lockup_detector: Cross arch compile fixes
[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 = raw_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(int cpu)
161 {
162         unsigned long hrint = per_cpu(hrtimer_interrupts, cpu);
163
164         if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
165                 return 1;
166
167         per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
168         return 0;
169 }
170 #endif
171
172 static int is_softlockup(unsigned long touch_ts, int cpu)
173 {
174         unsigned long now = get_timestamp(cpu);
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         int this_cpu = smp_processor_id();
210
211         if (__get_cpu_var(watchdog_nmi_touch) == true) {
212                 __get_cpu_var(watchdog_nmi_touch) = false;
213                 return;
214         }
215
216         /* check for a hardlockup
217          * This is done by making sure our timer interrupt
218          * is incrementing.  The timer interrupt should have
219          * fired multiple times before we overflow'd.  If it hasn't
220          * then this is a good indication the cpu is stuck
221          */
222         if (is_hardlockup(this_cpu)) {
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         int this_cpu = smp_processor_id();
251         unsigned long touch_ts = __get_cpu_var(watchdog_touch_ts);
252         struct pt_regs *regs = get_irq_regs();
253         int duration;
254
255         /* kick the hardlockup detector */
256         watchdog_interrupt_count();
257
258         /* kick the softlockup detector */
259         wake_up_process(__get_cpu_var(softlockup_watchdog));
260
261         /* .. and repeat */
262         hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
263
264         if (touch_ts == 0) {
265                 if (unlikely(per_cpu(softlockup_touch_sync, this_cpu))) {
266                         /*
267                          * If the time stamp was touched atomically
268                          * make sure the scheduler tick is up to date.
269                          */
270                         per_cpu(softlockup_touch_sync, this_cpu) = false;
271                         sched_clock_tick();
272                 }
273                 __touch_watchdog();
274                 return HRTIMER_RESTART;
275         }
276
277         /* check for a softlockup
278          * This is done by making sure a high priority task is
279          * being scheduled.  The task touches the watchdog to
280          * indicate it is getting cpu time.  If it hasn't then
281          * this is a good indication some task is hogging the cpu
282          */
283         duration = is_softlockup(touch_ts, this_cpu);
284         if (unlikely(duration)) {
285                 /* only warn once */
286                 if (__get_cpu_var(soft_watchdog_warn) == true)
287                         return HRTIMER_RESTART;
288
289                 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
290                         this_cpu, duration,
291                         current->comm, task_pid_nr(current));
292                 print_modules();
293                 print_irqtrace_events(current);
294                 if (regs)
295                         show_regs(regs);
296                 else
297                         dump_stack();
298
299                 if (softlockup_panic)
300                         panic("softlockup: hung tasks");
301                 __get_cpu_var(soft_watchdog_warn) = true;
302         } else
303                 __get_cpu_var(soft_watchdog_warn) = false;
304
305         return HRTIMER_RESTART;
306 }
307
308
309 /*
310  * The watchdog thread - touches the timestamp.
311  */
312 static int watchdog(void *__bind_cpu)
313 {
314         struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
315         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, (unsigned long)__bind_cpu);
316
317         sched_setscheduler(current, SCHED_FIFO, &param);
318
319         /* initialize timestamp */
320         __touch_watchdog();
321
322         /* kick off the timer for the hardlockup detector */
323         /* done here because hrtimer_start can only pin to smp_processor_id() */
324         hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
325                       HRTIMER_MODE_REL_PINNED);
326
327         set_current_state(TASK_INTERRUPTIBLE);
328         /*
329          * Run briefly once per second to reset the softlockup timestamp.
330          * If this gets delayed for more than 60 seconds then the
331          * debug-printout triggers in softlockup_tick().
