2 * Detect hard and soft lockups on a system
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 * Note: Most of this code is borrowed heavily from the original softlockup
7 * detector, so thanks to Ingo for the initial implementation.
8 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
9 * to those contributors as well.
12 #define pr_fmt(fmt) "NMI watchdog: " fmt
15 #include <linux/cpu.h>
16 #include <linux/nmi.h>
17 #include <linux/init.h>
18 #include <linux/delay.h>
19 #include <linux/freezer.h>
20 #include <linux/kthread.h>
21 #include <linux/lockdep.h>
22 #include <linux/notifier.h>
23 #include <linux/module.h>
24 #include <linux/sysctl.h>
25 #include <linux/smpboot.h>
26 #include <linux/sched/rt.h>
28 #include <asm/irq_regs.h>
29 #include <linux/kvm_para.h>
30 #include <linux/perf_event.h>
32 int watchdog_enabled
= 1;
33 int __read_mostly watchdog_thresh
= 10;
34 static int __read_mostly watchdog_disabled
;
35 static u64 __read_mostly sample_period
;
37 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
38 static DEFINE_PER_CPU(struct task_struct
*, softlockup_watchdog
);
39 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
40 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
41 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
43 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt
);
44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
45 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
46 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
47 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
49 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
50 static cpumask_t __read_mostly watchdog_cpus
;
52 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
53 static DEFINE_PER_CPU(struct perf_event
*, watchdog_ev
);
58 * Should we panic when a soft-lockup or hard-lockup occurs:
60 #ifdef CONFIG_HARDLOCKUP_DETECTOR
61 static int hardlockup_panic
=
62 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
64 static int __init
hardlockup_panic_setup(char *str
)
66 if (!strncmp(str
, "panic", 5))
68 else if (!strncmp(str
, "nopanic", 7))
70 else if (!strncmp(str
, "0", 1))
74 __setup("nmi_watchdog=", hardlockup_panic_setup
);
77 unsigned int __read_mostly softlockup_panic
=
78 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE
;
80 static int __init
softlockup_panic_setup(char *str
)
82 softlockup_panic
= simple_strtoul(str
, NULL
, 0);
86 __setup("softlockup_panic=", softlockup_panic_setup
);
88 static int __init
nowatchdog_setup(char *str
)
93 __setup("nowatchdog", nowatchdog_setup
);
96 static int __init
nosoftlockup_setup(char *str
)
101 __setup("nosoftlockup", nosoftlockup_setup
);
105 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
106 * lockups can have false positives under extreme conditions. So we generally
107 * want a higher threshold for soft lockups than for hard lockups. So we couple
108 * the thresholds with a factor: we make the soft threshold twice the amount of
109 * time the hard threshold is.
111 static int get_softlockup_thresh(void)
113 return watchdog_thresh
* 2;
117 * Returns seconds, approximately. We don't need nanosecond
118 * resolution, and we don't need to waste time with a big divide when
121 static unsigned long get_timestamp(void)
123 return local_clock() >> 30LL; /* 2^30 ~= 10^9 */
126 static void set_sample_period(void)
129 * convert watchdog_thresh from seconds to ns
130 * the divide by 5 is to give hrtimer several chances (two
131 * or three with the current relation between the soft
132 * and hard thresholds) to increment before the
133 * hardlockup detector generates a warning
135 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
138 /* Commands for resetting the watchdog */
139 static void __touch_watchdog(void)
141 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
144 void touch_softlockup_watchdog(void)
146 __this_cpu_write(watchdog_touch_ts
, 0);
148 EXPORT_SYMBOL(touch_softlockup_watchdog
);
150 void touch_all_softlockup_watchdogs(void)
155 * this is done lockless
156 * do we care if a 0 races with a timestamp?
