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) {
  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
 171         /* Clear out private data before diag gets it and
 172          * the ca has not been initialized.
 173          */
 174         if (ca->get_info)
 175                 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
 176         if (ca->flags & TCP_CONG_NEEDS_ECN)
 177                 INET_ECN_xmit(sk);
 178         else
 179                 INET_ECN_dontxmit(sk);
 180 }
 181
 182 void tcp_init_congestion_control(struct sock *sk)
 183 {
 184         const struct inet_connection_sock *icsk = inet_csk(sk);
 185
 186         tcp_sk(sk)->prior_ssthresh = 0;
 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                 tcp_init_congestion_control(sk);
 206 }
 207
 208 /* Manage refcounts on socket close. */
 209 void tcp_cleanup_congestion_control(struct sock *sk)
 210 {
 211         struct inet_connection_sock *icsk = inet_csk(sk);
 212
 213         if (icsk->icsk_ca_ops->release)
 214                 icsk->icsk_ca_ops->release(sk);
 215         module_put(icsk->icsk_ca_ops->owner);
 216 }
 217
 218 /* Used by sysctl to change default congestion control */
 219 int tcp_set_default_congestion_control(const char *name)
 220 {
 221         struct tcp_congestion_ops *ca;
 222         int ret = -ENOENT;
 223
 224         spin_lock(&tcp_cong_list_lock);
 225         ca = tcp_ca_find(name);
 226 #ifdef CONFIG_MODULES
 227         if (!ca && capable(CAP_NET_ADMIN)) {
 228                 spin_unlock(&tcp_cong_list_lock);
 229
 230                 request_module("tcp_%s", name);
 231                 spin_lock(&tcp_cong_list_lock);
 232                 ca = tcp_ca_find(name);
 233         }
 234 #endif
 235
 236         if (ca) {
 237                 ca->flags |= TCP_CONG_NON_RESTRICTED;   /* default is always allowed */
 238                 list_move(&ca->list, &tcp_cong_list);
 239                 ret = 0;
 240         }
 241         spin_unlock(&tcp_cong_list_lock);
 242
 243         return ret;
 244 }
 245
 246 /* Set default value from kernel configuration at bootup */
 247 static int __init tcp_congestion_default(void)
 248 {
 249         return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 250 }
 251 late_initcall(tcp_congestion_default);
 252
 253 /* Build string with list of available congestion control values */
 254 void tcp_get_available_congestion_control(char *buf, size_t maxlen)
 255 {
 256         struct tcp_congestion_ops *ca;
 257         size_t offs = 0;
 258
 259         rcu_read_lock();
 260         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 261                 offs += snprintf(buf + offs, maxlen - offs,
 262                                  "%s%s",
 263                                  offs == 0 ? "" : " ", ca->name);
 264         }
 265         rcu_read_unlock();
 266 }
 267
 268 /* Get current default congestion control */
 269 void tcp_get_default_congestion_control(char *name)
 270 {
 271         struct tcp_congestion_ops *ca;
 272         /* We will always have reno... */
 273         BUG_ON(list_empty(&tcp_cong_list));
 274
 275         rcu_read_lock();
 276         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 277         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 278         rcu_read_unlock();
 279 }
 280
 281 /* Built list of non-restricted congestion control values */
 282 void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
 283 {
 284         struct tcp_congestion_ops *ca;
 285         size_t offs = 0;
 286
 287         *buf = '\0';
 288         rcu_read_lock();
 289         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
 290                 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
 291                         continue;
 292                 offs += snprintf(buf + offs, maxlen - offs,
 293                                  "%s%s",
 294                                  offs == 0 ? "" : " ", ca->name);
 295         }
 296         rcu_read_unlock();
 297 }
 298
 299 /* Change list of non-restricted congestion control */
 300 int tcp_set_allowed_congestion_control(char *val)
 301 {
 302         struct tcp_congestion_ops *ca;
 303         char *saved_clone, *clone, *name;
 304         int ret = 0;
 305
 306         saved_clone = clone = kstrdup(val, GFP_USER);
 307         if (!clone)
 308                 return -ENOMEM;
 309
 310         spin_lock(&tcp_cong_list_lock);
 311         /* pass 1 check for bad entries */
 312         while ((name = strsep(&clone, " ")) && *name) {
 313                 ca = tcp_ca_find(name);
 314                 if (!ca) {
 315                         ret = -ENOENT;
 316                         goto out;
 317                 }
 318         }
 319
 320         /* pass 2 clear old values */
