softlockup: allow panic on lockup
[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/nmi.h>
12 #include <linux/init.h>
13 #include <linux/delay.h>
14 #include <linux/freezer.h>
15 #include <linux/kthread.h>
16 #include <linux/notifier.h>
17 #include <linux/module.h>
18
19 #include <asm/irq_regs.h>
20
21 static DEFINE_SPINLOCK(print_lock);
22
23 static DEFINE_PER_CPU(unsigned long, touch_timestamp);
24 static DEFINE_PER_CPU(unsigned long, print_timestamp);
25 static DEFINE_PER_CPU(struct task_struct *, watchdog_task);
26
27 static int __read_mostly did_panic;
28 unsigned long __read_mostly softlockup_thresh = 60;
29
30 /*
31  * Should we panic (and reboot, if panic_timeout= is set) when a
32  * soft-lockup occurs:
33  */
34 unsigned int __read_mostly softlockup_panic =
35                                 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
36
37 static int __init softlockup_panic_setup(char *str)
38 {
39         softlockup_panic = simple_strtoul(str, NULL, 0);
40
41         return 1;
42 }
43 __setup("softlockup_panic=", softlockup_panic_setup);
44
45 static int
46 softlock_panic(struct notifier_block *this, unsigned long event, void *ptr)
47 {
48         did_panic = 1;
49
50         return NOTIFY_DONE;
51 }
52
53 static struct notifier_block panic_block = {
54         .notifier_call = softlock_panic,
55 };
56
57 /*
58  * Returns seconds, approximately.  We don't need nanosecond
59  * resolution, and we don't need to waste time with a big divide when
60  * 2^30ns == 1.074s.
61  */
62 static unsigned long get_timestamp(int this_cpu)
63 {
64         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
65 }
66
67 void touch_softlockup_watchdog(void)
68 {
69         int this_cpu = raw_smp_processor_id();
70
71         __raw_get_cpu_var(touch_timestamp) = get_timestamp(this_cpu);
72 }
73 EXPORT_SYMBOL(touch_softlockup_watchdog);
74
75 void touch_all_softlockup_watchdogs(void)
76 {
77         int cpu;
78
79         /* Cause each CPU to re-update its timestamp rather than complain */
80         for_each_online_cpu(cpu)
81                 per_cpu(touch_timestamp, cpu) = 0;
82 }
83 EXPORT_SYMBOL(touch_all_softlockup_watchdogs);
84
85 /*
86  * This callback runs from the timer interrupt, and checks
87  * whether the watchdog thread has hung or not:
88  */
89 void softlockup_tick(void)
90 {
91         int this_cpu = smp_processor_id();
92         unsigned long touch_timestamp = per_cpu(touch_timestamp, this_cpu);
93         unsigned long print_timestamp;
94         struct pt_regs *regs = get_irq_regs();
95         unsigned long now;
96
97         if (touch_timestamp == 0) {
98                 touch_softlockup_watchdog();
99                 return;
100         }
101
102         print_timestamp = per_cpu(print_timestamp, this_cpu);
103
104         /* report at most once a second */
105         if ((print_timestamp >= touch_timestamp &&
106                         print_timestamp < (touch_timestamp + 1)) ||
107                         did_panic || !per_cpu(watchdog_task, this_cpu)) {
108                 return;
109         }
110
111         /* do not print during early bootup: */
112         if (unlikely(system_state != SYSTEM_RUNNING)) {
113                 touch_softlockup_watchdog();
114                 return;
115         }
116
117         now = get_timestamp(this_cpu);
118
119         /* Wake up the high-prio watchdog task every second: */
120         if (now > (touch_timestamp + 1))
121                 wake_up_process(per_cpu(watchdog_task, this_cpu));
122
123         /* Warn about unreasonable delays: */
124         if (now <= (touch_timestamp + softlockup_thresh))
125                 return;
126
127         per_cpu(print_timestamp, this_cpu) = touch_timestamp;
128
129         spin_lock(&print_lock);
130         printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %lus! [%s:%d]\n",
131                         this_cpu, now - touch_timestamp,
132                         current->comm, task_pid_nr(current));
133         if (regs)
134                 show_regs(regs);
135         else
136                 dump_stack();
137         spin_unlock(&print_lock);
138
139         if (softlockup_panic)
140                 panic("softlockup: hung tasks");
141 }
142
143 /*
144  * Have a reasonable limit on the number of tasks checked:
145  */
146 unsigned long __read_mostly sysctl_hung_task_check_count = 1024;
147
148 /*
149  * Zero means infinite timeout - no checking done:
150  */
151 unsigned long __read_mostly sysctl_hung_task_timeout_secs = 120;
152
153 unsigned long __read_mostly sysctl_hung_task_warnings = 10;
154
155 /*
156  * Only do the hung-tasks check on one CPU:
157  */
158 static int check_cpu __read_mostly = -1;
159
160 static void check_hung_task(struct task_struct *t, unsigned long now)
161 {
162         unsigned long switch_count = t->nvcsw + t->nivcsw;
163
164         if (t->flags & PF_FROZEN)
165                 return;
166
167         if (switch_count != t->last_switch_count || !t->last_switch_timestamp) {
168                 t->last_switch_count = switch_count;
169                 t->last_switch_timestamp = now;
170                 return;
171         }
172         if ((long)(now - t->last_switch_timestamp) <
173                                         sysctl_hung_task_timeout_secs)
174                 return;
175         if (sysctl_hung_task_warnings < 0)
176                 return;
177         sysctl_hung_task_warnings--;
178
179         /*
180          * Ok, the task did not get scheduled for more than 2 minutes,
181          * complain:
182          */
183         printk(KERN_ERR "INFO: task %s:%d blocked for more than "
184                         "%ld seconds.\n", t->comm, t->pid,
185                         sysctl_hung_task_timeout_secs);
186         printk(KERN_ERR "\"echo 0 > /proc/sys/kernel/hung_task_timeout_secs\""
187                         " disables this message.\n");
188         sched_show_task(t);
189         __debug_show_held_locks(t);
190
191         t->last_switch_timestamp = now;
192         touch_nmi_watchdog();
193
194         if (softlockup_panic)
195                 panic("softlockup: blocked tasks");
196 }
197
198 /*
199  * Check whether a TASK_UNINTERRUPTIBLE does not get woken up for
200  * a really long time (120 seconds). If that happens, print out
201  * a warning.
