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.
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>
31 int __read_mostly softlockup_thresh
= 60;
33 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
34 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
35 static DEFINE_PER_CPU(struct task_struct
*, softlockup_watchdog
);
36 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
37 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
38 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
39 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
40 #ifdef CONFIG_HARDLOCKUP_DETECTOR
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
);
46 static int __read_mostly did_panic
;
47 static int __initdata no_watchdog
;
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))
63 __setup("nmi_watchdog=", hardlockup_panic_setup
);
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);
75 __setup("softlockup_panic=", softlockup_panic_setup
);
77 static int __init
nowatchdog_setup(char *str
)
82 __setup("nowatchdog", nowatchdog_setup
);
85 static int __init
nosoftlockup_setup(char *str
)
90 __setup("nosoftlockup", nosoftlockup_setup
);
95 * Returns seconds, approximately. We don't need nanosecond
96 * resolution, and we don't need to waste time with a big divide when
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
112 return softlockup_thresh
/ 5 * NSEC_PER_SEC
;
115 /* Commands for resetting the watchdog */
116 static void __touch_watchdog(void)
118 int this_cpu
= raw_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)
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 void touch_nmi_watchdog(void)
144 __get_cpu_var(watchdog_nmi_touch
) = true;
145 touch_softlockup_watchdog();
147 EXPORT_SYMBOL(touch_nmi_watchdog
);
149 void touch_softlockup_watchdog_sync(void)
151 __raw_get_cpu_var(softlockup_touch_sync
) = true;
152 __raw_get_cpu_var(watchdog_touch_ts
) = 0;
155 #ifdef CONFIG_HARDLOCKUP_DETECTOR
156 /* watchdog detector functions */
157 static int is_hardlockup(int cpu
)
159 unsigned long hrint
= per_cpu(hrtimer_interrupts
, cpu
);
161 if (per_cpu(hrtimer_interrupts_saved
, cpu
) == hrint
)
164 per_cpu(hrtimer_interrupts_saved
, cpu
) = hrint
;
169 static int is_softlockup(unsigned long touch_ts
, int cpu
)
171 unsigned long now
= get_timestamp(cpu
);
173 /* Warn about unreasonable delays: */
174 if (time_after(now
, touch_ts
+ softlockup_thresh
))
175 return now
- touch_ts
;
181 watchdog_panic(struct notifier_block
*this, unsigned long event
, void *ptr
)
188 static struct notifier_block panic_block
= {
189 .notifier_call
= watchdog_panic
,
192 #ifdef CONFIG_HARDLOCKUP_DETECTOR
193 static struct perf_event_attr wd_hw_attr
= {
194 .type
= PERF_TYPE_HARDWARE
,
195 .config
= PERF_COUNT_HW_CPU_CYCLES
,
196 .size
= sizeof(struct perf_event_attr
),
201 /* Callback function for perf event subsystem */
202 void watchdog_overflow_callback(struct perf_event
*event
, int nmi
,
203 struct perf_sample_data
*data
,
204 struct pt_regs
*regs
)
206 int this_cpu
= smp_processor_id();
208 if (__get_cpu_var(watchdog_nmi_touch
) == true) {
209 __get_cpu_var(watchdog_nmi_touch
) = false;
213 /* check for a hardlockup
214 * This is done by making sure our timer interrupt
215 * is incrementing. The timer interrupt should have
216 * fired multiple times before we overflow'd. If it hasn't
217 * then this is a good indication the cpu is stuck
219 if (is_hardlockup(this_cpu
)) {
220 /* only print hardlockups once */
221 if (__get_cpu_var(hard_watchdog_warn
) == true)
224 if (hardlockup_panic
)
225 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
227 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
229 __get_cpu_var(hard_watchdog_warn
) = true;
233 __get_cpu_var(hard_watchdog_warn
) = false;
236 static void watchdog_interrupt_count(void)
238 __get_cpu_var(hrtimer_interrupts
)++;
241 static inline void watchdog_interrupt_count(void) { return; }
242 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
244 /* watchdog kicker functions */
245 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
247 int this_cpu
= smp_processor_id();
248 unsigned long touch_ts
= __get_cpu_var(watchdog_touch_ts
);
249 struct pt_regs
*regs
= get_irq_regs();
252 /* kick the hardlockup detector */
253 watchdog_interrupt_count();
255 /* kick the softlockup detector */
256 wake_up_process(__get_cpu_var(softlockup_watchdog
));
259 hrtimer_forward_now(hrtimer
, ns_to_ktime(get_sample_period()));
262 if (unlikely(per_cpu(softlockup_touch_sync
, this_cpu
))) {
264 * If the time stamp was touched atomically
265 * make sure the scheduler tick is up to date.