 321         list_for_each_entry_rcu(ca, &tcp_cong_list, list)
 322                 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
 323
 324         /* pass 3 mark as allowed */
 325         while ((name = strsep(&val, " ")) && *name) {
 326                 ca = tcp_ca_find(name);
 327                 WARN_ON(!ca);
 328                 if (ca)
 329                         ca->flags |= TCP_CONG_NON_RESTRICTED;
 330         }
 331 out:
 332         spin_unlock(&tcp_cong_list_lock);
 333         kfree(saved_clone);
 334
 335         return ret;
 336 }
 337
 338 /* Change congestion control for socket */
 339 int tcp_set_congestion_control(struct sock *sk, const char *name)
 340 {
 341         struct inet_connection_sock *icsk = inet_csk(sk);
 342         const struct tcp_congestion_ops *ca;
 343         int err = 0;
 344
 345         if (icsk->icsk_ca_dst_locked)
 346                 return -EPERM;
 347
 348         rcu_read_lock();
 349         ca = __tcp_ca_find_autoload(name);
 350         /* No change asking for existing value */
 351         if (ca == icsk->icsk_ca_ops) {
 352                 icsk->icsk_ca_setsockopt = 1;
 353                 goto out;
 354         }
 355         if (!ca)
 356                 err = -ENOENT;
 357         else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
 358                    ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
 359                 err = -EPERM;
 360         else if (!try_module_get(ca->owner))
 361                 err = -EBUSY;
 362         else
 363                 tcp_reinit_congestion_control(sk, ca);
 364  out:
 365         rcu_read_unlock();
 366         return err;
 367 }
 368
 369 /* Slow start is used when congestion window is no greater than the slow start
 370  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
 371  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
 372  * something better;) a packet is only considered (s)acked in its entirety to
 373  * defend the ACK attacks described in the RFC. Slow start processes a stretch
 374  * ACK of degree N as if N acks of degree 1 are received back to back except
 375  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
 376  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
 377  */
 378 u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
 379 {
 380         u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
 381
 382         acked -= cwnd - tp->snd_cwnd;
 383         tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
 384
 385         return acked;
 386 }
 387 EXPORT_SYMBOL_GPL(tcp_slow_start);
 388
 389 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
 390  * for every packet that was ACKed.
 391  */
 392 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
 393 {
 394         /* If credits accumulated at a higher w, apply them gently now. */
 395         if (tp->snd_cwnd_cnt >= w) {
 396                 tp->snd_cwnd_cnt = 0;
 397                 tp->snd_cwnd++;
 398         }
 399
 400         tp->snd_cwnd_cnt += acked;
 401         if (tp->snd_cwnd_cnt >= w) {
 402                 u32 delta = tp->snd_cwnd_cnt / w;
 403
 404                 tp->snd_cwnd_cnt -= delta * w;
 405                 tp->snd_cwnd += delta;
 406         }
 407         tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
 408 }
 409 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
 410
 411 /*
 412  * TCP Reno congestion control
 413  * This is special case used for fallback as well.
 414  */
 415 /* This is Jacobson's slow start and congestion avoidance.
 416  * SIGCOMM '88, p. 328.
 417  */
 418 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
 419 {
 420         struct tcp_sock *tp = tcp_sk(sk);
 421
 422         if (!tcp_is_cwnd_limited(sk))
 423                 return;
 424
 425         /* In "safe" area, increase. */
 426         if (tcp_in_slow_start(tp)) {
 427                 acked = tcp_slow_start(tp, acked);
 428                 if (!acked)
 429                         return;
 430         }
 431         /* In dangerous area, increase slowly. */
 432         tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
 433 }
 434 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 435
 436 /* Slow start threshold is half the congestion window (min 2) */
 437 u32 tcp_reno_ssthresh(struct sock *sk)
 438 {
 439         const struct tcp_sock *tp = tcp_sk(sk);
 440
 441         return max(tp->snd_cwnd >> 1U, 2U);
 442 }
 443 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 444
 445 struct tcp_congestion_ops tcp_reno = {
 446         .flags          = TCP_CONG_NON_RESTRICTED,
 447         .name           = "reno",
 448         .owner          = THIS_MODULE,
 449         .ssthresh       = tcp_reno_ssthresh,
 450         .cong_avoid     = tcp_reno_cong_avoid,
 451 };