net/ipv4/tcp_cong.c

   1 /*
   2  * Pluggable TCP congestion control support and newReno
   3  * congestion control.
   4  * Based on ideas from I/O scheduler support and Web100.
   5  *
   6  * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
   7  */
   8
   9 #define pr_fmt(fmt) "TCP: " fmt
  10
  11 #include <linux/module.h>
  12 #include <linux/mm.h>
  13 #include <linux/types.h>
  14 #include <linux/list.h>
  15 #include <linux/gfp.h>
  16 #include <linux/jhash.h>
  17 #include <net/tcp.h>
  18
  19 static DEFINE_SPINLOCK(tcp_cong_list_lock);
  20 static LIST_HEAD(tcp_cong_list);
  21
  22 /* Simple linear search, don't expect many entries! */
  23 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
  24 {
  25         struct tcp_congestion_ops *e;
  26
  27         list_for_each_entry_rcu(e, &tcp_cong_list, list) {
  28                 if (strcmp(e->name, name) == 0)
  29                         return e;
  30         }
  31
  32         return NULL;
  33 }
  34
  35 /* Must be called with rcu lock held */
  36 static const struct tcp_congestion_ops *__tcp_ca_find_autoload(const char *name)
  37 {
  38         const struct tcp_congestion_ops *ca = tcp_ca_find(name);
  39 #ifdef CONFIG_MODULES
  40         if (!ca && capable(CAP_NET_ADMIN)) {
  41                 rcu_read_unlock();
  42                 request_module("tcp_%s", name);
  43                 rcu_read_lock();
  44                 ca = tcp_ca_find(name);
  45         }
  46 #endif
  47         return ca;
  48 }
  49
  50 /* Simple linear search, not much in here. */
  51 struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
  52 {
  53         struct tcp_congestion_ops *e;
  54
  55         list_for_each_entry_rcu(e, &tcp_cong_list, list) {
  56                 if (e->key == key)
  57                         return e;
  58         }
  59
  60         return NULL;
  61 }
  62
  63 /*
  64  * Attach new congestion control algorithm to the list
  65  * of available options.
  66  */
  67 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
  68 {
  69         int ret = 0;
  70
  71         /* all algorithms must implement ssthresh and cong_avoid ops */
  72         if (!ca->ssthresh || !(ca->cong_avoid || ca->cong_control)) {
  73                 pr_err("%s does not implement required ops\n", ca->name);
  74                 return -EINVAL;
  75         }
  76
  77         ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
  78
  79         spin_lock(&tcp_cong_list_lock);
  80         if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
  81                 pr_notice("%s already registered or non-unique key\n",
  82                           ca->name);
  83                 ret = -EEXIST;
  84         } else {
  85                 list_add_tail_rcu(&ca->list, &tcp_cong_list);
  86                 pr_debug("%s registered\n", ca->name);
  87         }
  88         spin_unlock(&tcp_cong_list_lock);
  89
  90         return ret;
  91 }
  92 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
  93
  94 /*
  95  * Remove congestion control algorithm, called from
  96  * the module's remove function.  Module ref counts are used
  97  * to ensure that this can't be done till all sockets using
  98  * that method are closed.
  99  */
 100 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
 101 {
 102         spin_lock(&tcp_cong_list_lock);
 103         list_del_rcu(&ca->list);
 104         spin_unlock(&tcp_cong_list_lock);
 105
 106         /* Wait for outstanding readers to complete before the
 107          * module gets removed entirely.
 108          *
 109          * A try_module_get() should fail by now as our module is
 110          * in "going" state since no refs are held anymore and
 111          * module_exit() handler being called.
 112          */
 113         synchronize_rcu();
 114 }
 115 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
 116
 117 u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
 118 {
 119         const struct tcp_congestion_ops *ca;
 120         u32 key = TCP_CA_UNSPEC;
 121
 122         might_sleep();
 123
 124         rcu_read_lock();
 125         ca = __tcp_ca_find_autoload(name);
 126         if (ca) {
 127                 key = ca->key;
 128                 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
 129         }
 130         rcu_read_unlock();
 131
 132         return key;
 133 }
 134 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
 135
 136 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
 137 {
 138         const struct tcp_congestion_ops *ca;
 139         char *ret = NULL;
 140
 141         rcu_read_lock();
 142         ca = tcp_ca_find_key(key);
 143         if (ca)
 144                 ret = strncpy(buffer, ca->name,
 145                               TCP_CA_NAME_MAX);
 146         rcu_read_unlock();
 147
 148         return ret;
 149 }
 150 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
 151
 152 /* Assign choice of congestion control. */
 153 void tcp_assign_congestion_control(struct sock *sk)
 154 {
 155         struct inet_connection_sock *icsk = inet_csk(sk);
 156         struct tcp_congestion_ops *ca;
 157
 158         rcu_read_lock();
 159         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 160                 if (likely(try_module_get(ca->owner))) {
 161                         icsk->icsk_ca_ops = ca;
 162                         goto out;
 163                 }
 164                 /* Fallback to next available. The last really
 165                  * guaranteed fallback is Reno from this list.
