drm/i915: Enable MI_FLUSH on Sandybridge
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / kernel / watchdog.c
blob0d53c8e853b12450cf0c74665d13a22e91a47543
1 /*
2 * Detect hard and soft lockups on a system
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 * this code detects hard lockups: incidents in where on a CPU
7 * the kernel does not respond to anything except NMI.
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.
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>
27 #include <asm/irq_regs.h>
28 #include <linux/perf_event.h>
30 int watchdog_enabled;
31 int __read_mostly softlockup_thresh = 60;
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
46 static int __read_mostly did_panic;
47 static int __initdata no_watchdog;
50 /* boot commands */
52 * Should we panic when a soft-lockup or hard-lockup occurs:
54 #ifdef CONFIG_HARDLOCKUP_DETECTOR
55 static int hardlockup_panic;
57 static int __init hardlockup_panic_setup(char *str)
59 if (!strncmp(str, "panic", 5))
60 hardlockup_panic = 1;
61 return 1;
63 __setup("nmi_watchdog=", hardlockup_panic_setup);
64 #endif
66 unsigned int __read_mostly softlockup_panic =
67 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
69 static int __init softlockup_panic_setup(char *str)
71 softlockup_panic = simple_strtoul(str, NULL, 0);
73 return 1;
75 __setup("softlockup_panic=", softlockup_panic_setup);
77 static int __init nowatchdog_setup(char *str)
79 no_watchdog = 1;
80 return 1;
82 __setup("nowatchdog", nowatchdog_setup);
84 /* deprecated */
85 static int __init nosoftlockup_setup(char *str)
87 no_watchdog = 1;
88 return 1;
90 __setup("nosoftlockup", nosoftlockup_setup);
91 /* */
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.
99 static unsigned long get_timestamp(int this_cpu)
101 return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */
104 static unsigned long get_sample_period(void)
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
112 return softlockup_thresh / 5 * NSEC_PER_SEC;
115 /* Commands for resetting the watchdog */
116 static void __touch_watchdog(void)
118 int this_cpu = smp_processor_id();
120 __get_cpu_var(watchdog_touch_ts) = get_timestamp(this_cpu);
123 void touch_softlockup_watchdog(void)
125 __get_cpu_var(watchdog_touch_ts) = 0;
127 EXPORT_SYMBOL(touch_softlockup_watchdog);
129 void touch_all_softlockup_watchdogs(void)
131 int cpu;
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
138 for_each_online_cpu(cpu)
139 per_cpu(watchdog_touch_ts, cpu) = 0;
142 #ifdef CONFIG_HARDLOCKUP_DETECTOR
143 void touch_nmi_watchdog(void)
145 __get_cpu_var(watchdog_nmi_touch) = true;
146 touch_softlockup_watchdog();
148 EXPORT_SYMBOL(touch_nmi_watchdog);
150 #endif
152 void touch_softlockup_watchdog_sync(void)
154 __raw_get_cpu_var(softlockup_touch_sync) = true;
155 __raw_get_cpu_var(watchdog_touch_ts) = 0;
158 #ifdef CONFIG_HARDLOCKUP_DETECTOR
159 /* watchdog detector functions */
160 static int is_hardlockup(void)
162 unsigned long hrint = __get_cpu_var(hrtimer_interrupts);
164 if (__get_cpu_var(hrtimer_interrupts_saved) == hrint)
165 return 1;
167 __get_cpu_var(hrtimer_interrupts_saved) = hrint;
168 return 0;
170 #endif
172 static int is_softlockup(unsigned long touch_ts)
174 unsigned long now = get_timestamp(smp_processor_id());
176 /* Warn about unreasonable delays: */
177 if (time_after(now, touch_ts + softlockup_thresh))
178 return now - touch_ts;
180 return 0;
183 static int
184 watchdog_panic(struct notifier_block *this, unsigned long event, void *ptr)
186 did_panic = 1;
188 return NOTIFY_DONE;
191 static struct notifier_block panic_block = {
192 .notifier_call = watchdog_panic,
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,
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)
209 /* Ensure the watchdog never gets throttled */
210 event->hw.interrupts = 0;
212 if (__get_cpu_var(watchdog_nmi_touch) == true) {
213 __get_cpu_var(watchdog_nmi_touch) = false;
214 return;
217 /* check for a hardlockup
218 * This is done by making sure our timer interrupt
219 * is incrementing. The timer interrupt should have
220 * fired multiple times before we overflow'd. If it hasn't
221 * then this is a good indication the cpu is stuck
223 if (is_hardlockup()) {
224 int this_cpu = smp_processor_id();
226 /* only print hardlockups once */
227 if (__get_cpu_var(hard_watchdog_warn) == true)
228 return;
230 if (hardlockup_panic)
231 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
232 else
233 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
235 __get_cpu_var(hard_watchdog_warn) = true;
236 return;
239 __get_cpu_var(hard_watchdog_warn) = false;
240 return;
242 static void watchdog_interrupt_count(void)
244 __get_cpu_var(hrtimer_interrupts)++;
246 #else
247 static inline void watchdog_interrupt_count(void) { return; }
248 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
250 /* watchdog kicker functions */
251 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
253 unsigned long touch_ts = __get_cpu_var(watchdog_touch_ts);
254 struct pt_regs *regs = get_irq_regs();
255 int duration;
257 /* kick the hardlockup detector */
258 watchdog_interrupt_count();
260 /* kick the softlockup detector */
261 wake_up_process(__get_cpu_var(softlockup_watchdog));
263 /* .. and repeat */
264 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
266 if (touch_ts == 0) {
267 if (unlikely(__get_cpu_var(softlockup_touch_sync))) {
269 * If the time stamp was touched atomically
270 * make sure the scheduler tick is up to date.
