[NET] link_watch: Eliminate potential delay on wrap-around
[linux-2.6] / 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/slab.h>
23 #include <linux/workqueue.h>
24 #include <linux/bitops.h>
25 #include <asm/types.h>
26
27
28 enum lw_bits {
29         LW_RUNNING = 0,
30 };
31
32 static unsigned long linkwatch_flags;
33 static unsigned long linkwatch_nextevent;
34
35 static void linkwatch_event(struct work_struct *dummy);
36 static DECLARE_DELAYED_WORK(linkwatch_work, linkwatch_event);
37
38 static struct net_device *lweventlist;
39 static DEFINE_SPINLOCK(lweventlist_lock);
40
41 static unsigned char default_operstate(const struct net_device *dev)
42 {
43         if (!netif_carrier_ok(dev))
44                 return (dev->ifindex != dev->iflink ?
45                         IF_OPER_LOWERLAYERDOWN : IF_OPER_DOWN);
46
47         if (netif_dormant(dev))
48                 return IF_OPER_DORMANT;
49
50         return IF_OPER_UP;
51 }
52
53
54 static void rfc2863_policy(struct net_device *dev)
55 {
56         unsigned char operstate = default_operstate(dev);
57
58         if (operstate == dev->operstate)
59                 return;
60
61         write_lock_bh(&dev_base_lock);
62
63         switch(dev->link_mode) {
64         case IF_LINK_MODE_DORMANT:
65                 if (operstate == IF_OPER_UP)
66                         operstate = IF_OPER_DORMANT;
67                 break;
68
69         case IF_LINK_MODE_DEFAULT:
70         default:
71                 break;
72         }
73
74         dev->operstate = operstate;
75
76         write_unlock_bh(&dev_base_lock);
77 }
78
79
80 static int linkwatch_urgent_event(struct net_device *dev)
81 {
82         return netif_running(dev) && netif_carrier_ok(dev) &&
83                dev->qdisc != dev->qdisc_sleeping;
84 }
85
86
87 static void linkwatch_add_event(struct net_device *dev)
88 {
89         unsigned long flags;
90
91         spin_lock_irqsave(&lweventlist_lock, flags);
92         dev->link_watch_next = lweventlist;
93         lweventlist = dev;
94         spin_unlock_irqrestore(&lweventlist_lock, flags);
95 }
96
97
98 static void linkwatch_schedule_work(unsigned long delay)
99 {
100         if (test_and_set_bit(LW_RUNNING, &linkwatch_flags))
101                 return;
102
103         /* If we wrap around we'll delay it by at most HZ. */
104         if (delay > HZ) {
105                 linkwatch_nextevent = jiffies;
106                 delay = 0;
107         }
108
109         schedule_delayed_work(&linkwatch_work, delay);
110 }
111
112
113 static void __linkwatch_run_queue(int urgent_only)
114 {
115         struct net_device *next;
116
117         /*
118          * Limit the number of linkwatch events to one
119          * per second so that a runaway driver does not
120          * cause a storm of messages on the netlink
121          * socket.  This limit does not apply to up events
122          * while the device qdisc is down.
123          */
124         if (!urgent_only)
125                 linkwatch_nextevent = jiffies + HZ;
126         clear_bit(LW_RUNNING, &linkwatch_flags);
127
128         spin_lock_irq(&lweventlist_lock);
129         next = lweventlist;
130         lweventlist = NULL;
131         spin_unlock_irq(&lweventlist_lock);
132
133         while (next) {
134                 struct net_device *dev = next;
135
136                 next = dev->link_watch_next;
137
138                 if (urgent_only && !linkwatch_urgent_event(dev)) {
139                         linkwatch_add_event(dev);
140                         continue;
141                 }
142
143                 /*
144                  * Make sure the above read is complete since it can be
145                  * rewritten as soon as we clear the bit below.
146                  */
147                 smp_mb__before_clear_bit();
148
149                 /* We are about to handle this device,
150                  * so new events can be accepted
151                  */
152                 clear_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state);
153
154                 rfc2863_policy(dev);
155                 if (dev->flags & IFF_UP) {
156                         if (netif_carrier_ok(dev)) {
157                                 WARN_ON(dev->qdisc_sleeping == &noop_qdisc);
158                                 dev_activate(dev);
159                         } else
160                                 dev_deactivate(dev);
161
162                         netdev_state_change(dev);
163                 }
164
165                 dev_put(dev);
166         }
167
168         if (lweventlist)
169                 linkwatch_schedule_work(linkwatch_nextevent - jiffies);
170 }
171
172
173 /* Must be called with the rtnl semaphore held */
174 void linkwatch_run_queue(void)
175 {
176         __linkwatch_run_queue(0);
177 }
178
179
180 static void linkwatch_event(struct work_struct *dummy)
181 {
182         rtnl_lock();
183         __linkwatch_run_queue(time_after(linkwatch_nextevent, jiffies));
184         rtnl_unlock();
185 }
186
187
188 void linkwatch_fire_event(struct net_device *dev)
189 {
190         if (!test_and_set_bit(__LINK_STATE_LINKWATCH_PENDING, &dev->state)) {
191                 unsigned long delay;
192
193                 dev_hold(dev);
194
195                 linkwatch_add_event(dev);
196
197                 delay = linkwatch_nextevent - jiffies;
198
199                 /* Minimise down-time: drop delay for up event. */
200                 if (linkwatch_urgent_event(dev))
201                         delay = 0;
202
203                 linkwatch_schedule_work(delay);
204         }
205 }
206
207 EXPORT_SYMBOL(linkwatch_fire_event);