fix the softlockup watchdog to actually work
[linux-2.6.git] / kernel / softlockup.c
1 /*
2  * Detect Soft Lockups
3  *
4  * started by Ingo Molnar, Copyright (C) 2005, 2006 Red Hat, Inc.
5  *
6  * this code detects soft lockups: incidents in where on a CPU
7  * the kernel does not reschedule for 10 seconds or more.
8  */
9 #include <linux/mm.h>
10 #include <linux/cpu.h>
11 #include <linux/init.h>
12 #include <linux/delay.h>
13 #include <linux/freezer.h>
14 #include <linux/kthread.h>
15 #include <linux/notifier.h>
16 #include <linux/module.h>
17
18 static DEFINE_SPINLOCK(print_lock);
19
20 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
21 static DEFINE_PER_CPU(unsigned long, print_timestamp);
22 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
23
24 static int did_panic = 0;
25
26 static int
27 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
28 {
29         did_panic = 1;
30
31         return NOTIFY_DONE;
32 }
33
34 static struct notifier_block panic_block = {
35         .notifier_call = softlock_panic,
36 };
37
38 /*
39  * Returns seconds, approximately.  We don't need nanosecond
40  * resolution, and we don't need to waste time with a big divide when
41  * 2^30ns == 1.074s.
42  */
43 static unsigned long get_timestamp(int this_cpu)
44 {
45         return cpu_clock(this_cpu) >> 30;  /* 2^30 ~= 10^9 */
46 }
47
48 void touch_softlockup_watchdog(void)
49 {
50         int this_cpu = raw_smp_processor_id();
51
52         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
53 }
54 EXPORT_SYMBOL(touch_softlockup_watchdog);
55
56 void touch_all_softlockup_watchdogs(void)
57 {
58         int cpu;
59
60         /* Cause each CPU to re-update its timestamp rather than complain */
61         for_each_online_cpu(cpu)
62                 per_cpu(touch_timestamp, cpu) = 0;
63 }
64 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
65
66 /*
67  * This callback runs from the timer interrupt, and checks
68  * whether the watchdog thread has hung or not:
69  */
70 void softlockup_tick(void)
71 {
72         int this_cpu = smp_processor_id();
73         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
74         unsigned long print_timestamp;
75         unsigned long now;
76
77         if (touch_timestamp == 0) {
78                 touch_softlockup_watchdog();
79                 return;
80         }
81
82         print_timestamp = per_cpu(print_timestamp, this_cpu);
83
84         /* report at most once a second */
85         if ((print_timestamp >= touch_timestamp &&
86                         print_timestamp < (touch_timestamp + 1)) ||
87                         did_panic || !per_cpu(watchdog_task, this_cpu)) {
88                 return;
89         }
90
91         /* do not print during early bootup: */
92         if (unlikely(system_state != SYSTEM_RUNNING)) {
93                 touch_softlockup_watchdog();
94                 return;
95         }
96
97         now = get_timestamp(this_cpu);
98
99         /* Wake up the high-prio watchdog task every second: */
100         if (now > (touch_timestamp + 1))
101                 wake_up_process(per_cpu(watchdog_task, this_cpu));
102
103         /* Warn about unreasonable 10+ seconds delays: */
104         if (now > (touch_timestamp + 10)) {
105                 per_cpu(print_timestamp, this_cpu) = touch_timestamp;
106
107                 spin_lock(&print_lock);
108                 printk(KERN_ERR "BUG: soft lockup detected on CPU#%d!\n",
109                         this_cpu);
110                 dump_stack();
111                 spin_unlock(&print_lock);
112         }
113 }
114
115 /*
116  * The watchdog thread - runs every second and touches the timestamp.
117  */
118 static int watchdog(void * __bind_cpu)
119 {
120         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
121
122         sched_setscheduler(current, SCHED_FIFO, &param);
123
124         /* initialize timestamp */
125         touch_softlockup_watchdog();
126
127         /*
128          * Run briefly once per second to reset the softlockup timestamp.
129          * If this gets delayed for more than 10 seconds then the
130          * debug-printout triggers in softlockup_tick().
131          */
132         while (!kthread_should_stop()) {
133                 set_current_state(TASK_INTERRUPTIBLE);
134                 touch_softlockup_watchdog();
135                 schedule();
136         }
137
138         return 0;
139 }
140
141 /*
142  * Create/destroy watchdog threads as CPUs come and go:
143  */
144 static int __cpuinit
145 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
146 {
147         int hotcpu = (unsigned long)hcpu;
148         struct task_struct *p;
149
150         switch (action) {
151         case CPU_UP_PREPARE:
152         case CPU_UP_PREPARE_FROZEN:
153                 BUG_ON(per_cpu(watchdog_task, hotcpu));
154                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
155                 if (IS_ERR(p)) {
156                         printk("watchdog for %i failed\n", hotcpu);
157                         return NOTIFY_BAD;
158                 }
159                 per_cpu(touch_timestamp, hotcpu) = 0;
160                 per_cpu(watchdog_task, hotcpu) = p;
161                 kthread_bind(p, hotcpu);
162                 break;
163         case CPU_ONLINE:
164         case CPU_ONLINE_FROZEN:
165                 wake_up_process(per_cpu(watchdog_task, hotcpu));
166                 break;
167 #ifdef CONFIG_HOTPLUG_CPU
168         case CPU_UP_CANCELED:
169         case CPU_UP_CANCELED_FROZEN:
170                 if (!per_cpu(watchdog_task, hotcpu))
171                         break;
172                 /* Unbind so it can run.  Fall thru. */
173                 kthread_bind(per_cpu(watchdog_task, hotcpu),
174                              any_online_cpu(cpu_online_map));
175         case CPU_DEAD:
176         case CPU_DEAD_FROZEN:
177                 p = per_cpu(watchdog_task, hotcpu);
178                 per_cpu(watchdog_task, hotcpu) = NULL;
179                 kthread_stop(p);
180                 break;
181 #endif /* CONFIG_HOTPLUG_CPU */
182         }
183         return NOTIFY_OK;
184 }
185
186 static struct notifier_block __cpuinitdata cpu_nfb = {
187         .notifier_call = cpu_callback
188 };
189
190 __init void spawn_softlockup_task(void)
191 {
192         void *cpu = (void *)(long)smp_processor_id();
193         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
194
195         BUG_ON(err == NOTIFY_BAD);
196         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
197         register_cpu_notifier(&cpu_nfb);
198
199         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
200 }