Bi) / 4;
 136                        if (ca->Bi > ca->maxB)
 137                                ca->maxB = ca->Bi;
 138                        if (ca->minB > ca->maxB)
 139                                ca->minB = ca->maxB;
 140                }
 141                ca->packetcount = 0;
 142                ca->lasttime = now;
 143        }
 144}
 145
 146static inline void htcp_beta_update(struct htcp *ca, u32 minRTT, u32 maxRTT)
 147{
 148        if (use_bandwidth_switch) {
 149                u32 maxB = ca->maxB;
 150                u32 old_maxB = ca->old_maxB;
 151                ca->old_maxB = ca->maxB;
 152
 153                if (!between(5 * maxB, 4 * old_maxB, 6 * old_maxB)) {
 154                        ca->beta = BETA_MIN;
 155                        ca->modeswitch = 0;
 156                        return;
 157                }
 158        }
 159
 160        if (ca->modeswitch && minRTT > msecs_to_jiffies(10) && maxRTT) {
 161                ca->beta = (minRTT << 7) / maxRTT;
 162                if (ca->beta < BETA_MIN)
 163                        ca->beta = BETA_MIN;
 164                else if (ca->beta > BETA_MAX)
 165                        ca->beta = BETA_MAX;
 166        } else {
 167                ca->beta = BETA_MIN;
 168                ca->modeswitch = 1;
 169        }
 170}
 171
 172static inline void htcp_alpha_update(struct htcp *ca)
 173{
 174        u32 minRTT = ca->minRTT;
 175        u32 factor = 1;
 176        u32 diff = htcp_cong_time(ca);
 177
 178        if (diff > HZ) {
 179                diff -= HZ;
 180                factor = 1 + (10 * diff + ((diff / 2) * (diff / 2) / HZ)) / HZ;
 181        }
 182
 183        if (use_rtt_scaling && minRTT) {
 184                u32 scale = (HZ << 3) / (10 * minRTT);
 185
 186                /* clamping ratio to interval [0.5,10]<<3 */
 187                scale = min(max(scale, 1U << 2), 10U << 3);
 188                factor = (factor << 3) / scale;
 189                if (!factor)
 190                        factor = 1;
 191        }
 192
 193        ca->alpha = 2 * factor * ((1 << 7) - ca->beta);
 194        if (!ca->alpha)
 195                ca->alpha = ALPHA_BASE;
 196}
 197
 198/*
 199 * After we have the rtt data to calculate beta, we'd still prefer to wait one
 200 * rtt before we adjust our beta to ensure we are working from a consistent
 201 * data.
 202 *
 203 * This function should be called when we hit a congestion event since only at
 204 * that point do we really have a real sense of maxRTT (the queues en route
 205 * were getting just too full now).
 206 */
 207static void htcp_param_update(struct sock *sk)
 208{
 209        struct htcp *ca = inet_csk_ca(sk);
 210        u32 minRTT = ca->minRTT;
 211        u32 maxRTT = ca->maxRTT;
 212
 213        htcp_beta_update(ca, minRTT, maxRTT);
 214        htcp_alpha_update(ca);
 215
 216        /* add slowly fading memory for maxRTT to accommodate routing changes */
 217        if (minRTT > 0 && maxRTT > minRTT)
 218                ca->maxRTT = minRTT + ((maxRTT - minRTT) * 95) / 100;
 219}
 220
 221static u32 htcp_recalc_ssthresh(struct sock *sk)
 222{
 223        const struct tcp_sock *tp = tcp_sk(sk);
 224        const struct htcp *ca = inet_csk_ca(sk);
 225
 226        htcp_param_update(sk);
 227        return max((tp->snd_cwnd * ca->beta) >> 7, 2U);
 228}
 229
 230static void htcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
 231{
 232        struct tcp_sock *tp = tcp_sk(sk);
 233        struct htcp *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                /* In dangerous area, increase slowly.
 242                 * In theory this is tp->snd_cwnd += alpha / tp->snd_cwnd
 243                 */
 244                if ((tp->snd_cwnd_cnt * ca->alpha)>>7 >= tp->snd_cwnd) {
 245                        if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 246                                tp->snd_cwnd++;
 247                        tp->snd_cwnd_cnt = 0;
 248                        htcp_alpha_update(ca);
 249                } else
 250                        tp->snd_cwnd_cnt += ca->pkts_acked;
 251
 252                ca->pkts_acked = 1;
 253        }
 254}
 255
 256static void htcp_init(struct sock *sk)
 257{
 258        struct htcp *ca = inet_csk_ca(sk);
 259
 260        memset(ca, 0, sizeof(struct htcp));
 261        ca->alpha = ALPHA_BASE;
 262        ca->beta = BETA_MIN;
 263        ca->pkts_acked = 1;
 264        ca->last_cong = jiffies;
 265}
 266
 267static void htcp_state(struct sock *sk, u8 new_state)
 268{
 269        switch (new_state) {
 270        case TCP_CA_Open:
 271                {
 272                        struct htcp *ca = inet_csk_ca(sk);
 273                        if (ca->undo_last_cong) {
 274                                ca->last_cong = jiffies;
 275                                ca->undo_last_cong = 0;
 276                        }
 277                }
 278                break;
 279        case TCP_CA_CWR:
 280        case TCP_CA_Recovery:
 281        case TCP_CA_Loss:
 282                htcp_reset(inet_csk_ca(sk));
 283                break;
 284        }
 285}
 286
 287static struct tcp_congestion_ops htcp __read_mostly = {
 288        .init           = htcp_init,
 289        .ssthresh       = htcp_recalc_ssthresh,
 290        .cong_avoid     = htcp_cong_avoid,
 291        .set_state      = htcp_state,
 292        .undo_cwnd      = htcp_cwnd_undo,
 293        .pkts_acked     = measure_achieved_throughput,
 294        .owner          = THIS_MODULE,
 295        .name           = "htcp",
 296};
 297
 298static int __init htcp_register(void)
 299{
 300        BUILD_BUG_ON(sizeof(struct htcp) > ICSK_CA_PRIV_SIZE);
 301        BUILD_BUG_ON(BETA_MIN >= BETA_MAX);
 302        return tcp_register_congestion_control(&htcp);
 303}
 304
 305static void __exit htcp_unregister(void)
 306{
 307        tcp_unregister_congestion_control(&htcp);
 308}
 309
 310module_init(htcp_register);
 311module_exit(htcp_unregister);
 312
 313MODULE_AUTHOR("Baruch Even");
 314MODULE_LICENSE("GPL");
 315MODULE_DESCRIPTION("H-TCP");
 316
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.