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 these */
  72         if (!ca->ssthresh || !ca->undo_cwnd ||
  73             !(ca->cong_avoid || ca->cong_control)) {
  74                 pr_err("%s does not implement required ops\n", ca->name);
  75                 return -EINVAL;
  76         }
  77
  78         ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
  79
  80         spin_lock(&tcp_cong_list_lock);
  81         if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
  82                 pr_notice("%s already registered or non-unique key\n",
  83                           ca->name);
  84                 ret = -EEXIST;
  85         } else {
  86                 list_add_tail_rcu(&ca->list, &tcp_cong_list);
  87                 pr_debug("%s registered\n", ca->name);
  88         }
  89         spin_unlock(&tcp_cong_list_lock);
  90
  91         return ret;
  92 }
  93 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
  94
  95 /*
  96  * Remove congestion control algorithm, called from
  97  * the module's remove function.  Module ref counts are used
  98  * to ensure that this can't be done till all sockets using
  99  * that method are closed.
 100  */
 101 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
 102 {
 103         spin_lock(&tcp_cong_list_lock);
 104         list_del_rcu(&ca->list);
 105         spin_unlock(&tcp_cong_list_lock);
 106
 107         /* Wait for outstanding readers to complete before the
 108          * module gets removed entirely.
 109          *
 110          * A try_module_get() should fail by now as our module is
 111          * in "going" state since no refs are held anymore and
 112          * module_exit() handler being called.
 113          */
 114         synchronize_rcu();
 115 }
 116 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
 117
 118 u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
 119 {
 120         const struct tcp_congestion_ops *ca;
 121         u32 key = TCP_CA_UNSPEC;
 122
 123         might_sleep();
 124
 125         rcu_read_lock();
 126         ca = __tcp_ca_find_autoload(name);
 127         if (ca) {
 128                 key = ca->key;
 129                 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
 130         }
 131         rcu_read_unlock();
 132
 133         return key;
 134 }
 135 EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
 136
 137 char *tcp_ca_get_name_by_key(u32 key, char *buffer)
 138 {
 139         const struct tcp_congestion_ops *ca;
 140         char *ret = NULL;
 141
 142         rcu_read_lock();
 143         ca = tcp_ca_find_key(key);
 144         if (ca)
 145                 ret = strncpy(buffer, ca->name,
 146                               TCP_CA_NAME_MAX);
 147         rcu_read_unlock();
 148
 149         return ret;
 150 }
 151 EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
 152
 153 /* Assign choice of congestion control. */
 154 void tcp_assign_congestion_control(struct sock *sk)
 155 {
 156         struct inet_connection_sock *icsk = inet_csk(sk);
 157         struct tcp_congestion_ops *ca;
 158
 159         rcu_read_lock();
 160         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 161                 if (likely(try_module_get(ca->owner))) {
 162                         icsk->icsk_ca_ops = ca;
 163                         goto out;
 164                 }
 165                 /* Fallback to next available. The last really
 166                  * guaranteed fallback is Reno from this list.
 167                  */
 168         }
 169 out:
 170         rcu_read_unlock();
 171
 172         /* Clear out private data before diag gets it and
 173          * the ca has not been initialized.