157 * all it means is the softlock check starts one cycle later
159 for_each_online_cpu(cpu
)
160 per_cpu(watchdog_touch_ts
, cpu
) = 0;
163 #ifdef CONFIG_HARDLOCKUP_DETECTOR
164 void touch_nmi_watchdog(void)
166 if (watchdog_enabled
) {
169 for_each_present_cpu(cpu
) {
170 if (per_cpu(watchdog_nmi_touch
, cpu
) != true)
171 per_cpu(watchdog_nmi_touch
, cpu
) = true;
174 touch_softlockup_watchdog();
176 EXPORT_SYMBOL(touch_nmi_watchdog
);
180 void touch_softlockup_watchdog_sync(void)
182 __raw_get_cpu_var(softlockup_touch_sync
) = true;
183 __raw_get_cpu_var(watchdog_touch_ts
) = 0;
186 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
187 /* watchdog detector functions */
188 static int is_hardlockup(void)
190 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
192 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
195 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
200 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
201 static unsigned int watchdog_next_cpu(unsigned int cpu
)
203 cpumask_t cpus
= watchdog_cpus
;
204 unsigned int next_cpu
;
206 next_cpu
= cpumask_next(cpu
, &cpus
);
207 if (next_cpu
>= nr_cpu_ids
)
208 next_cpu
= cpumask_first(&cpus
);
216 static int is_hardlockup_other_cpu(unsigned int cpu
)
218 unsigned long hrint
= per_cpu(hrtimer_interrupts
, cpu
);
220 if (per_cpu(hrtimer_interrupts_saved
, cpu
) == hrint
)
223 per_cpu(hrtimer_interrupts_saved
, cpu
) = hrint
;
227 static void watchdog_check_hardlockup_other_cpu(void)
229 unsigned int next_cpu
;
232 * Test for hardlockups every 3 samples. The sample period is
233 * watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over
234 * watchdog_thresh (over by 20%).
236 if (__this_cpu_read(hrtimer_interrupts
) % 3 != 0)
239 /* check for a hardlockup on the next cpu */
240 next_cpu
= watchdog_next_cpu(smp_processor_id());
241 if (next_cpu
>= nr_cpu_ids
)
246 if (per_cpu(watchdog_nmi_touch
, next_cpu
) == true) {
247 per_cpu(watchdog_nmi_touch
, next_cpu
) = false;
251 if (is_hardlockup_other_cpu(next_cpu
)) {
253 if (per_cpu(hard_watchdog_warn
, next_cpu
) == true)
256 if (hardlockup_panic
)
257 panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu
);
259 WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu
);
261 per_cpu(hard_watchdog_warn
, next_cpu
) = true;
263 per_cpu(hard_watchdog_warn
, next_cpu
) = false;
267 static inline void watchdog_check_hardlockup_other_cpu(void) { return; }
270 static int is_softlockup(unsigned long touch_ts
)
272 unsigned long now
= get_timestamp();
274 /* Warn about unreasonable delays: */
275 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
276 return now
- touch_ts
;
281 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
283 static struct perf_event_attr wd_hw_attr
= {
284 .type
= PERF_TYPE_HARDWARE
,
285 .config
= PERF_COUNT_HW_CPU_CYCLES
,
286 .size
= sizeof(struct perf_event_attr
),
291 /* Callback function for perf event subsystem */
292 static void watchdog_overflow_callback(struct perf_event
*event
,
293 struct perf_sample_data
*data
,
294 struct pt_regs
*regs
)
296 /* Ensure the watchdog never gets throttled */
297 event
->hw
.interrupts
= 0;
299 if (__this_cpu_read(watchdog_nmi_touch
) == true) {
300 __this_cpu_write(watchdog_nmi_touch
, false);
304 /* check for a hardlockup
305 * This is done by making sure our timer interrupt
306 * is incrementing. The timer interrupt should have
307 * fired multiple times before we overflow'd. If it hasn't
308 * then this is a good indication the cpu is stuck
310 if (is_hardlockup()) {
311 int this_cpu
= smp_processor_id();
313 /* only print hardlockups once */
314 if (__this_cpu_read(hard_watchdog_warn
) == true)
317 if (hardlockup_panic
)
318 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
320 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
322 __this_cpu_write(hard_watchdog_warn
, true);
326 __this_cpu_write(hard_watchdog_warn
, false);
329 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
331 static void watchdog_interrupt_count(void)
333 __this_cpu_inc(hrtimer_interrupts
);
336 static int watchdog_nmi_enable(unsigned int cpu
);
337 static void watchdog_nmi_disable(unsigned int cpu
);
339 /* watchdog kicker functions */
340 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
342 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
343 struct pt_regs
*regs
= get_irq_regs();
346 /* kick the hardlockup detector */
347 watchdog_interrupt_count();
349 /* test for hardlockups on the next cpu */
350 watchdog_check_hardlockup_other_cpu();
352 /* kick the softlockup detector */
353 wake_up_process(__this_cpu_read(softlockup_watchdog
));
356 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
359 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
361 * If the time stamp was touched atomically
362 * make sure the scheduler tick is up to date.