267 per_cpu(softlockup_touch_sync
, this_cpu
) = false;
271 return HRTIMER_RESTART
;
274 /* check for a softlockup
275 * This is done by making sure a high priority task is
276 * being scheduled. The task touches the watchdog to
277 * indicate it is getting cpu time. If it hasn't then
278 * this is a good indication some task is hogging the cpu
280 duration
= is_softlockup(touch_ts
, this_cpu
);
281 if (unlikely(duration
)) {
283 if (__get_cpu_var(soft_watchdog_warn
) == true)
284 return HRTIMER_RESTART
;
286 printk(KERN_ERR
"BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
288 current
->comm
, task_pid_nr(current
));
290 print_irqtrace_events(current
);
296 if (softlockup_panic
)
297 panic("softlockup: hung tasks");
298 __get_cpu_var(soft_watchdog_warn
) = true;
300 __get_cpu_var(soft_watchdog_warn
) = false;
302 return HRTIMER_RESTART
;
307 * The watchdog thread - touches the timestamp.
309 static int watchdog(void *__bind_cpu
)
311 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
-1 };
312 struct hrtimer
*hrtimer
= &per_cpu(watchdog_hrtimer
, (unsigned long)__bind_cpu
);
314 sched_setscheduler(current
, SCHED_FIFO
, ¶m
);
316 /* initialize timestamp */
319 /* kick off the timer for the hardlockup detector */
320 /* done here because hrtimer_start can only pin to smp_processor_id() */
321 hrtimer_start(hrtimer
, ns_to_ktime(get_sample_period()),
322 HRTIMER_MODE_REL_PINNED
);
324 set_current_state(TASK_INTERRUPTIBLE
);
326 * Run briefly once per second to reset the softlockup timestamp.
327 * If this gets delayed for more than 60 seconds then the
328 * debug-printout triggers in softlockup_tick().
330 while (!kthread_should_stop()) {
334 if (kthread_should_stop())
337 set_current_state(TASK_INTERRUPTIBLE
);
339 __set_current_state(TASK_RUNNING
);
345 #ifdef CONFIG_HARDLOCKUP_DETECTOR
346 static int watchdog_nmi_enable(int cpu
)
348 struct perf_event_attr
*wd_attr
;
349 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
351 /* is it already setup and enabled? */
352 if (event
&& event
->state
> PERF_EVENT_STATE_OFF
)
355 /* it is setup but not enabled */
359 /* Try to register using hardware perf events */
360 wd_attr
= &wd_hw_attr
;
361 wd_attr
->sample_period
= hw_nmi_get_sample_period();
362 event
= perf_event_create_kernel_counter(wd_attr
, cpu
, -1, watchdog_overflow_callback
);
363 if (!IS_ERR(event
)) {
364 printk(KERN_INFO
"NMI watchdog enabled, takes one hw-pmu counter.\n");
368 printk(KERN_ERR
"NMI watchdog failed to create perf event on cpu%i: %p\n", cpu
, event
);
373 per_cpu(watchdog_ev
, cpu
) = event
;
375 perf_event_enable(per_cpu(watchdog_ev
, cpu
));
380 static void watchdog_nmi_disable(int cpu
)
382 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
385 perf_event_disable(event
);
386 per_cpu(watchdog_ev
, cpu
) = NULL
;
388 /* should be in cleanup, but blocks oprofile */
389 perf_event_release_kernel(event
);
394 static int watchdog_nmi_enable(int cpu
) { return 0; }
395 static void watchdog_nmi_disable(int cpu
) { return; }
396 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
398 /* prepare/enable/disable routines */
399 static int watchdog_prepare_cpu(int cpu
)
401 struct hrtimer
