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 algorthim 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 (icsk->icsk_ca_ops != &tcp_init_congestion_ops)
  81                 return;
  82
  83         rcu_read_lock();
  84         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
  85                 if (try_module_get(ca->owner)) {
  86                         icsk->icsk_ca_ops = ca;
  87                         break;
  88                 }
  89
  90         }
  91         rcu_read_unlock();
  92
  93         if (icsk->icsk_ca_ops->init)
  94                 icsk->icsk_ca_ops->init(sk);
  95 }
  96
  97 /* Manage refcounts on socket close. */
  98 void tcp_cleanup_congestion_control(struct sock *sk)
  99 {
 100         struct inet_connection_sock *icsk = inet_csk(sk);
 101
 102         if (icsk->icsk_ca_ops->release)
 103                 icsk->icsk_ca_ops->release(sk);
 104         module_put(icsk->icsk_ca_ops->owner);
 105 }
 106
 107 /* Used by sysctl to change default congestion control */
 108 int tcp_set_default_congestion_control(const char *name)
 109 {
 110         struct tcp_congestion_ops *ca;
 111         int ret = -ENOENT;
 112
 113         spin_lock(&tcp_cong_list_lock);
 114         ca = tcp_ca_find(name);
 115 #ifdef CONFIG_KMOD
 116         if (!ca) {
 117                 spin_unlock(&tcp_cong_list_lock);
 118
 119                 request_module("tcp_%s", name);
 120                 spin_lock(&tcp_cong_list_lock);
 121                 ca = tcp_ca_find(name);
 122         }
 123 #endif
 124
 125         if (ca) {
 126                 list_move(&ca->list, &tcp_cong_list);
 127                 ret = 0;
 128         }
 129         spin_unlock(&tcp_cong_list_lock);
 130
 131         return ret;
 132 }
 133
 134 /* Set default value from kernel configuration at bootup */
 135 static int __init tcp_congestion_default(void)
 136 {
 137         return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
 138 }
 139 late_initcall(tcp_congestion_default);
 140
 141
 142 /* Get current default congestion control */
 143 void tcp_get_default_congestion_control(char *name)
 144 {
 145         struct tcp_congestion_ops *ca;
 146         /* We will always have reno... */
 147         BUG_ON(list_empty(&tcp_cong_list));
 148
 149         rcu_read_lock();
 150         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 151         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 152         rcu_read_unlock();
 153 }
 154
 155 /* Change congestion control for socket */
 156 int tcp_set_congestion_control(struct sock *sk, const char *name)
 157 {
 158         struct inet_connection_sock *icsk = inet_csk(sk);
 159         struct tcp_congestion_ops *ca;
 160         int err = 0;
 161
 162         rcu_read_lock();
 163         ca = tcp_ca_find(name);
 164         if (ca == icsk->icsk_ca_ops)
 165                 goto out;
 166
 167         if (!ca)
 168                 err = -ENOENT;
 169
 170         else if (!try_module_get(ca->owner))
 171                 err = -EBUSY;
 172
 173         else {
 174                 tcp_cleanup_congestion_control(sk);
 175                 icsk->icsk_ca_ops = ca;
 176                 if (icsk->icsk_ca_ops->init)
 177                         icsk->icsk_ca_ops->init(sk);
 178         }
 179  out:
 180         rcu_read_unlock();
 181         return err;
 182 }
 183
 184
 185 /*
 186  * Linear increase during slow start
 187  */
 188 void tcp_slow_start(struct tcp_sock *tp)
 189 {
 190         if (sysctl_tcp_abc) {
 191                 /* RFC3465: Slow Start
 192                  * TCP sender SHOULD increase cwnd by the number of
 193                  * previously unacknowledged bytes ACKed by each incoming
 194                  * acknowledgment, provided the increase is not more than L
 195                  */
 196                 if (tp->bytes_acked < tp->mss_cache)
 197                         return;
 198
 199                 /* We MAY increase by 2 if discovered delayed ack */
 200                 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache) {
 201                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 202                                 tp->snd_cwnd++;
 203                 }
 204         }
 205         tp->bytes_acked = 0;
 206
 207         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 208                 tp->snd_cwnd++;
 209 }
 210 EXPORT_SYMBOL_GPL(tcp_slow_start);
 211
 212 /*
 213  * TCP Reno congestion control
 214  * This is special case used for fallback as well.
 215  */
 216 /* This is Jacobson's slow start and congestion avoidance.
 217  * SIGCOMM '88, p. 328.
 218  */
 219 void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 rtt, u32 in_flight,
 220                          int flag)
 221 {
 222         struct tcp_sock *tp = tcp_sk(sk);
 223
 224         if (!tcp_is_cwnd_limited(sk, in_flight))
 225                 return;
 226
 227         /* In "safe" area, increase. */
 228         if (tp->snd_cwnd <= tp->snd_ssthresh)
 229                 tcp_slow_start(tp);
 230
 231         /* In dangerous area, increase slowly. */
 232         else if (sysctl_tcp_abc) {
 233                 /* RFC3465: Appropriate Byte Count
 234                  * increase once for each full cwnd acked
 235                  */
 236                 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
 237                         tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
 238                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 239                                 tp->snd_cwnd++;
 240                 }
 241         } else {
 242                 /* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd */
 243                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 244                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 245                                 tp->snd_cwnd++;
 246                         tp->snd_cwnd_cnt = 0;
 247                 } else
 248                         tp->snd_cwnd_cnt++;
 249         }
 250 }
 251 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 252
 253 /* Slow start threshold is half the congestion window (min 2) */
 254 u32 tcp_reno_ssthresh(struct sock *sk)
 255 {
 256         const struct tcp_sock *tp = tcp_sk(sk);
 257         return max(tp->snd_cwnd >> 1U, 2U);
 258 }
 259 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 260
 261 /* Lower bound on congestion window with halving. */
 262 u32 tcp_reno_min_cwnd(const struct sock *sk)
 263 {
 264         const struct tcp_sock *tp = tcp_sk(sk);
 265         return tp->snd_ssthresh/2;
 266 }
 267 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
 268
 269 struct tcp_congestion_ops tcp_reno = {
 270         .name           = "reno",
 271         .owner          = THIS_MODULE,
 272         .ssthresh       = tcp_reno_ssthresh,
 273         .cong_avoid     = tcp_reno_cong_avoid,
 274         .min_cwnd       = tcp_reno_min_cwnd,
 275 };
 276
 277 /* Initial congestion control used (until SYN)
 278  * really reno under another name so we can tell difference
 279  * during tcp_set_default_congestion_control
 280  */
 281 struct tcp_congestion_ops tcp_init_congestion_ops  = {
 282         .name           = "",
 283         .owner          = THIS_MODULE,
 284         .ssthresh       = tcp_reno_ssthresh,
 285         .cong_avoid     = tcp_reno_cong_avoid,
 286         .min_cwnd       = tcp_reno_min_cwnd,
 287 };
 288 EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);