364 __this_cpu_write(softlockup_touch_sync
, false);
368 /* Clear the guest paused flag on watchdog reset */
369 kvm_check_and_clear_guest_paused();
371 return HRTIMER_RESTART
;
374 /* check for a softlockup
375 * This is done by making sure a high priority task is
376 * being scheduled. The task touches the watchdog to
377 * indicate it is getting cpu time. If it hasn't then
378 * this is a good indication some task is hogging the cpu
380 duration
= is_softlockup(touch_ts
);
381 if (unlikely(duration
)) {
383 * If a virtual machine is stopped by the host it can look to
384 * the watchdog like a soft lockup, check to see if the host
385 * stopped the vm before we issue the warning
387 if (kvm_check_and_clear_guest_paused())
388 return HRTIMER_RESTART
;
391 if (__this_cpu_read(soft_watchdog_warn
) == true)
392 return HRTIMER_RESTART
;
394 printk(KERN_EMERG
"BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
395 smp_processor_id(), duration
,
396 current
->comm
, task_pid_nr(current
));
398 print_irqtrace_events(current
);
404 if (softlockup_panic
)
405 panic("softlockup: hung tasks");
406 __this_cpu_write(soft_watchdog_warn
, true);
408 __this_cpu_write(soft_watchdog_warn
, false);
410 return HRTIMER_RESTART
;
413 static void watchdog_set_prio(unsigned int policy
, unsigned int prio
)
415 struct sched_param param
= { .sched_priority
= prio
};
417 sched_setscheduler(current
, policy
, ¶m
);
420 static void watchdog_enable(unsigned int cpu
)
422 struct hrtimer
*hrtimer
= &__raw_get_cpu_var(watchdog_hrtimer
);
424 /* kick off the timer for the hardlockup detector */
425 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
426 hrtimer
->function
= watchdog_timer_fn
;
428 if (!watchdog_enabled
) {
429 kthread_park(current
);
433 /* Enable the perf event */
434 watchdog_nmi_enable(cpu
);
436 /* done here because hrtimer_start can only pin to smp_processor_id() */
437 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
438 HRTIMER_MODE_REL_PINNED
);
440 /* initialize timestamp */
441 watchdog_set_prio(SCHED_FIFO
, MAX_RT_PRIO
- 1);
445 static void watchdog_disable(unsigned int cpu
)
447 struct hrtimer
*hrtimer
= &__raw_get_cpu_var(watchdog_hrtimer
);
449 watchdog_set_prio(SCHED_NORMAL
, 0);
450 hrtimer_cancel(hrtimer
);
451 /* disable the perf event */
452 watchdog_nmi_disable(cpu
);
455 static int watchdog_should_run(unsigned int cpu
)
457 return __this_cpu_read(hrtimer_interrupts
) !=
458 __this_cpu_read(soft_lockup_hrtimer_cnt
);
462 * The watchdog thread function - touches the timestamp.
464 * It only runs once every sample_period seconds (4 seconds by
465 * default) to reset the softlockup timestamp. If this gets delayed
466 * for more than 2*watchdog_thresh seconds then the debug-printout
467 * triggers in watchdog_timer_fn().
469 static void watchdog(unsigned int cpu
)
471 __this_cpu_write(soft_lockup_hrtimer_cnt
,
472 __this_cpu_read(hrtimer_interrupts
));
476 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
478 * People like the simple clean cpu node info on boot.
479 * Reduce the watchdog noise by only printing messages
480 * that are different from what cpu0 displayed.
