release/src-rt/linux/linux-2.6/net/ipv4/tcp_cong.c

   1 /*
   2  * Plugable TCP congestion control support and newReno
   3  * congestion control.
   4  * Based on ideas from I/O scheduler suport and Web100.
   5  *
   6  * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
   7  */
   8
   9 #include <linux/module.h>
  10 #include <linux/mm.h>
  11 #include <linux/types.h>
  12 #include <linux/list.h>
  13 #include <net/tcp.h>
  14
  15 int sysctl_tcp_max_ssthresh = 0;
  16
  17 static DEFINE_SPINLOCK(tcp_cong_list_lock);
  18 static LIST_HEAD(tcp_cong_list);
  19
  20 /* Simple linear search, don't expect many entries! */
  21 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
  22 {
  23         struct tcp_congestion_ops *e;
  24
  25         list_for_each_entry_rcu(e, &tcp_cong_list, list) {
  26                 if (strcmp(e->name, name) == 0)
  27                         return e;
  28         }
  29
  30         return NULL;
  31 }
  32
  33 /*
  34  * Attach new congestion control algorithm to the list
  35  * of available options.
  36  */
  37 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
  38 {
  39         int ret = 0;
  40
  41         /* all algorithms must implement ssthresh and cong_avoid ops */
  42         if (!ca->ssthresh || !ca->cong_avoid) {
  43                 printk(KERN_ERR "TCP %s does not implement required ops\n",
  44                        ca->name);
  45                 return -EINVAL;
  46         }
  47
  48         spin_lock(&tcp_cong_list_lock);
  49         if (tcp_ca_find(ca->name)) {
  50                 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
  51                 ret = -EEXIST;
  52         } else {
  53                 list_add_tail_rcu(&ca->list, &tcp_cong_list);
  54                 printk(KERN_INFO "TCP %s registered\n", ca->name);
  55         }
  56         spin_unlock(&tcp_cong_list_lock);
  57
  58         return ret;
  59 }
  60 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
  61
  62 /*
  63  * Remove congestion control algorithm, called from
  64  * the module's remove function.  Module ref counts are used
  65  * to ensure that this can't be done till all sockets using
  66  * that method are closed.
  67  */
  68 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
  69 {
  70         spin_lock(&tcp_cong_list_lock);
  71         list_del_rcu(&ca->list);
  72         spin_unlock(&tcp_cong_list_lock);
  73 }
  74 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
  75
  76 /* Assign choice of congestion control. */
  77 void tcp_init_congestion_control(struct sock *sk)
  78 {
  79         struct inet_connection_sock *icsk = inet_csk(sk);
  80         struct tcp_congestion_ops *ca;
  81
  82         /* if no choice made yet assign the current value set as default */
  83         if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
  84                 rcu_read_lock();
  85                 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
  86                         if (try_module_get(ca->owner)) {
  87                                 icsk->icsk_ca_ops = ca;
  88                                 break;
  89                         }
  90
  91                         /* fallback to next available */
  92                 }
  93                 rcu_read_unlock();
  94         }
  95
  96         if (icsk->icsk_ca_ops->init)
  97                 icsk->icsk_ca_ops->init(sk);
  98 }
  99
 100 /* Manage refcounts on socket close. */
 101 void tcp_cleanup_congestion_control(struct sock *sk)
 102 {
 103         struct inet_connection_sock *icsk = inet_csk(sk);
 104
 105         if (icsk->icsk_ca_ops->release)
 106                 icsk->icsk_ca_ops->release(sk);
 107         module_put(icsk->icsk_ca_ops->owner);
 108 }
 109
 110 /* Used by sysctl to change default congestion control */
 111 int tcp_set_default_congestion_control(const char *name)
 112 {
 113         struct tcp_congestion_ops *ca;
 114         int ret = -ENOENT;
 115
 116         spin_lock(&tcp_cong_list_lock);
 117         ca = tcp_ca_find(name);
 118 #ifdef CONFIG_KMOD
 119         if (!ca && capable(CAP_SYS_MODULE)) {
 120                 spin_unlock(&tcp_cong_list_lock);
 121
 122                 request_module("tcp_%s", name);
 123                 spin_lock(&tcp_cong_list_lock);
 124                 ca = tcp_ca_find(name);
 125         }
 126 #endif
 127
 128         if (ca) {
 129                 ca->flags |= TCP_CONG_NON_RESTRICTED;   /* default is always allowed */
 130                 list_move(&ca->list, &tcp_cong_list);
 131                 ret = 0;
 132         }
 133         spin_unlock(&tcp_cong_list_lock);
 134
 135         return ret;
 136 }
 137
 138 /* Set default value from kernel configuration at bootup */