202  */
203 static void check_hung_uninterruptible_tasks(int this_cpu)
204 {
205         int max_count = sysctl_hung_task_check_count;
206         unsigned long now = get_timestamp(this_cpu);
207         struct task_struct *g, *t;
208
209         /*
210          * If the system crashed already then all bets are off,
211          * do not report extra hung tasks:
212          */
213         if ((tainted & TAINT_DIE) || did_panic)
214                 return;
215
216         read_lock(&tasklist_lock);
217         do_each_thread(g, t) {
218                 if (!--max_count)
219                         goto unlock;
220                 if (t->state & TASK_UNINTERRUPTIBLE)
221                         check_hung_task(t, now);
222         } while_each_thread(g, t);
223  unlock:
224         read_unlock(&tasklist_lock);
225 }
226
227 /*
228  * The watchdog thread - runs every second and touches the timestamp.
229  */
230 static int watchdog(void *__bind_cpu)
231 {
232         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
233         int this_cpu = (long)__bind_cpu;
234
235         sched_setscheduler(current, SCHED_FIFO, &param);
236
237         /* initialize timestamp */
238         touch_softlockup_watchdog();
239
240         set_current_state(TASK_INTERRUPTIBLE);
241         /*
242          * Run briefly once per second to reset the softlockup timestamp.
243          * If this gets delayed for more than 60 seconds then the
244          * debug-printout triggers in softlockup_tick().
245          */
246         while (!kthread_should_stop()) {
247                 touch_softlockup_watchdog();
248                 schedule();
249
250                 if (kthread_should_stop())
251                         break;
252
253                 if (this_cpu == check_cpu) {
254                         if (sysctl_hung_task_timeout_secs)
255                                 check_hung_uninterruptible_tasks(this_cpu);
256                 }
257
258                 set_current_state(TASK_INTERRUPTIBLE);
259         }
260         __set_current_state(TASK_RUNNING);
261
262         return 0;
263 }
264
265 /*
266  * Create/destroy watchdog threads as CPUs come and go:
267  */
268 static int __cpuinit
269 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
270 {
271         int hotcpu = (unsigned long)hcpu;
272         struct task_struct *p;
273
274         switch (action) {
275         case CPU_UP_PREPARE:
276         case CPU_UP_PREPARE_FROZEN:
277                 BUG_ON(per_cpu(watchdog_task, hotcpu));
278                 p = kthread_create(watchdog, hcpu, "watchdog/%d", hotcpu);
279                 if (IS_ERR(p)) {
280                         printk(KERN_ERR "watchdog for %i failed\n", hotcpu);
281                         return NOTIFY_BAD;
282                 }
283                 per_cpu(touch_timestamp, hotcpu) = 0;
284                 per_cpu(watchdog_task, hotcpu) = p;
285                 kthread_bind(p, hotcpu);
286                 break;
287         case CPU_ONLINE:
288         case CPU_ONLINE_FROZEN:
289                 check_cpu = any_online_cpu(cpu_online_map);
290                 wake_up_process(per_cpu(watchdog_task, hotcpu));
291                 break;
292 #ifdef CONFIG_HOTPLUG_CPU
293         case CPU_DOWN_PREPARE:
294         case CPU_DOWN_PREPARE_FROZEN:
295                 if (hotcpu == check_cpu) {
296                         cpumask_t temp_cpu_online_map = cpu_online_map;
297
298                         cpu_clear(hotcpu, temp_cpu_online_map);
299                         check_cpu = any_online_cpu(temp_cpu_online_map);
300                 }
301                 break;
302
303         case CPU_UP_CANCELED:
304         case CPU_UP_CANCELED_FROZEN:
305                 if (!per_cpu(watchdog_task, hotcpu))
306                         break;
307                 /* Unbind so it can run.  Fall thru. */
308                 kthread_bind(per_cpu(watchdog_task, hotcpu),
309                              any_online_cpu(cpu_online_map));
310         case CPU_DEAD:
311         case CPU_DEAD_FROZEN:
312                 p = per_cpu(watchdog_task, hotcpu);
313                 per_cpu(watchdog_task, hotcpu) = NULL;
314                 kthread_stop(p);
315                 break;
316 #endif /* CONFIG_HOTPLUG_CPU */
317         }
318         return NOTIFY_OK;
319 }
320
321 static struct notifier_block __cpuinitdata cpu_nfb = {
322         .notifier_call = cpu_callback
323 };
324
325 __init void spawn_softlockup_task(void)
326 {
327         void *cpu = (void *)(long)smp_processor_id();
328         int err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
329
330         BUG_ON(err == NOTIFY_BAD);
331         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
332         register_cpu_notifier(&cpu_nfb);
333
334         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
335 }