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