482 static unsigned long cpu0_err
;
484 static int watchdog_nmi_enable(unsigned int cpu
)
486 struct perf_event_attr
*wd_attr
;
487 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
489 /* is it already setup and enabled? */
490 if (event
&& event
->state
> PERF_EVENT_STATE_OFF
)
493 /* it is setup but not enabled */
497 wd_attr
= &wd_hw_attr
;
498 wd_attr
->sample_period
= hw_nmi_get_sample_period(watchdog_thresh
);
500 /* Try to register using hardware perf events */
501 event
= perf_event_create_kernel_counter(wd_attr
, cpu
, NULL
, watchdog_overflow_callback
, NULL
);
503 /* save cpu0 error for future comparision */
504 if (cpu
== 0 && IS_ERR(event
))
505 cpu0_err
= PTR_ERR(event
);
507 if (!IS_ERR(event
)) {
508 /* only print for cpu0 or different than cpu0 */
509 if (cpu
== 0 || cpu0_err
)
510 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
514 /* skip displaying the same error again */
515 if (cpu
> 0 && (PTR_ERR(event
) == cpu0_err
))
516 return PTR_ERR(event
);
518 /* vary the KERN level based on the returned errno */
519 if (PTR_ERR(event
) == -EOPNOTSUPP
)
520 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu
);
521 else if (PTR_ERR(event
) == -ENOENT
)
522 pr_warning("disabled (cpu%i): hardware events not enabled\n",
525 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
526 cpu
, PTR_ERR(event
));
527 return PTR_ERR(event
);
531 per_cpu(watchdog_ev
, cpu
) = event
;
533 perf_event_enable(per_cpu(watchdog_ev
, cpu
));
538 static void watchdog_nmi_disable(unsigned int cpu
)
540 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
543 perf_event_disable(event
);
544 per_cpu(watchdog_ev
, cpu
) = NULL
;
546 /* should be in cleanup, but blocks oprofile */
547 perf_event_release_kernel(event
);
552 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
553 static int watchdog_nmi_enable(unsigned int cpu
)
556 * The new cpu will be marked online before the first hrtimer interrupt
557 * runs on it. If another cpu tests for a hardlockup on the new cpu
558 * before it has run its first hrtimer, it will get a false positive.
559 * Touch the watchdog on the new cpu to delay the first check for at
560 * least 3 sampling periods to guarantee one hrtimer has run on the new
563 per_cpu(watchdog_nmi_touch
, cpu
) = true;
565 cpumask_set_cpu(cpu
, &watchdog_cpus
);
569 static void watchdog_nmi_disable(unsigned int cpu
)
571 unsigned int next_cpu
= watchdog_next_cpu(cpu
);
574 * Offlining this cpu will cause the cpu before this one to start
575 * checking the one after this one. If this cpu just finished checking
576 * the next cpu and updating hrtimer_interrupts_saved, and then the
577 * previous cpu checks it within one sample period, it will trigger a
578 * false positive. Touch the watchdog on the next cpu to prevent it.
580 if (next_cpu
< nr_cpu_ids
)
581 per_cpu(watchdog_nmi_touch
, next_cpu
) = true;
583 cpumask_clear_cpu(cpu
, &watchdog_cpus
);
586 static int watchdog_nmi_enable(unsigned int cpu
) { return 0; }
587 static void watchdog_nmi_disable(unsigned int cpu
) { return; }
588 #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */
589 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
591 /* prepare/enable/disable routines */
592 /* sysctl functions */
594 static void watchdog_enable_all_cpus(void)
598 if (watchdog_disabled
) {
599 watchdog_disabled
= 0;
600 for_each_online_cpu(cpu
)
601 kthread_unpark(per_cpu(softlockup_watchdog
, cpu
));
605 static void watchdog_disable_all_cpus(void)
609 if (!watchdog_disabled
) {
610 watchdog_disabled
= 1;
611 for_each_online_cpu(cpu
)
612 kthread_park(per_cpu(softlockup_watchdog
, cpu
));
617 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
620 int proc_dowatchdog(struct ctl_table
*table
, int write
,
621 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
625 if (watchdog_disabled
< 0)
628 ret
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
634 * Watchdog threads shouldn't be enabled if they are
635 * disabled. The 'watchdog_disabled' variable check in
636 * watchdog_*_all_cpus() function takes care of this.
638 if (watchdog_enabled
&& watchdog_thresh
)
639 watchdog_enable_all_cpus();
641 watchdog_disable_all_cpus();
645 #endif /* CONFIG_SYSCTL */
647 static struct smp_hotplug_thread watchdog_threads
= {
648 .store
= &softlockup_watchdog
,
649 .thread_should_run
= watchdog_should_run
,
650 .thread_fn
= watchdog
,
651 .thread_comm
= "watchdog/%u",
652 .setup
= watchdog_enable
,
653 .park
= watchdog_disable
,
654 .unpark
= watchdog_enable
,
657 void __init
lockup_detector_init(void)
660 if (smpboot_register_percpu_thread(&watchdog_threads
)) {
661 pr_err("Failed to create watchdog threads, disabled\n");
662 watchdog_disabled
= -ENODEV
;