watchdog: Avoid kernel crash when disabling watchdog
[linux-2.6.git] / kernel / watchdog.c
1 /*
2  * Detect hard and soft lockups on a system
3  *
4  * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5  *
6  * this code detects hard lockups: incidents in where on a CPU
7  * the kernel does not respond to anything except NMI.
8  *
9  * Note: Most of this code is borrowed heavily from softlockup.c,
10  * so thanks to Ingo for the initial implementation.
11  * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
12  * to those contributors as well.
13  */
14
15 #include <linux/mm.h>
16 #include <linux/cpu.h>
17 #include <linux/nmi.h>
18 #include <linux/init.h>
19 #include <linux/delay.h>
20 #include <linux/freezer.h>
21 #include <linux/kthread.h>
22 #include <linux/lockdep.h>
23 #include <linux/notifier.h>
24 #include <linux/module.h>
25 #include <linux/sysctl.h>
26
27 #include <asm/irq_regs.h>
28 #include <linux/perf_event.h>
29
30 int watchdog_enabled;
31 int __read_mostly softlockup_thresh = 60;
32
33 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
34 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
35 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
36 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
37 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
38 #ifdef CONFIG_HARDLOCKUP_DETECTOR
39 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
41 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
43 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
44 #endif
45
46 static int __initdata no_watchdog;
47
48
49 /* boot commands */
50 /*
51  * Should we panic when a soft-lockup or hard-lockup occurs:
52  */
53 #ifdef CONFIG_HARDLOCKUP_DETECTOR
54 static int hardlockup_panic;
55
56 static int __init hardlockup_panic_setup(char *str)
57 {
58         if (!strncmp(str, "panic", 5))
59                 hardlockup_panic = 1;
60         return 1;
61 }
62 __setup("nmi_watchdog=", hardlockup_panic_setup);
63 #endif
64
65 unsigned int __read_mostly softlockup_panic =
66                         CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
67
68 static int __init softlockup_panic_setup(char *str)
69 {
70         softlockup_panic = simple_strtoul(str, NULL, 0);
71
72         return 1;
73 }
74 __setup("softlockup_panic=", softlockup_panic_setup);
75
76 static int __init nowatchdog_setup(char *str)
77 {
78         no_watchdog = 1;
79         return 1;
80 }
81 __setup("nowatchdog", nowatchdog_setup);
82
83 /* deprecated */
84 static int __init nosoftlockup_setup(char *str)
85 {
86         no_watchdog = 1;
87         return 1;
88 }
89 __setup("nosoftlockup", nosoftlockup_setup);
90 /*  */
91
92
93 /*
94  * Returns seconds, approximately.  We don't need nanosecond
95  * resolution, and we don't need to waste time with a big divide when
96  * 2^30ns == 1.074s.
97  */
98 static unsigned long get_timestamp(int this_cpu)
99 {
100         return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
101 }
102
103 static unsigned long get_sample_period(void)
104 {
105         /*
106          * convert softlockup_thresh from seconds to ns
107          * the divide by 5 is to give hrtimer 5 chances to
108          * increment before the hardlockup detector generates
109          * a warning
110          */
111         return softlockup_thresh / 5 * NSEC_PER_SEC;
112 }
113
114 /* Commands for resetting the watchdog */
115 static void __touch_watchdog(void)
116 {
117         int this_cpu = smp_processor_id();
118
119         __get_cpu_var(watchdog_touch_ts) = get_timestamp(this_cpu);
120 }
121
122 void touch_softlockup_watchdog(void)
123 {
124         __raw_get_cpu_var(watchdog_touch_ts) = 0;
125 }
126 EXPORT_SYMBOL(touch_softlockup_watchdog);
127
128 void touch_all_softlockup_watchdogs(void)
129 {
130         int cpu;
131
132         /*
133          * this is done lockless
134          * do we care if a 0 races with a timestamp?