272 __get_cpu_var(softlockup_touch_sync) = false;
273 sched_clock_tick();
275 __touch_watchdog();
276 return HRTIMER_RESTART;
279 /* check for a softlockup
280 * This is done by making sure a high priority task is
281 * being scheduled. The task touches the watchdog to
282 * indicate it is getting cpu time. If it hasn't then
283 * this is a good indication some task is hogging the cpu
285 duration = is_softlockup(touch_ts);
286 if (unlikely(duration)) {
287 /* only warn once */
288 if (__get_cpu_var(soft_watchdog_warn) == true)
289 return HRTIMER_RESTART;
291 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
292 smp_processor_id(), duration,
293 current->comm, task_pid_nr(current));
294 print_modules();
295 print_irqtrace_events(current);
296 if (regs)
297 show_regs(regs);
298 else
299 dump_stack();
301 if (softlockup_panic)
302 panic("softlockup: hung tasks");
303 __get_cpu_var(soft_watchdog_warn) = true;
304 } else
305 __get_cpu_var(soft_watchdog_warn) = false;
307 return HRTIMER_RESTART;
312 * The watchdog thread - touches the timestamp.
314 static int watchdog(void *unused)
316 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
317 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
319 sched_setscheduler(current, SCHED_FIFO, &param);
321 /* initialize timestamp */
322 __touch_watchdog();
324 /* kick off the timer for the hardlockup detector */
325 /* done here because hrtimer_start can only pin to smp_processor_id() */
326 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
327 HRTIMER_MODE_REL_PINNED);
329 set_current_state(TASK_INTERRUPTIBLE);
331 * Run briefly once per second to reset the softlockup timestamp.
332 * If this gets delayed for more than 60 seconds then the
333 * debug-printout triggers in watchdog_timer_fn().
335 while (!kthread_should_stop()) {
336 __touch_watchdog();
337 schedule();
339 if (kthread_should_stop())
340 break;
342 set_current_state(TASK_INTERRUPTIBLE);
344 __set_current_state(TASK_RUNNING);
346 return 0;
350 #ifdef CONFIG_HARDLOCKUP_DETECTOR
351 static int watchdog_nmi_enable(int cpu)
353 struct perf_event_attr *wd_attr;
354 struct perf_event *event = per_cpu(watchdog_ev, cpu);
356 /* is it already setup and enabled? */
357 if (event && event->state > PERF_EVENT_STATE_OFF)
358 goto out;
360 /* it is setup but not enabled */
361 if (event != NULL)
362 goto out_enable;
364 /* Try to register using hardware perf events */
365 wd_attr = &wd_hw_attr;
366 wd_attr->sample_period = hw_nmi_get_sample_period();
367 event = perf_event_create_kernel_counter(wd_attr, cpu, -1, watchdog_overflow_callback);
368 if (!IS_ERR(event)) {
369 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
370 goto out_save;
373 printk(KERN_ERR "NMI watchdog failed to create perf event on cpu%i: %p\n", cpu, event);
374 return -1;
376 /* success path */
377 out_save:
378 per_cpu(watchdog_ev, cpu) = event;
379 out_enable:
380 perf_event_enable(per_cpu(watchdog_ev, cpu));
381 out:
382 return 0;
385 static void watchdog_nmi_disable(int cpu)
387 struct perf_event *event = per_cpu(watchdog_ev, cpu);
389 if (event) {
390 perf_event_disable(event);
391 per_cpu(watchdog_ev, cpu) = NULL;
393 /* should be in cleanup, but blocks oprofile */
394 perf_event_release_kernel(event);
396 return;
398 #else
399 static int watchdog_nmi_enable(int cpu) { return 0; }
400 static void watchdog_nmi_disable(int cpu) { return; }
401 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
403 /* prepare/enable/disable routines */
404 static int watchdog_prepare_cpu(int cpu)
406 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