 166                  */
 167         }
 168 out:
 169         rcu_read_unlock();
 170         memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
 171
 172         if (ca->flags & TCP_CONG_NEEDS_ECN)
 173                 INET_ECN_xmit(sk);
 174         else
 175                 INET_ECN_dontxmit(sk);
 176 }
 177
 178 void tcp_init_congestion_control(struct sock *sk)
 179 {
 180         const struct inet_connection_sock *icsk = inet_csk(sk);
 181
 182         tcp_sk(sk)->prior_ssthresh = 0;
 183         if (icsk->icsk_ca_ops->init)
 184                 icsk->icsk_ca_ops->init(sk);
 185         if (tcp_ca_needs_ecn(sk))
 186                 INET_ECN_xmit(sk);
 187         else
 188                 INET_ECN_dontxmit(sk);
 189 }
 190
 191 static void tcp_reinit_congestion_control(struct sock *sk,
 192                                           const struct tcp_congestion_ops *ca)
 193 {
 194         struct inet_connection_sock *icsk = inet_csk(sk);
 195
 196         tcp_cleanup_congestion_control(sk);
 197         icsk->icsk_ca_ops = ca;
 198         icsk->icsk_ca_setsockopt = 1;
 199         memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
 200
 201         if (sk->sk_state != TCP_CLOSE)
 202                 tcp_init_congestion_control(sk);
 203 }
 204
 205 /* Manage refcounts on socket close. */
 206 void tcp_cleanup_congestion_control(struct sock *sk)
 207 {
 208         struct inet_connection_sock *icsk = inet_csk(sk);
 209
 210         if (icsk->icsk_ca_ops->release)
 211                 icsk->icsk_ca_ops->release(sk);
 212         module_put(icsk->icsk_ca_ops->owner);
 213 }
 214
 215 /* Used by sysctl to change default congestion control */
 216 int tcp_set_default_congestion_control(const char *name)
 217 {
 218         struct tcp_congestion_ops *ca;
 219         int ret = -ENOENT;
 220
 221         spin_lock(&tcp_cong_list_lock);
 222         ca = tcp_ca_find(name);
 223 #ifdef CONFIG_MODULES
 224         if (!ca && capable(CAP_NET_ADMIN)) {
 225                 spin_unlock(&tcp_cong_list_lock);
 226
 227                 request_module("tcp_%s", name);
 228                 spin_lock(&tcp_cong_list_lock);
 229                 ca = tcp_ca_find(name);
 230         }
 231 #endif
 232
 233         if (ca) {
 234                 ca->flags |= TCP_CONG_NON_RESTRICTED;   /* default is always allowed */
 235                 list_move(&ca->list, &tcp_cong_list);
 236                 ret = 0;
 237         }
 238         spin_unlock(&tcp_cong_list_lock);
 239
 240         return ret;
 241 }
 242
 243 /* Set default value from kernel configuration at bootup */
 244 static int __init tcp_congestion_default(void)
 245 {
 246         return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 247 }
 248 late_initcall(tcp_congestion_default);
 249
 250 /* Build string with list of available congestion control values */
 251 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
 252 {
 253         struct tcp_congestion_ops *ca;
 254         size_t offs = 0;
 255
 256         rcu_read_lock();
 257         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 258                 offs += snprintf(buf + offs, maxlen - offs,
 259                                  "%s%s",
 260                                  offs == 0 ? "" : " ", ca->name);
 261         }
 262         rcu_read_unlock();
 263 }
 264
 265 /* Get current default congestion control */
 266 void tcp_get_default_congestion_control(char *name)
 267 {
 268         struct tcp_congestion_ops *ca;
 269         /* We will always have reno... */
 270         BUG_ON(list_empty(&tcp_cong_list));
 271
 272         rcu_read_lock();
 273         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 274         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 275         rcu_read_unlock();
 276 }
 277
 278 /* Built list of non-restricted congestion control values */
 279 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
 280 {
 281         struct tcp_congestion_ops *ca;
 282         size_t offs = 0;
 283
 284         *buf = '\0';
 285         rcu_read_lock();
 286         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 287                 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
 288                         continue;
 289                 offs += snprintf(buf + offs, maxlen - offs,
 290                                  "%s%s",
 291                                  offs == 0 ? "" : " ", ca->name);
 292         }
 293         rcu_read_unlock();
 294 }
 295
 296 /* Change list of non-restricted congestion control */
 297 int tcp_set_allowed_congestion_control(char *val)
 298 {
 299         struct tcp_congestion_ops *ca;
 300         char *saved_clone, *clone, *name;
 301         int ret = 0;
 302
 303         saved_clone = clone = kstrdup(val, GFP_USER);
 304         if (!clone)
 305                 return -ENOMEM;
 306
 307         spin_lock(&tcp_cong_list_lock);
 308         /* pass 1 check for bad entries */
 309         while ((name = strsep(&clone, " ")) && *name) {