135          * all it means is the softlock check starts one cycle later
136          */
137         for_each_online_cpu(cpu)
138                 per_cpu(watchdog_touch_ts, cpu) = 0;
139 }
140
141 #ifdef CONFIG_HARDLOCKUP_DETECTOR
142 void touch_nmi_watchdog(void)
143 {
144         if (watchdog_enabled) {
145                 unsigned cpu;
146
147                 for_each_present_cpu(cpu) {
148                         if (per_cpu(watchdog_nmi_touch, cpu) != true)
149                                 per_cpu(watchdog_nmi_touch, cpu) = true;
150                 }
151         }
152         touch_softlockup_watchdog();
153 }
154 EXPORT_SYMBOL(touch_nmi_watchdog);
155
156 #endif
157
158 void touch_softlockup_watchdog_sync(void)
159 {
160         __raw_get_cpu_var(softlockup_touch_sync) = true;
161         __raw_get_cpu_var(watchdog_touch_ts) = 0;
162 }
163
164 #ifdef CONFIG_HARDLOCKUP_DETECTOR
165 /* watchdog detector functions */
166 static int is_hardlockup(void)
167 {
168         unsigned long hrint = __get_cpu_var(hrtimer_interrupts);
169
170         if (__get_cpu_var(hrtimer_interrupts_saved) == hrint)
171                 return 1;
172
173         __get_cpu_var(hrtimer_interrupts_saved) = hrint;
174         return 0;
175 }
176 #endif
177
178 static int is_softlockup(unsigned long touch_ts)
179 {
180         unsigned long now = get_timestamp(smp_processor_id());
181
182         /* Warn about unreasonable delays: */
183         if (time_after(now, touch_ts + softlockup_thresh))
184                 return now - touch_ts;
185
186         return 0;
187 }
188
189 #ifdef CONFIG_HARDLOCKUP_DETECTOR
190 static struct perf_event_attr wd_hw_attr = {
191         .type           = PERF_TYPE_HARDWARE,
192         .config         = PERF_COUNT_HW_CPU_CYCLES,
193         .size           = sizeof(struct perf_event_attr),
194         .pinned         = 1,
195         .disabled       = 1,
196 };
197
198 /* Callback function for perf event subsystem */
199 void watchdog_overflow_callback(struct perf_event *event, int nmi,
200                  struct perf_sample_data *data,
201                  struct pt_regs *regs)
202 {
203         /* Ensure the watchdog never gets throttled */
204         event->hw.interrupts = 0;
205
206         if (__get_cpu_var(watchdog_nmi_touch) == true) {
207                 __get_cpu_var(watchdog_nmi_touch) = false;
208                 return;
209         }
210
211         /* check for a hardlockup
212          * This is done by making sure our timer interrupt
213          * is incrementing.  The timer interrupt should have
214          * fired multiple times before we overflow'd.  If it hasn't
215          * then this is a good indication the cpu is stuck
216          */
217         if (is_hardlockup()) {
218                 int this_cpu = smp_processor_id();
219
220                 /* only print hardlockups once */
221                 if (__get_cpu_var(hard_watchdog_warn) == true)
222                         return;
223
224                 if (hardlockup_panic)
225                         panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
226                 else
227                         WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
228
229                 __get_cpu_var(hard_watchdog_warn) = true;
230                 return;
231         }
232
233         __get_cpu_var(hard_watchdog_warn) = false;
234         return;
235 }
236 static void watchdog_interrupt_count(void)
237 {
238         __get_cpu_var(hrtimer_interrupts)++;
239 }
240 #else
241 static inline void watchdog_interrupt_count(void) { return; }
242 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
243
244 /* watchdog kicker functions */
245 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
246 {
247         unsigned long touch_ts = __get_cpu_var(watchdog_touch_ts);
248         struct pt_regs *regs = get_irq_regs();
249         int duration;
250
251         /* kick the hardlockup detector */
252         watchdog_interrupt_count();
253
254         /* kick the softlockup detector */
255         wake_up_process(__get_cpu_var(softlockup_watchdog));
256
257         /* .. and repeat */
258         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
259
260         if (touch_ts == 0) {
261                 if (unlikely(__get_cpu_var(softlockup_touch_sync))) {
262                         /*
263                          * If the time stamp was touched atomically
264                          * make sure the scheduler tick is up to date.