 174          */
 175         if (ca->get_info)
 176                 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
 177         if (ca->flags & TCP_CONG_NEEDS_ECN)
 178                 INET_ECN_xmit(sk);
 179         else
 180                 INET_ECN_dontxmit(sk);
 181 }
 182
 183 void tcp_init_congestion_control(struct sock *sk)
 184 {
 185         const struct inet_connection_sock *icsk = inet_csk(sk);
 186
 187         if (icsk->icsk_ca_ops->init)
 188                 icsk->icsk_ca_ops->init(sk);
 189         if (tcp_ca_needs_ecn(sk))
 190                 INET_ECN_xmit(sk);
 191         else
 192                 INET_ECN_dontxmit(sk);
 193 }
 194
 195 static void tcp_reinit_congestion_control(struct sock *sk,
 196                                           const struct tcp_congestion_ops *ca)
 197 {
 198         struct inet_connection_sock *icsk = inet_csk(sk);
 199
 200         tcp_cleanup_congestion_control(sk);
 201         icsk->icsk_ca_ops = ca;
 202         icsk->icsk_ca_setsockopt = 1;
 203
 204         if (sk->sk_state != TCP_CLOSE) {
 205                 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
 206                 tcp_init_congestion_control(sk);
 207         }
 208 }
 209
 210 /* Manage refcounts on socket close. */
 211 void tcp_cleanup_congestion_control(struct sock *sk)
 212 {
 213         struct inet_connection_sock *icsk = inet_csk(sk);
 214
 215         if (icsk->icsk_ca_ops->release)
 216                 icsk->icsk_ca_ops->release(sk);
 217         module_put(icsk->icsk_ca_ops->owner);
 218 }
 219
 220 /* Used by sysctl to change default congestion control */
 221 int tcp_set_default_congestion_control(const char *name)
 222 {
 223         struct tcp_congestion_ops *ca;
 224         int ret = -ENOENT;
 225
 226         spin_lock(&tcp_cong_list_lock);
 227         ca = tcp_ca_find(name);
 228 #ifdef CONFIG_MODULES
 229         if (!ca && capable(CAP_NET_ADMIN)) {
 230                 spin_unlock(&tcp_cong_list_lock);
 231
 232                 request_module("tcp_%s", name);
 233                 spin_lock(&tcp_cong_list_lock);
 234                 ca = tcp_ca_find(name);
 235         }
 236 #endif
 237
 238         if (ca) {
 239                 ca->flags |= TCP_CONG_NON_RESTRICTED;   /* default is always allowed */
 240                 list_move(&ca->list, &tcp_cong_list);
 241                 ret = 0;
 242         }
 243         spin_unlock(&tcp_cong_list_lock);
 244
 245         return ret;
 246 }
 247
 248 /* Set default value from kernel configuration at bootup */
 249 static int __init tcp_congestion_default(void)
 250 {
 251         return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 252 }
 253 late_initcall(tcp_congestion_default);
 254
 255 /* Build string with list of available congestion control values */
 256 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
 257 {
 258         struct tcp_congestion_ops *ca;
 259         size_t offs = 0;
 260
 261         rcu_read_lock();
 262         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 263                 offs += snprintf(buf + offs, maxlen - offs,
 264                                  "%s%s",
 265                                  offs == 0 ? "" : " ", ca->name);
 266         }
 267         rcu_read_unlock();
 268 }
 269
 270 /* Get current default congestion control */
 271 void tcp_get_default_congestion_control(char *name)
 272 {
 273         struct tcp_congestion_ops *ca;
 274         /* We will always have reno... */
 275         BUG_ON(list_empty(&tcp_cong_list));
 276
 277         rcu_read_lock();
 278         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 279         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 280         rcu_read_unlock();
 281 }
 282
 283 /* Built list of non-restricted congestion control values */
 284 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
 285 {
 286         struct tcp_congestion_ops *ca;
 287         size_t offs = 0;
 288
 289         *buf = '\0';
 290         rcu_read_lock();
 291         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 292                 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
 293                         continue;
 294                 offs += snprintf(buf + offs, maxlen - offs,
 295                                  "%s%s",
 296                                  offs == 0 ? "" : " ", ca->name);
 297         }
 298         rcu_read_unlock();
 299 }
 300
 301 /* Change list of non-restricted congestion control */
 302 int tcp_set_allowed_congestion_control(char *val)
 303 {
 304         struct tcp_congestion_ops *ca;
 305         char *saved_clone, *clone, *name;
 306         int ret = 0;
 307
 308         saved_clone = clone = kstrdup(val, GFP_USER);
 309         if (!clone)
 310                 return -ENOMEM;
 311
 312         spin_lock(&tcp_cong_list_lock);
 313         /* pass 1 check for bad entries */
 314         while ((name = strsep(&clone, " ")) && *name) {
 315                 ca = tcp_ca_find(name);
 316                 if (!ca) {
 317                         ret = -ENOENT;
 318                         goto out;
 319                 }