 139 static int __init tcp_congestion_default(void)
 140 {
 141         return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 142 }
 143 late_initcall(tcp_congestion_default);
 144
 145
 146 /* Build string with list of available congestion control values */
 147 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
 148 {
 149         struct tcp_congestion_ops *ca;
 150         size_t offs = 0;
 151
 152         rcu_read_lock();
 153         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 154                 offs += snprintf(buf + offs, maxlen - offs,
 155                                  "%s%s",
 156                                  offs == 0 ? "" : " ", ca->name);
 157
 158         }
 159         rcu_read_unlock();
 160 }
 161
 162 /* Get current default congestion control */
 163 void tcp_get_default_congestion_control(char *name)
 164 {
 165         struct tcp_congestion_ops *ca;
 166         /* We will always have reno... */
 167         BUG_ON(list_empty(&tcp_cong_list));
 168
 169         rcu_read_lock();
 170         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 171         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 172         rcu_read_unlock();
 173 }
 174
 175 /* Built list of non-restricted congestion control values */
 176 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
 177 {
 178         struct tcp_congestion_ops *ca;
 179         size_t offs = 0;
 180
 181         *buf = '\0';
 182         rcu_read_lock();
 183         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 184                 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
 185                         continue;
 186                 offs += snprintf(buf + offs, maxlen - offs,
 187                                  "%s%s",
 188                                  offs == 0 ? "" : " ", ca->name);
 189
 190         }
 191         rcu_read_unlock();
 192 }
 193
 194 /* Change list of non-restricted congestion control */
 195 int tcp_set_allowed_congestion_control(char *val)
 196 {
 197         struct tcp_congestion_ops *ca;
 198         char *saved_clone, *clone, *name;
 199         int ret = 0;
 200
 201         saved_clone = clone = kstrdup(val, GFP_USER);
 202         if (!clone)
 203                 return -ENOMEM;
 204
 205         spin_lock(&tcp_cong_list_lock);
 206         /* pass 1 check for bad entries */
 207         while ((name = strsep(&clone, " ")) && *name) {
 208                 ca = tcp_ca_find(name);
 209                 if (!ca) {
 210                         ret = -ENOENT;
 211                         goto out;
 212                 }
 213         }
 214
 215         /* pass 2 clear old values */
 216         list_for_each_entry_rcu(ca, &tcp_cong_list, list)
 217                 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
 218
 219         /* pass 3 mark as allowed */
 220         while ((name = strsep(&val, " ")) && *name) {
 221                 ca = tcp_ca_find(name);
 222                 WARN_ON(!ca);
 223                 if (ca)
 224                         ca->flags |= TCP_CONG_NON_RESTRICTED;
 225         }
 226 out:
 227         spin_unlock(&tcp_cong_list_lock);
 228         kfree(saved_clone);
 229
 230         return ret;
 231 }
 232
 233
 234 /* Change congestion control for socket */
 235 int tcp_set_congestion_control(struct sock *sk, const char *name)
 236 {
 237         struct inet_connection_sock *icsk = inet_csk(sk);
 238         struct tcp_congestion_ops *ca;
 239         int err = 0;
 240
 241         rcu_read_lock();
 242         ca = tcp_ca_find(name);
 243
 244         /* no change asking for existing value */
 245         if (ca == icsk->icsk_ca_ops)
 246                 goto out;
 247
 248 #ifdef CONFIG_KMOD
 249         /* not found attempt to autoload module */
 250         if (!ca && capable(CAP_SYS_MODULE)) {
 251                 rcu_read_unlock();
 252                 request_module("tcp_%s", name);
 253                 rcu_read_lock();
 254                 ca = tcp_ca_find(name);
 255         }
 256 #endif
 257         if (!ca)
 258                 err = -ENOENT;
 259
 260         else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
 261                 err = -EPERM;
 262
 263         else if (!try_module_get(ca->owner))
 264                 err = -EBUSY;
 265
 266         else {
 267                 tcp_cleanup_congestion_control(sk);
 268                 icsk->icsk_ca_ops = ca;
 269
 270                 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
 271                         icsk->icsk_ca_ops->init(sk);
 272         }
 273  out:
 274         rcu_read_unlock();
 275         return err;
 276 }
 277
 278
 279 /*
 280  * Slow start is used when congestion window is less than slow start
