net/core/link_watch.c

   1 /*
   2  * Linux network device link state notification
   3  *
   4  * Author:
   5  *     Stefan Rompf <sux@loplof.de>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; either version
  10  * 2 of the License, or (at your option) any later version.
  11  *
  12  */
  13
  14 #include <linux/module.h>
  15 #include <linux/netdevice.h>
  16 #include <linux/if.h>
  17 #include <net/sock.h>
  18 #include <net/pkt_sched.h>
  19 #include <linux/rtnetlink.h>
  20 #include <linux/jiffies.h>
  21 #include <linux/spinlock.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/bitops.h>
  24 #include <linux/types.h>
  25
  26
  27 enum lw_bits {
  28         LW_URGENT = 0,
  29 };
  30
  31 static unsigned long linkwatch_flags;
  32 static unsigned long linkwatch_nextevent;
  33
  34 static void linkwatch_event(struct work_struct *dummy);
  35 static DECLARE_DELAYED_WORK(linkwatch_work, linkwatch_event);
  36
  37 static LIST_HEAD(lweventlist);
  38 static DEFINE_SPINLOCK(lweventlist_lock);
  39
  40 static unsigned char default_operstate(const struct net_device *dev)
  41 {
  42         if (!netif_carrier_ok(dev))
  43                 return (dev->ifindex != dev_get_iflink(dev) ?
  44                         IF_OPER_LOWERLAYERDOWN : IF_OPER_DOWN);
  45
  46         if (netif_dormant(dev))
  47                 return IF_OPER_DORMANT;
  48
  49         return IF_OPER_UP;
  50 }
  51
  52
  53 static void rfc2863_policy(struct net_device *dev)
  54 {
  55         unsigned char operstate = default_operstate(dev);
  56
  57         if (operstate == dev->operstate)
  58                 return;
  59
  60         write_lock_bh(&dev_base_lock);
  61
  62         switch(dev->link_mode) {
  63         case IF_LINK_MODE_DORMANT:
  64                 if (operstate == IF_OPER_UP)
  65                         operstate = IF_OPER_DORMANT;
  66                 break;
  67
  68         case IF_LINK_MODE_DEFAULT:
  69         default:
  70                 break;
  71         }
  72
  73         dev->operstate = operstate;
  74
  75         write_unlock_bh(&dev_base_lock);
  76 }
  77
  78
  79 void linkwatch_init_dev(struct net_device *dev)
  80 {
  81         /* Handle pre-registration link state changes */
  82         if (!netif_carrier_ok(dev) || netif_dormant(dev))
  83                 rfc2863_policy(dev);
  84 }
  85
  86
  87 static bool linkwatch_urgent_event(struct net_device *dev)
  88 {
  89         if (!netif_running(dev))
  90                 return false;
  91
  92         if (dev->ifindex != dev_get_iflink(dev))
  93                 return true;
  94
  95         if (netif_is_lag_port(dev) || netif_is_lag_master(dev))
  96                 return true;
  97
  98         return netif_carrier_ok(dev) && qdisc_tx_changing(dev);
  99 }
 100
 101
 102 static void linkwatch_add_event(struct net_device *dev)
 103 {
 104         unsigned long flags;
 105
 106         spin_lock_irqsave(&lweventlist_lock, flags);
 107         if (list_empty(&dev->link_watch_list)) {
 108                 list_add_tail(&dev->link_watch_list, &lweventlist);
 109                 dev_hold(dev);
 110         }
 111         spin_unlock_irqrestore(&lweventlist_lock, flags);
 112 }
 113
 114
 115 static void linkwatch_schedule_work(int urgent)
 116 {
 117         unsigned long delay = linkwatch_nextevent - jiffies;
 118
 119         if (test_bit(LW_URGENT, &linkwatch_flags))
 120                 return;
 121
 122         /* Minimise down-time: drop delay for up event. */
 123         if (urgent) {
 124                 if (test_and_set_bit(LW_URGENT, &linkwatch_flags))
 125                         return;
 126                 delay = 0;
 127         }
 128
 129         /* If we wrap around we'll delay it by at most HZ. */
 130         if (delay > HZ)
 131                 delay = 0;
 132
 133         /*
 134          * If urgent, schedule immediate execution; otherwise, don't
 135          * override the existing timer.