 320         }
 321
 322         /* pass 2 clear old values */
 323         list_for_each_entry_rcu(ca, &tcp_cong_list, list)
 324                 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
 325
 326         /* pass 3 mark as allowed */
 327         while ((name = strsep(&val, " ")) && *name) {
 328                 ca = tcp_ca_find(name);
 329                 WARN_ON(!ca);
 330                 if (ca)
 331                         ca->flags |= TCP_CONG_NON_RESTRICTED;
 332         }
 333 out:
 334         spin_unlock(&tcp_cong_list_lock);
 335         kfree(saved_clone);
 336
 337         return ret;
 338 }
 339
 340 /* Change congestion control for socket */
 341 int tcp_set_congestion_control(struct sock *sk, const char *name)
 342 {
 343         struct inet_connection_sock *icsk = inet_csk(sk);
 344         const struct tcp_congestion_ops *ca;
 345         int err = 0;
 346
 347         if (icsk->icsk_ca_dst_locked)
 348                 return -EPERM;
 349
 350         rcu_read_lock();
 351         ca = __tcp_ca_find_autoload(name);
 352         /* No change asking for existing value */
 353         if (ca == icsk->icsk_ca_ops) {
 354                 icsk->icsk_ca_setsockopt = 1;
 355                 goto out;
 356         }
 357         if (!ca)
 358                 err = -ENOENT;
 359         else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
 360                    ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
 361                 err = -EPERM;
 362         else if (!try_module_get(ca->owner))
 363                 err = -EBUSY;
 364         else
 365                 tcp_reinit_congestion_control(sk, ca);
 366  out:
 367         rcu_read_unlock();
 368         return err;
 369 }
 370
 371 /* Slow start is used when congestion window is no greater than the slow start
 372  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
 373  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
 374  * something better;) a packet is only considered (s)acked in its entirety to
 375  * defend the ACK attacks described in the RFC. Slow start processes a stretch
 376  * ACK of degree N as if N acks of degree 1 are received back to back except
 377  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
 378  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
 379  */
 380 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
 381 {
 382         u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
 383
 384         acked -= cwnd - tp->snd_cwnd;
 385         tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
 386
 387         return acked;
 388 }
 389 EXPORT_SYMBOL_GPL(tcp_slow_start);
 390
 391 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
 392  * for every packet that was ACKed.
 393  */
 394 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
 395 {
 396         /* If credits accumulated at a higher w, apply them gently now. */
 397         if (tp->snd_cwnd_cnt >= w) {
 398                 tp->snd_cwnd_cnt = 0;
 399                 tp->snd_cwnd++;
 400         }
 401
 402         tp->snd_cwnd_cnt += acked;
 403         if (tp->snd_cwnd_cnt >= w) {
 404                 u32 delta = tp->snd_cwnd_cnt / w;
 405
 406                 tp->snd_cwnd_cnt -= delta * w;
 407                 tp->snd_cwnd += delta;
 408         }
 409         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
 410 }
 411 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
 412
 413 /*
 414  * TCP Reno congestion control
 415  * This is special case used for fallback as well.
 416  */
 417 /* This is Jacobson's slow start and congestion avoidance.
 418  * SIGCOMM '88, p. 328.
 419  */
 420 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 421 {
 422         struct tcp_sock *tp = tcp_sk(sk);
 423
 424         if (!tcp_is_cwnd_limited(sk))
 425                 return;
 426
 427         /* In "safe" area, increase. */
 428         if (tcp_in_slow_start(tp)) {
 429                 acked = tcp_slow_start(tp, acked);
 430                 if (!acked)
 431                         return;
 432         }
 433         /* In dangerous area, increase slowly. */
 434         tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
 435 }
 436 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 437
 438 /* Slow start threshold is half the congestion window (min 2) */
 439 u32 tcp_reno_ssthresh(struct sock *sk)
 440 {
 441         const struct tcp_sock *tp = tcp_sk(sk);
 442
 443         return max(tp->snd_cwnd >> 1U, 2U);
 444 }
 445 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 446
 447 u32 tcp_reno_undo_cwnd(struct sock *sk)
 448 {
 449         const struct tcp_sock *tp = tcp_sk(sk);
 450
 451         return max(tp->snd_cwnd, tp->snd_ssthresh << 1);
 452 }
 453 EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
 454
 455 struct tcp_congestion_ops tcp_reno = {
 456         .flags          = TCP_CONG_NON_RESTRICTED,
 457         .name           = "reno",
 458         .owner          = THIS_MODULE,
 459         .ssthresh       = tcp_reno_ssthresh,
 460         .cong_avoid     = tcp_reno_cong_avoid,
 461         .undo_cwnd      = tcp_reno_undo_cwnd,
 462 };