*hrtimer
= &per_cpu(watchdog_hrtimer
, cpu
);
403 WARN_ON(per_cpu(softlockup_watchdog
, cpu
));
404 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
405 hrtimer
->function
= watchdog_timer_fn
;
410 static int watchdog_enable(int cpu
)
412 struct task_struct
*p
= per_cpu(softlockup_watchdog
, cpu
);
414 /* enable the perf event */
415 if (watchdog_nmi_enable(cpu
) != 0)
418 /* create the watchdog thread */
420 p
= kthread_create(watchdog
, (void *)(unsigned long)cpu
, "watchdog/%d", cpu
);
422 printk(KERN_ERR
"softlockup watchdog for %i failed\n", cpu
);
425 kthread_bind(p
, cpu
);
426 per_cpu(watchdog_touch_ts
, cpu
) = 0;
427 per_cpu(softlockup_watchdog
, cpu
) = p
;
434 static void watchdog_disable(int cpu
)
436 struct task_struct
*p
= per_cpu(softlockup_watchdog
, cpu
);
437 struct hrtimer
*hrtimer
= &per_cpu(watchdog_hrtimer
, cpu
);
440 * cancel the timer first to stop incrementing the stats
441 * and waking up the kthread
443 hrtimer_cancel(hrtimer
);
445 /* disable the perf event */
446 watchdog_nmi_disable(cpu
);
448 /* stop the watchdog thread */
450 per_cpu(softlockup_watchdog
, cpu
) = NULL
;
454 /* if any cpu succeeds, watchdog is considered enabled for the system */
455 watchdog_enabled
= 1;
458 static void watchdog_enable_all_cpus(void)
463 for_each_online_cpu(cpu
)
464 result
+= watchdog_enable(cpu
);
467 printk(KERN_ERR
"watchdog: failed to be enabled on some cpus\n");
471 static void watchdog_disable_all_cpus(void)
475 for_each_online_cpu(cpu
)
476 watchdog_disable(cpu
);
478 /* if all watchdogs are disabled, then they are disabled for the system */
479 watchdog_enabled
= 0;
483 /* sysctl functions */
486 * proc handler for /proc/sys/kernel/nmi_watchdog
489 int proc_dowatchdog_enabled(struct ctl_table
*table
, int write
,
490 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
492 proc_dointvec(table
, write
, buffer
, length
, ppos
);
494 if (watchdog_enabled
)
495 watchdog_enable_all_cpus();
497 watchdog_disable_all_cpus();
501 int proc_dowatchdog_thresh(struct ctl_table
*table
, int write
,
503 size_t *lenp
, loff_t
*ppos
)
505 return proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
507 #endif /* CONFIG_SYSCTL */
511 * Create/destroy watchdog threads as CPUs come and go:
514 cpu_callback(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
516 int hotcpu
= (unsigned long)hcpu
;
520 case CPU_UP_PREPARE_FROZEN
:
521 if (watchdog_prepare_cpu(hotcpu
))
525 case CPU_ONLINE_FROZEN
:
526 if (watchdog_enable(hotcpu
))
529 #ifdef CONFIG_HOTPLUG_CPU
530 case CPU_UP_CANCELED
:
531 case CPU_UP_CANCELED_FROZEN
:
532 watchdog_disable(hotcpu
);
535 case CPU_DEAD_FROZEN
:
536 watchdog_disable(hotcpu
);
538 #endif /* CONFIG_HOTPLUG_CPU */
543 static struct notifier_block __cpuinitdata cpu_nfb
= {
544 .notifier_call
= cpu_callback
547 static int __init
spawn_watchdog_task(void)
549 void *cpu
= (void *)(long)smp_processor_id();
555 err
= cpu_callback(&cpu_nfb
, CPU_UP_PREPARE
, cpu
);
556 WARN_ON(err
== NOTIFY_BAD
);
558 cpu_callback(&cpu_nfb
, CPU_ONLINE
, cpu
);
559 register_cpu_notifier(&cpu_nfb
);
561 atomic_notifier_chain_register(&panic_notifier_list
, &panic_block
);
565 early_initcall(spawn_watchdog_task
);