 136          */
 137         if (test_bit(LW_URGENT, &linkwatch_flags))
 138                 mod_delayed_work(system_wq, &linkwatch_work, 0);
 139         else
 140                 schedule_delayed_work(&linkwatch_work, delay);
 141 }
 142
 143
 144 static void linkwatch_do_dev(struct net_device *dev)
 145 {
 146         /*
 147          * Make sure the above read is complete since it can be
 148          * rewritten as soon as we clear the bit below.
 149          */
 150         smp_mb__before_atomic();
 151
 152         /* We are about to handle this device,
 153          * so new events can be accepted
 154          */
 155         clear_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state);
 156
 157         rfc2863_policy(dev);
 158         if (dev->flags & IFF_UP) {
 159                 if (netif_carrier_ok(dev))
 160                         dev_activate(dev);
 161                 else
 162                         dev_deactivate(dev);
 163
 164                 netdev_state_change(dev);
 165         }
 166         dev_put(dev);
 167 }
 168
 169 static void __linkwatch_run_queue(int urgent_only)
 170 {
 171         struct net_device *dev;
 172         LIST_HEAD(wrk);
 173
 174         /*
 175          * Limit the number of linkwatch events to one
 176          * per second so that a runaway driver does not
 177          * cause a storm of messages on the netlink
 178          * socket.  This limit does not apply to up events
 179          * while the device qdisc is down.
 180          */
 181         if (!urgent_only)
 182                 linkwatch_nextevent = jiffies + HZ;
 183         /* Limit wrap-around effect on delay. */
 184         else if (time_after(linkwatch_nextevent, jiffies + HZ))
 185                 linkwatch_nextevent = jiffies;
 186
 187         clear_bit(LW_URGENT, &linkwatch_flags);
 188
 189         spin_lock_irq(&lweventlist_lock);
 190         list_splice_init(&lweventlist, &wrk);
 191
 192         while (!list_empty(&wrk)) {
 193
 194                 dev = list_first_entry(&wrk, struct net_device, link_watch_list);
 195                 list_del_init(&dev->link_watch_list);
 196
 197                 if (urgent_only && !linkwatch_urgent_event(dev)) {
 198                         list_add_tail(&dev->link_watch_list, &lweventlist);
 199                         continue;
 200                 }
 201                 spin_unlock_irq(&lweventlist_lock);
 202                 linkwatch_do_dev(dev);
 203                 spin_lock_irq(&lweventlist_lock);
 204         }
 205
 206         if (!list_empty(&lweventlist))
 207                 linkwatch_schedule_work(0);
 208         spin_unlock_irq(&lweventlist_lock);
 209 }
 210
 211 void linkwatch_forget_dev(struct net_device *dev)
 212 {
 213         unsigned long flags;
 214         int clean = 0;
 215
 216         spin_lock_irqsave(&lweventlist_lock, flags);
 217         if (!list_empty(&dev->link_watch_list)) {
 218                 list_del_init(&dev->link_watch_list);
 219                 clean = 1;
 220         }
 221         spin_unlock_irqrestore(&lweventlist_lock, flags);
 222         if (clean)
 223                 linkwatch_do_dev(dev);
 224 }
 225
 226
 227 /* Must be called with the rtnl semaphore held */
 228 void linkwatch_run_queue(void)
 229 {
 230         __linkwatch_run_queue(0);
 231 }
 232
 233
 234 static void linkwatch_event(struct work_struct *dummy)
 235 {
 236         rtnl_lock();
 237         __linkwatch_run_queue(time_after(linkwatch_nextevent, jiffies));
 238         rtnl_unlock();
 239 }
 240
 241
 242 void linkwatch_fire_event(struct net_device *dev)
 243 {
 244         bool urgent = linkwatch_urgent_event(dev);
 245
 246         if (!test_and_set_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state)) {
 247                 linkwatch_add_event(dev);
 248         } else if (!urgent)
 249                 return;
 250
 251         linkwatch_schedule_work(urgent);
 252 }
 253 EXPORT_SYMBOL(linkwatch_fire_event);