 281  * threshold. This version implements the basic RFC2581 version
 282  * and optionally supports:
 283  *      RFC3742 Limited Slow Start        - growth limited to max_ssthresh
 284  *      RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
 285  */
 286 void tcp_slow_start(struct tcp_sock *tp)
 287 {
 288         int cnt; /* increase in packets */
 289
 290         /* RFC3465: ABC Slow start
 291          * Increase only after a full MSS of bytes is acked
 292          *
 293          * TCP sender SHOULD increase cwnd by the number of
 294          * previously unacknowledged bytes ACKed by each incoming
 295          * acknowledgment, provided the increase is not more than L
 296          */
 297         if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
 298                 return;
 299
 300         if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
 301                 cnt = sysctl_tcp_max_ssthresh >> 1;     /* limited slow start */
 302         else
 303                 cnt = tp->snd_cwnd;                     /* exponential increase */
 304
 305         /* RFC3465: ABC
 306          * We MAY increase by 2 if discovered delayed ack
 307          */
 308         if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
 309                 cnt <<= 1;
 310         tp->bytes_acked = 0;
 311
 312         tp->snd_cwnd_cnt += cnt;
 313         while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 314                 tp->snd_cwnd_cnt -= tp->snd_cwnd;
 315                 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 316                         tp->snd_cwnd++;
 317         }
 318 }
 319 EXPORT_SYMBOL_GPL(tcp_slow_start);
 320
 321 /*
 322  * TCP Reno congestion control
 323  * This is special case used for fallback as well.
 324  */
 325 /* This is Jacobson's slow start and congestion avoidance.
 326  * SIGCOMM '88, p. 328.
 327  */
 328 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
 329                          int flag)
 330 {
 331         struct tcp_sock *tp = tcp_sk(sk);
 332
 333         if (!tcp_is_cwnd_limited(sk, in_flight))
 334                 return;
 335
 336         /* In "safe" area, increase. */
 337         if (tp->snd_cwnd <= tp->snd_ssthresh)
 338                 tcp_slow_start(tp);
 339
 340         /* In dangerous area, increase slowly. */
 341         else if (sysctl_tcp_abc) {
 342                 /* RFC3465: Appropriate Byte Count
 343                  * increase once for each full cwnd acked
 344                  */
 345                 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
 346                         tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
 347                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 348                                 tp->snd_cwnd++;
 349                 }
 350         } else {
 351                 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
 352                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 353                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 354                                 tp->snd_cwnd++;
 355                         tp->snd_cwnd_cnt = 0;
 356                 } else
 357                         tp->snd_cwnd_cnt++;
 358         }
 359 }
 360 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 361
 362 /* Slow start threshold is half the congestion window (min 2) */
 363 u32 tcp_reno_ssthresh(struct sock *sk)
 364 {
 365         const struct tcp_sock *tp = tcp_sk(sk);
 366         return max(tp->snd_cwnd >> 1U, 2U);
 367 }
 368 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 369
 370 /* Lower bound on congestion window with halving. */
 371 u32 tcp_reno_min_cwnd(const struct sock *sk)
 372 {
 373         const struct tcp_sock *tp = tcp_sk(sk);
 374         return tp->snd_ssthresh/2;
 375 }
 376 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
 377
 378 struct tcp_congestion_ops tcp_reno = {
 379         .flags          = TCP_CONG_NON_RESTRICTED,
 380         .name           = "reno",
 381         .owner          = THIS_MODULE,
 382         .ssthresh       = tcp_reno_ssthresh,
 383         .cong_avoid     = tcp_reno_cong_avoid,
 384         .min_cwnd       = tcp_reno_min_cwnd,
 385 };
 386
 387 /* Initial congestion control used (until SYN)
 388  * really reno under another name so we can tell difference
 389  * during tcp_set_default_congestion_control
 390  */
 391 struct tcp_congestion_ops tcp_init_congestion_ops  = {
 392         .name           = "",
 393         .owner          = THIS_MODULE,
 394         .ssthresh       = tcp_reno_ssthresh,
 395         .cong_avoid     = tcp_reno_cong_avoid,
 396         .min_cwnd       = tcp_reno_min_cwnd,
 397 };
 398 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);