332          */
333         while (!kthread_should_stop()) {
334                 __touch_watchdog();
335                 schedule();
336
337                 if (kthread_should_stop())
338                         break;
339
340                 set_current_state(TASK_INTERRUPTIBLE);
341         }
342         __set_current_state(TASK_RUNNING);
343
344         return 0;
345 }
346
347
348 #ifdef CONFIG_HARDLOCKUP_DETECTOR
349 static int watchdog_nmi_enable(int cpu)
350 {
351         struct perf_event_attr *wd_attr;
352         struct perf_event *event = per_cpu(watchdog_ev, cpu);
353
354         /* is it already setup and enabled? */
355         if (event && event->state > PERF_EVENT_STATE_OFF)
356                 goto out;
357
358         /* it is setup but not enabled */
359         if (event != NULL)
360                 goto out_enable;
361
362         /* Try to register using hardware perf events */
363         wd_attr = &wd_hw_attr;
364         wd_attr->sample_period = hw_nmi_get_sample_period();
365         event = perf_event_create_kernel_counter(wd_attr, cpu, -1, watchdog_overflow_callback);
366         if (!IS_ERR(event)) {
367                 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
368                 goto out_save;
369         }
370
371         printk(KERN_ERR "NMI watchdog failed to create perf event on cpu%i: %p\n", cpu, event);
372         return -1;
373
374         /* success path */
375 out_save:
376         per_cpu(watchdog_ev, cpu) = event;
377 out_enable:
378         perf_event_enable(per_cpu(watchdog_ev, cpu));
379 out:
380         return 0;
381 }
382
383 static void watchdog_nmi_disable(int cpu)
384 {
385         struct perf_event *event = per_cpu(watchdog_ev, cpu);
386
387         if (event) {
388                 perf_event_disable(event);
389                 per_cpu(watchdog_ev, cpu) = NULL;
390
391                 /* should be in cleanup, but blocks oprofile */
392                 perf_event_release_kernel(event);
393         }
394         return;
395 }
396 #else
397 static int watchdog_nmi_enable(int cpu) { return 0; }
398 static void watchdog_nmi_disable(int cpu) { return; }
399 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
400
401 /* prepare/enable/disable routines */
402 static int watchdog_prepare_cpu(int cpu)
403 {
404         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
405
406         WARN_ON(per_cpu(softlockup_watchdog, cpu));
407         hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
408         hrtimer->function = watchdog_timer_fn;
409
410         return 0;
411 }
412
413 static int watchdog_enable(int cpu)
414 {
415         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
416
417         /* enable the perf event */
418         if (watchdog_nmi_enable(cpu) != 0)
419                 return -1;
420
421         /* create the watchdog thread */
422         if (!p) {
423                 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
424                 if (IS_ERR(p)) {
425                         printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
426                         return -1;
427                 }
428                 kthread_bind(p, cpu);
429                 per_cpu(watchdog_touch_ts, cpu) = 0;
430                 per_cpu(softlockup_watchdog, cpu) = p;
431                 wake_up_process(p);
432         }
433
434         return 0;
435 }
436
437 static void watchdog_disable(int cpu)
438 {
439         struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
440         struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
441
442         /*
443          * cancel the timer first to stop incrementing the stats
444          * and waking up the kthread
445          */
446         hrtimer_cancel(hrtimer);
447
448         /* disable the perf event */
449         watchdog_nmi_disable(cpu);
450
451         /* stop the watchdog thread */
452         if (p) {
453                 per_cpu(softlockup_watchdog, cpu) = NULL;
454                 kthread_stop(p);
455         }
456
457         /* if any cpu succeeds, watchdog is considered enabled for the system */
458         watchdog_enabled = 1;
459 }
460
461 static void watchdog_enable_all_cpus(void)
462 {
463         int cpu;
464         int result;
465
466         for_each_online_cpu(cpu)
467                 result += watchdog_enable(cpu);
468
469         if (result)
470                 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
471
472 }
473
474 static void watchdog_disable_all_cpus(void)
475 {
476         int cpu;
477
478         for_each_online_cpu(cpu)
479                 watchdog_disable(cpu);
480
481         /* if all watchdogs are disabled, then they are disabled for the system */
482         watchdog_enabled = 0;
483 }
484
485
486 /* sysctl functions */
487 #ifdef CONFIG_SYSCTL
488 /*
489  * proc handler for /proc/sys/kernel/nmi_watchdog
490  */
491
492 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
493                      void __user *buffer, size_t *length, loff_t *ppos)
494 {
495         proc_dointvec(table, write, buffer, length, ppos);
496
497         if (watchdog_enabled)
498                 watchdog_enable_all_cpus();
499         else
500                 watchdog_disable_all_cpus();
501         return 0;
502 }
503
504 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
505                              void __user *buffer,
506                              size_t *lenp, loff_t *ppos)
507 {
508         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
509 }
510 #endif /* CONFIG_SYSCTL */
511
512
513 /*
514  * Create/destroy watchdog threads as CPUs come and go:
515  */
516 static int __cpuinit
517 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
518 {
519         int hotcpu = (unsigned long)hcpu;
520
521         switch (action) {
522         case CPU_UP_PREPARE:
523         case CPU_UP_PREPARE_FROZEN:
524                 if (watchdog_prepare_cpu(hotcpu))
525                         return NOTIFY_BAD;
526                 break;
527         case CPU_ONLINE:
528         case CPU_ONLINE_FROZEN:
529                 if (watchdog_enable(hotcpu))
530                         return NOTIFY_BAD;
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 NOTIFY_OK;
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(err == NOTIFY_BAD);
560
561         cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
562         register_cpu_notifier(&cpu_nfb);
563
564         atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
565
566         return 0;
567 }
568 early_initcall(spawn_watchdog_task);