408 WARN_ON(per_cpu(softlockup_watchdog, cpu));
409 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
410 hrtimer->function = watchdog_timer_fn;
412 return 0;
415 static int watchdog_enable(int cpu)
417 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
419 /* enable the perf event */
420 if (watchdog_nmi_enable(cpu) != 0)
421 return -1;
423 /* create the watchdog thread */
424 if (!p) {
425 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
426 if (IS_ERR(p)) {
427 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
428 return -1;
430 kthread_bind(p, cpu);
431 per_cpu(watchdog_touch_ts, cpu) = 0;
432 per_cpu(softlockup_watchdog, cpu) = p;
433 wake_up_process(p);
436 return 0;
439 static void watchdog_disable(int cpu)
441 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
442 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
445 * cancel the timer first to stop incrementing the stats
446 * and waking up the kthread
448 hrtimer_cancel(hrtimer);
450 /* disable the perf event */
451 watchdog_nmi_disable(cpu);
453 /* stop the watchdog thread */
454 if (p) {
455 per_cpu(softlockup_watchdog, cpu) = NULL;
456 kthread_stop(p);
459 /* if any cpu succeeds, watchdog is considered enabled for the system */
460 watchdog_enabled = 1;
463 static void watchdog_enable_all_cpus(void)
465 int cpu;
466 int result = 0;
468 for_each_online_cpu(cpu)
469 result += watchdog_enable(cpu);
471 if (result)
472 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
476 static void watchdog_disable_all_cpus(void)
478 int cpu;
480 for_each_online_cpu(cpu)
481 watchdog_disable(cpu);
483 /* if all watchdogs are disabled, then they are disabled for the system */
484 watchdog_enabled = 0;
488 /* sysctl functions */
489 #ifdef CONFIG_SYSCTL
491 * proc handler for /proc/sys/kernel/nmi_watchdog
494 int proc_dowatchdog_enabled(struct ctl_table *table, int write,
495 void __user *buffer, size_t *length, loff_t *ppos)
497 proc_dointvec(table, write, buffer, length, ppos);
499 if (watchdog_enabled)
500 watchdog_enable_all_cpus();
501 else
502 watchdog_disable_all_cpus();
503 return 0;
506 int proc_dowatchdog_thresh(struct ctl_table *table, int write,
507 void __user *buffer,
508 size_t *lenp, loff_t *ppos)
510 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
512 #endif /* CONFIG_SYSCTL */
516 * Create/destroy watchdog threads as CPUs come and go:
518 static int __cpuinit
519 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
521 int hotcpu = (unsigned long)hcpu;
523 switch (action) {
524 case CPU_UP_PREPARE:
525 case CPU_UP_PREPARE_FROZEN:
526 if (watchdog_prepare_cpu(hotcpu))
527 return NOTIFY_BAD;
528 break;
529 case CPU_ONLINE:
530 case CPU_ONLINE_FROZEN:
531 if (watchdog_enable(hotcpu))
532 return NOTIFY_BAD;
533 break;
534 #ifdef CONFIG_HOTPLUG_CPU
535 case CPU_UP_CANCELED:
536 case CPU_UP_CANCELED_FROZEN:
537 watchdog_disable(hotcpu);
538 break;
539 case CPU_DEAD:
540 case CPU_DEAD_FROZEN:
541 watchdog_disable(hotcpu);
542 break;
543 #endif /* CONFIG_HOTPLUG_CPU */
545 return NOTIFY_OK;
548 static struct notifier_block __cpuinitdata cpu_nfb = {
549 .notifier_call = cpu_callback
552 static int __init spawn_watchdog_task(void)
554 void *cpu = (void *)(long)smp_processor_id();
555 int err;
557 if (no_watchdog)
558 return 0;
560 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
561 WARN_ON(err == NOTIFY_BAD);
563 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
564 register_cpu_notifier(&cpu_nfb);
566 atomic_notifier_chain_register(&panic_notifier_list, &panic_block);
568 return 0;
570 early_initcall(spawn_watchdog_task);