 310                 ca = tcp_ca_find(name);
 311                 if (!ca) {
 312                         ret = -ENOENT;
 313                         goto out;
 314                 }
 315         }
 316
 317         /* pass 2 clear old values */
 318         list_for_each_entry_rcu(ca, &tcp_cong_list, list)
 319                 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
 320
 321         /* pass 3 mark as allowed */
 322         while ((name = strsep(&val, " ")) && *name) {
 323                 ca = tcp_ca_find(name);
 324                 WARN_ON(!ca);
 325                 if (ca)
 326                         ca->flags |= TCP_CONG_NON_RESTRICTED;
 327         }
 328 out:
 329         spin_unlock(&tcp_cong_list_lock);
 330         kfree(saved_clone);
 331
 332         return ret;
 333 }
 334
 335 /* Change congestion control for socket */
 336 int tcp_set_congestion_control(struct sock *sk, const char *name)
 337 {
 338         struct inet_connection_sock *icsk = inet_csk(sk);
 339         const struct tcp_congestion_ops *ca;
 340         int err = 0;
 341
 342         if (icsk->icsk_ca_dst_locked)
 343                 return -EPERM;
 344
 345         rcu_read_lock();
 346         ca = __tcp_ca_find_autoload(name);
 347         /* No change asking for existing value */
 348         if (ca == icsk->icsk_ca_ops) {
 349                 icsk->icsk_ca_setsockopt = 1;
 350                 goto out;
 351         }
 352         if (!ca)
 353                 err = -ENOENT;
 354         else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
 355                    ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
 356                 err = -EPERM;
 357         else if (!try_module_get(ca->owner))
 358                 err = -EBUSY;
 359         else
 360                 tcp_reinit_congestion_control(sk, ca);
 361  out:
 362         rcu_read_unlock();
 363         return err;
 364 }
 365
 366 /* Slow start is used when congestion window is no greater than the slow start
 367  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
 368  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
 369  * something better;) a packet is only considered (s)acked in its entirety to
 370  * defend the ACK attacks described in the RFC. Slow start processes a stretch
 371  * ACK of degree N as if N acks of degree 1 are received back to back except
 372  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
 373  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
 374  */
 375 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
 376 {
 377         u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
 378
 379         acked -= cwnd - tp->snd_cwnd;
 380         tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
 381
 382         return acked;
 383 }
 384 EXPORT_SYMBOL_GPL(tcp_slow_start);
 385
 386 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
 387  * for every packet that was ACKed.
 388  */
 389 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
 390 {
 391         /* If credits accumulated at a higher w, apply them gently now. */
 392         if (tp->snd_cwnd_cnt >= w) {
 393                 tp->snd_cwnd_cnt = 0;
 394                 tp->snd_cwnd++;
 395         }
 396
 397         tp->snd_cwnd_cnt += acked;
 398         if (tp->snd_cwnd_cnt >= w) {
 399                 u32 delta = tp->snd_cwnd_cnt / w;
 400
 401                 tp->snd_cwnd_cnt -= delta * w;
 402                 tp->snd_cwnd += delta;
 403         }
 404         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
 405 }
 406 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
 407
 408 /*
 409  * TCP Reno congestion control
 410  * This is special case used for fallback as well.
 411  */
 412 /* This is Jacobson's slow start and congestion avoidance.
 413  * SIGCOMM '88, p. 328.
 414  */
 415 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 416 {
 417         struct tcp_sock *tp = tcp_sk(sk);
 418
 419         if (!tcp_is_cwnd_limited(sk))
 420                 return;
 421
 422         /* In "safe" area, increase. */
 423         if (tcp_in_slow_start(tp)) {
 424                 acked = tcp_slow_start(tp, acked);
 425                 if (!acked)
 426                         return;
 427         }
 428         /* In dangerous area, increase slowly. */
 429         tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
 430 }
 431 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 432
 433 /* Slow start threshold is half the congestion window (min 2) */
 434 u32 tcp_reno_ssthresh(struct sock *sk)
 435 {
 436         const struct tcp_sock *tp = tcp_sk(sk);
 437
 438         return max(tp->snd_cwnd >> 1U, 2U);
 439 }
 440 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 441
 442 struct tcp_congestion_ops tcp_reno = {
 443         .flags          = TCP_CONG_NON_RESTRICTED,
 444         .name           = "reno",
 445         .owner          = THIS_MODULE,
 446         .ssthresh       = tcp_reno_ssthresh,
 447         .cong_avoid     = tcp_reno_cong_avoid,
 448 };