265                          */
266                         __get_cpu_var(softlockup_touch_sync) = false;
267                         sched_clock_tick();
268                 }
269                 __touch_watchdog();
270                 return HRTIMER_RESTART;
271         }
272
273         /* check for a softlockup
274          * This is done by making sure a high priority task is
275          * being scheduled.  The task touches the watchdog to
276          * indicate it is getting cpu time.  If it hasn't then
277          * this is a good indication some task is hogging the cpu
278          */
279         duration = is_softlockup(touch_ts);
280         if (unlikely(duration)) {
281                 /* only warn once */
282                 if (__get_cpu_var(soft_watchdog_warn) == true)
283                         return HRTIMER_RESTART;
284
285                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
286                         smp_processor_id(), duration,
287                         current->comm, task_pid_nr(current));
288                 print_modules();
289                 print_irqtrace_events(current);
290                 if (regs)
291                         show_regs(regs);
292                 else
293                         dump_stack();
294
295                 if (softlockup_panic)
296                         panic("softlockup: hung tasks");
297                 __get_cpu_var(soft_watchdog_warn) = true;
298         } else
299                 __get_cpu_var(soft_watchdog_warn) = false;
300
301         return HRTIMER_RESTART;
302 }
303
304
305 /*
306  * The watchdog thread - touches the timestamp.
307  */
308 static int watchdog(void *unused)
309 {
310         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
311         struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
312
313         sched_setscheduler(current, SCHED_FIFO, &param);
314
315         /* initialize timestamp */
316         __touch_watchdog();
317
318         /* kick off the timer for the hardlockup detector */
319         /* done here because hrtimer_start can only pin to smp_processor_id() */
320         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
321                       HRTIMER_MODE_REL_PINNED);
322
323         set_current_state(TASK_INTERRUPTIBLE);
324         /*
325          * Run briefly once per second to reset the softlockup timestamp.
326          * If this gets delayed for more than 60 seconds then the
327          * debug-printout triggers in watchdog_timer_fn().
328          */
329         while (!kthread_should_stop()) {
330                 __touch_watchdog();
331                 schedule();
332
333                 if (kthread_should_stop())
334                         break;
335
336                 set_current_state(TASK_INTERRUPTIBLE);
337         }
338         __set_current_state(TASK_RUNNING);
339
340         return 0;
341 }
342
343
344 #ifdef CONFIG_HARDLOCKUP_DETECTOR
345 static int watchdog_nmi_enable(int cpu)
346 {
347         struct perf_event_attr *wd_attr;
348         struct perf_event *event = per_cpu(watchdog_ev, cpu);
349
350         /* is it already setup and enabled? */
351         if (event && event->state > PERF_EVENT_STATE_OFF)
352                 goto out;
353
354         /* it is setup but not enabled */
355         if (event != NULL)
356                 goto out_enable;
357
358         /* Try to register using hardware perf events */
359         wd_attr = &wd_hw_attr;
360         wd_attr->sample_period = hw_nmi_get_sample_period();
361         event = perf_event_create_kernel_counter(wd_attr, cpu, -1, watchdog_overflow_callback);
362         if (!IS_ERR(event)) {
363                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
364                 goto out_save;
365         }
366
367         printk(KERN_ERR "NMI watchdog failed to create perf event on cpu%i: %p\n", cpu, event);
368         return PTR_ERR(event);
369
370         /* success path */
371 out_save:
372         per_cpu(watchdog_ev, cpu) = event;
373 out_enable:
374         perf_event_enable(per_cpu(watchdog_ev, cpu));
375 out:
376         return 0;
377 }
378
379 static void watchdog_nmi_disable(int cpu)
380 {
381         struct perf_event *event = per_cpu(watchdog_ev, cpu);
382
383         if (event) {
384                 perf_event_disable(event);
385                 per_cpu(watchdog_ev, cpu) = NULL;
386
387                 /* should be in cleanup, but blocks oprofile */
388                 perf_event_release_kernel(event);
389         }
390         return;
391 }
392 #else
393 static int watchdog_nmi_enable(int cpu) { return 0; }
394 static void watchdog_nmi_disable(int cpu) { return; }
395 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
396
397 /* prepare/enable/disable routines */
398 static int watchdog_prepare_cpu(int cpu)
399 {
400         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
