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/config.h>
  10 #include <linux/module.h>
  11 #include <linux/mm.h>
  12 #include <linux/types.h>
  13 #include <linux/list.h>
  14 #include <net/tcp.h>
  15
  16 static DEFINE_SPINLOCK(tcp_cong_list_lock);
  17 static LIST_HEAD(tcp_cong_list);
  18
  19 /* Simple linear search, don't expect many entries! */
  20 static struct tcp_congestion_ops *tcp_ca_find(const char *name)
  21 {
  22         struct tcp_congestion_ops *e;
  23
  24         list_for_each_entry(e, &tcp_cong_list, list) {
  25                 if (strcmp(e->name, name) == 0)
  26                         return e;
  27         }
  28
  29         return NULL;
  30 }
  31
  32 /*
  33  * Attach new congestion control algorthim to the list
  34  * of available options.
  35  */
  36 int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
  37 {
  38         int ret = 0;
  39
  40         /* all algorithms must implement ssthresh and cong_avoid ops */
  41         if (!ca->ssthresh || !ca->cong_avoid || !ca->min_cwnd) {
  42                 printk(KERN_ERR "TCP %s does not implement required ops\n",
  43                        ca->name);
  44                 return -EINVAL;
  45         }
  46
  47         spin_lock(&tcp_cong_list_lock);
  48         if (tcp_ca_find(ca->name)) {
  49                 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
  50                 ret = -EEXIST;
  51         } else {
  52                 list_add_rcu(&ca->list, &tcp_cong_list);
  53                 printk(KERN_INFO "TCP %s registered\n", ca->name);
  54         }
  55         spin_unlock(&tcp_cong_list_lock);
  56
  57         return ret;
  58 }
  59 EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
  60
  61 /*
  62  * Remove congestion control algorithm, called from
  63  * the module's remove function.  Module ref counts are used
  64  * to ensure that this can't be done till all sockets using
  65  * that method are closed.
  66  */
  67 void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
  68 {
  69         spin_lock(&tcp_cong_list_lock);
  70         list_del_rcu(&ca->list);
  71         spin_unlock(&tcp_cong_list_lock);
  72 }
  73 EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
  74
  75 /* Assign choice of congestion control. */
  76 void tcp_init_congestion_control(struct tcp_sock *tp)
  77 {
  78         struct tcp_congestion_ops *ca;
  79
  80         rcu_read_lock();
  81         list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
  82                 if (try_module_get(ca->owner)) {
  83                         tp->ca_ops = ca;
  84                         break;
  85                 }
  86
  87         }
  88         rcu_read_unlock();
  89
  90         if (tp->ca_ops->init)
  91                 tp->ca_ops->init(tp);
  92 }
  93
  94 /* Manage refcounts on socket close. */
  95 void tcp_cleanup_congestion_control(struct tcp_sock *tp)
  96 {
  97         if (tp->ca_ops->release)
  98                 tp->ca_ops->release(tp);
  99         module_put(tp->ca_ops->owner);
 100 }
 101
 102 /* Used by sysctl to change default congestion control */
 103 int tcp_set_default_congestion_control(const char *name)
 104 {
 105         struct tcp_congestion_ops *ca;
 106         int ret = -ENOENT;
 107
 108         spin_lock(&tcp_cong_list_lock);
 109         ca = tcp_ca_find(name);
 110 #ifdef CONFIG_KMOD
 111         if (!ca) {
 112                 spin_unlock(&tcp_cong_list_lock);
 113
 114                 request_module("tcp_%s", name);
 115                 spin_lock(&tcp_cong_list_lock);
 116                 ca = tcp_ca_find(name);
 117         }
 118 #endif
 119
 120         if (ca) {
 121                 list_move(&ca->list, &tcp_cong_list);
 122                 ret = 0;
 123         }
 124         spin_unlock(&tcp_cong_list_lock);
 125
 126         return ret;
 127 }
 128
 129 /* Get current default congestion control */
 130 void tcp_get_default_congestion_control(char *name)
 131 {
 132         struct tcp_congestion_ops *ca;
 133         /* We will always have reno... */
 134         BUG_ON(list_empty(&tcp_cong_list));
 135
 136         rcu_read_lock();
 137         ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
 138         strncpy(name, ca->name, TCP_CA_NAME_MAX);
 139         rcu_read_unlock();
 140 }
 141
 142 /*
 143  * TCP Reno congestion control
 144  * This is special case used for fallback as well.
 145  */
 146 /* This is Jacobson's slow start and congestion avoidance.
 147  * SIGCOMM '88, p. 328.
 148  */
 149 void tcp_reno_cong_avoid(struct tcp_sock *tp, u32 ack, u32 rtt, u32 in_flight,
 150                          int flag)
 151 {
 152         if (in_flight < tp->snd_cwnd)
 153                 return;
 154
 155         if (tp->snd_cwnd <= tp->snd_ssthresh) {
 156                 /* In "safe" area, increase. */
 157                 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 158                         tp->snd_cwnd++;
 159         } else {
 160                 /* In dangerous area, increase slowly.
 161                  * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd
 162                  */
 163                 if (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
 164                         if (tp->snd_cwnd < tp->snd_cwnd_clamp)
 165                                 tp->snd_cwnd++;
 166                         tp->snd_cwnd_cnt = 0;
 167                 } else
 168                         tp->snd_cwnd_cnt++;
 169         }
 170 }
 171 EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
 172
 173 /* Slow start threshold is half the congestion window (min 2) */
 174 u32 tcp_reno_ssthresh(struct tcp_sock *tp)
 175 {
 176         return max(tp->snd_cwnd >> 1U, 2U);
 177 }
 178 EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
 179
 180 /* Lower bound on congestion window. */
 181 u32 tcp_reno_min_cwnd(struct tcp_sock *tp)
 182 {
 183         return tp->snd_ssthresh/2;
 184 }
 185 EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
 186
 187 struct tcp_congestion_ops tcp_reno = {
 188         .name           = "reno",
 189         .owner          = THIS_MODULE,
 190         .ssthresh       = tcp_reno_ssthresh,
 191         .cong_avoid     = tcp_reno_cong_avoid,
 192         .min_cwnd       = tcp_reno_min_cwnd,
 193 };
 194
 195 EXPORT_SYMBOL_GPL(tcp_reno);