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