401
402         WARN_ON(per_cpu(softlockup_watchdog, cpu));
403         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
404         hrtimer->function = watchdog_timer_fn;
405
406         return 0;
407 }
408
409 static int watchdog_enable(int cpu)
410 {
411         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
412         int err;
413
414         /* enable the perf event */
415         err = watchdog_nmi_enable(cpu);
416         if (err)
417                 return err;
418
419         /* create the watchdog thread */
420         if (!p) {
421                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
422                 if (IS_ERR(p)) {
423                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
424                         return PTR_ERR(p);
425                 }
426                 kthread_bind(p, cpu);
427                 per_cpu(watchdog_touch_ts, cpu) = 0;
428                 per_cpu(softlockup_watchdog, cpu) = p;
429                 wake_up_process(p);
430         }
431
432         /* if any cpu succeeds, watchdog is considered enabled for the system */
433         watchdog_enabled = 1;
434
435         return 0;
436 }
437
438 static void watchdog_disable(int cpu)
439 {
440         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
441         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
442
443         /*
444          * cancel the timer first to stop incrementing the stats
445          * and waking up the kthread
446          */
447         hrtimer_cancel(hrtimer);
448
449         /* disable the perf event */
450         watchdog_nmi_disable(cpu);
451
452         /* stop the watchdog thread */
453         if (p) {
454                 per_cpu(softlockup_watchdog, cpu) = NULL;
455                 kthread_stop(p);
456         }
457 }
458
459 static void watchdog_enable_all_cpus(void)
460 {
461         int cpu;
462         int result = 0;
463
464         for_each_online_cpu(cpu)
465                 result += watchdog_enable(cpu);
466
467         if (result)
468                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
469
470 }
471
472 static void watchdog_disable_all_cpus(void)
473 {
474         int cpu;
475
476         if (no_watchdog)
477                 return;
478
479         for_each_online_cpu(cpu)
480                 watchdog_disable(cpu);
481
482         /* if all watchdogs are disabled, then they are disabled for the system */
483         watchdog_enabled = 0;
484 }
485
486
487 /* sysctl functions */
488 #ifdef CONFIG_SYSCTL
489 /*
490  * proc handler for /proc/sys/kernel/nmi_watchdog
491  */
492
493 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
494                      void __user *buffer, size_t *length, loff_t *ppos)
495 {
496         proc_dointvec(table, write, buffer, length, ppos);
497
498         if (watchdog_enabled)
499                 watchdog_enable_all_cpus();
500         else
501                 watchdog_disable_all_cpus();
502         return 0;
503 }
504
505 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
506                              void __user *buffer,
507                              size_t *lenp, loff_t *ppos)
508 {
509         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
510 }
511 #endif /* CONFIG_SYSCTL */
512
513
514 /*
515  * Create/destroy watchdog threads as CPUs come and go:
516  */
517 static int __cpuinit
518 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
519 {
520         int hotcpu = (unsigned long)hcpu;
521         int err = 0;
522
523         switch (action) {
524         case CPU_UP_PREPARE:
525         case CPU_UP_PREPARE_FROZEN:
526                 err = watchdog_prepare_cpu(hotcpu);
527                 break;
528         case CPU_ONLINE:
529         case CPU_ONLINE_FROZEN:
530                 err = watchdog_enable(hotcpu);
531                 break;
532 #ifdef CONFIG_HOTPLUG_CPU
533         case CPU_UP_CANCELED:
534         case CPU_UP_CANCELED_FROZEN:
535                 watchdog_disable(hotcpu);
536                 break;
537         case CPU_DEAD:
538         case CPU_DEAD_FROZEN:
539                 watchdog_disable(hotcpu);
540                 break;
541 #endif /* CONFIG_HOTPLUG_CPU */
542         }
543         return notifier_from_errno(err);
544 }
545
546 static struct notifier_block __cpuinitdata cpu_nfb = {
547         .notifier_call = cpu_callback
548 };
549
550 static int __init spawn_watchdog_task(void)
551 {
552         void *cpu = (void *)(long)smp_processor_id();
553         int err;
554
555         if (no_watchdog)
556                 return 0;
557
558         err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
559         WARN_ON(notifier_to_errno(err));
560
561         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
562         register_cpu_notifier(&cpu_nfb);
563
564         return 0;
565 }
566 early_initcall(spawn_watchdog_task);