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 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 <net/tcp.h>
  17
  18 static DEFINE_SPINLOCK(tcp_cong_list_lock);
  19 static LIST_HEAD(tcp_cong_list);
  20
  21 /* Simple linear search, don't expect many entries! */
  22 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
  23 {
  24         struct tcp_congestion_ops *e;
  25
  26         list_for_each_entry_rcu(e, &tcp_cong_list, list) {
  27                 if (strcmp(e->name, name) == 0)
  28                         return e;
  29         }
  30
  31         return NULL;
  32 }
  33
  34 /*
  35  * Attach new congestion control algorithm to the list
  36  * of available options.
  37  */
  38 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
  39 {
  40         int ret = 0;
  41
  42         /* all algorithms must implement ssthresh and cong_avoid ops */
  43         if (!ca->ssthresh || !ca->cong_avoid) {
  44                 pr_err("%s does not implement required ops\n", ca->name);
  45                 return -EINVAL;
  46         }
  47
  48         spin_lock(&tcp_cong_list_lock);
  49         if (tcp_ca_find(ca->name)) {
  50                 pr_notice("%s already registered\n", ca->name);
  51                 ret = -EEXIST;
  52         } else {
  53                 list_add_tail_rcu(&ca->list, &tcp_cong_list);
  54                 pr_info("%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_MODULES
 119         if (!ca && capable(CAP_NET_ADMIN)) {
 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_MODULES
 249         /* not found attempt to autoload module */
 250         if (!ca && capable(CAP_NET_ADMIN)) {
 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) ||
 261                    ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
 262                 err = -EPERM;
 263
 264         else if (!try_module_get(ca->owner))
 265                 err = -EBUSY;
 266
 267         else {
 268                 tcp_cleanup_congestion_control(sk);
 269                 icsk->icsk_ca_ops = ca;
 270
 271                 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
 272                         icsk->icsk_ca_ops->init(sk);
 273         }
 274  out:
 275         rcu_read_unlock();
 276         return err;
 277 }
 278
 279 /* RFC2861 Check whether we are limited by application or congestion window
 280  * This is the inverse of cwnd check in tcp_tso_should_defer
 281  */
 282 bool tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
 283 {
 284         const struct tcp_sock *tp = tcp_sk(sk);
 285         u32 left;
 286
 287         if (in_flight >= tp->snd_cwnd)
 288                 return true;
 289
 290         left = tp->snd_cwnd - in_flight;
 291         if (sk_can_gso(sk) &&
 292             left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
 293             left * tp->mss_cache < sk->sk_gso_max_size &&
 294             left < sk->sk_gso_max_segs)
 295                 return true;
 296         return left <= tcp_max_tso_deferred_mss(tp);
 297 }
 298 EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
 299
 300 /* Slow start is used when congestion window is no greater than the slow start
 301  * threshold. We base on RFC2581 and also handle stretch ACKs properly.
 302  * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
 303  * something better;) a packet is only considered (s)acked in its entirety to
 304  * defend the ACK attacks described in the RFC. Slow start processes a stretch
 305  * ACK of degree N as if N acks of degree 1 are received back to back except
 306  * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
 307  * returns the leftover acks to adjust cwnd in congestion avoidance mode.
 308  */
 309 int tcp_slow_start(struct tcp_sock *tp, u32 acked)
 310 {
 311         u32 cwnd = tp->snd_cwnd + acked;
 312
 313         if (cwnd > tp->snd_ssthresh)
 314                 cwnd = tp->snd_ssthresh + 1;
 315         acked -= cwnd - tp->snd_cwnd;
 316         tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
 317         return acked;
 318 }
 319 EXPORT_SYMBOL_GPL(tcp_slow_start);
 320
 321 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
 322 void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
 323 {
 324         if (tp->snd_cwnd_cnt >= w) {
 325                 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 326                         tp->snd_cwnd++;
 327                 tp->snd_cwnd_cnt = 0;
 328         } else {
 329                 tp->snd_cwnd_cnt++;
 330         }
 331 }
 332 EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
 333
 334 /*
 335  * TCP Reno congestion control
 336  * This is special case used for fallback as well.
 337  */
 338 /* This is Jacobson's slow start and congestion avoidance.
 339  * SIGCOMM '88, p. 328.
 340  */
 341 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 342 {
 343         struct tcp_sock *tp = tcp_sk(sk);
 344
 345         if (!tcp_is_cwnd_limited(sk, in_flight))
 346                 return;
 347
 348         /* In "safe" area, increase. */
 349         if (tp->snd_cwnd <= tp->snd_ssthresh)
 350                 tcp_slow_start(tp, acked);
 351         /* In dangerous area, increase slowly. */
 352         else
 353                 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
 354 }
 355 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 356
 357 /* Slow start threshold is half the congestion window (min 2) */
 358 u32 tcp_reno_ssthresh(struct sock *sk)
 359 {
 360         const struct tcp_sock *tp = tcp_sk(sk);
 361         return max(tp->snd_cwnd >> 1U, 2U);
 362 }
 363 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 364
 365 /* Lower bound on congestion window with halving. */
 366 u32 tcp_reno_min_cwnd(const struct sock *sk)
 367 {
 368         const struct tcp_sock *tp = tcp_sk(sk);
 369         return tp->snd_ssthresh/2;
 370 }
 371 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
 372
 373 struct tcp_congestion_ops tcp_reno = {
 374         .flags          = TCP_CONG_NON_RESTRICTED,
 375         .name           = "reno",
 376         .owner          = THIS_MODULE,
 377         .ssthresh       = tcp_reno_ssthresh,
 378         .cong_avoid     = tcp_reno_cong_avoid,
 379         .min_cwnd       = tcp_reno_min_cwnd,
 380 };
 381
 382 /* Initial congestion control used (until SYN)
 383  * really reno under another name so we can tell difference
 384  * during tcp_set_default_congestion_control
 385  */
 386 struct tcp_congestion_ops tcp_init_congestion_ops  = {
 387         .name           = "",
 388         .owner          = THIS_MODULE,
 389         .ssthresh       = tcp_reno_ssthresh,
 390         .cong_avoid     = tcp_reno_cong_avoid,
 391         .min_cwnd       = tcp_reno_min_cwnd,
 392 };
 393 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);