Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
7 * Note: Most of this code is borrowed heavily from the original softlockup
8 * detector, so thanks to Ingo for the initial implementation.
9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
10 * to those contributors as well.
11 */
12
13#define pr_fmt(fmt) "watchdog: " fmt
14
15#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/module.h>
20#include <linux/sysctl.h>
21#include <linux/tick.h>
22#include <linux/sched/clock.h>
23#include <linux/sched/debug.h>
24#include <linux/sched/isolation.h>
25#include <linux/stop_machine.h>
26
27#include <asm/irq_regs.h>
28#include <linux/kvm_para.h>
29
30static DEFINE_MUTEX(watchdog_mutex);
31
32#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
33# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
34# define NMI_WATCHDOG_DEFAULT 1
35#else
36# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37# define NMI_WATCHDOG_DEFAULT 0
38#endif
39
40unsigned long __read_mostly watchdog_enabled;
41int __read_mostly watchdog_user_enabled = 1;
42int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43int __read_mostly soft_watchdog_user_enabled = 1;
44int __read_mostly watchdog_thresh = 10;
45static int __read_mostly nmi_watchdog_available;
46
47struct cpumask watchdog_cpumask __read_mostly;
48unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
49
50#ifdef CONFIG_HARDLOCKUP_DETECTOR
51
52# ifdef CONFIG_SMP
53int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
54# endif /* CONFIG_SMP */
55
56/*
57 * Should we panic when a soft-lockup or hard-lockup occurs:
58 */
59unsigned int __read_mostly hardlockup_panic =
60 IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
61/*
62 * We may not want to enable hard lockup detection by default in all cases,
63 * for example when running the kernel as a guest on a hypervisor. In these
64 * cases this function can be called to disable hard lockup detection. This
65 * function should only be executed once by the boot processor before the
66 * kernel command line parameters are parsed, because otherwise it is not
67 * possible to override this in hardlockup_panic_setup().
68 */
69void __init hardlockup_detector_disable(void)
70{
71 nmi_watchdog_user_enabled = 0;
72}
73
74static int __init hardlockup_panic_setup(char *str)
75{
76 if (!strncmp(str, "panic", 5))
77 hardlockup_panic = 1;
78 else if (!strncmp(str, "nopanic", 7))
79 hardlockup_panic = 0;
80 else if (!strncmp(str, "0", 1))
81 nmi_watchdog_user_enabled = 0;
82 else if (!strncmp(str, "1", 1))
83 nmi_watchdog_user_enabled = 1;
84 return 1;
85}
86__setup("nmi_watchdog=", hardlockup_panic_setup);
87
88#endif /* CONFIG_HARDLOCKUP_DETECTOR */
89
90/*
91 * These functions can be overridden if an architecture implements its
92 * own hardlockup detector.
93 *
94 * watchdog_nmi_enable/disable can be implemented to start and stop when
95 * softlockup watchdog start and stop. The arch must select the
96 * SOFTLOCKUP_DETECTOR Kconfig.
97 */
98int __weak watchdog_nmi_enable(unsigned int cpu)
99{
100 hardlockup_detector_perf_enable();
101 return 0;
102}
103
104void __weak watchdog_nmi_disable(unsigned int cpu)
105{
106 hardlockup_detector_perf_disable();
107}
108
109/* Return 0, if a NMI watchdog is available. Error code otherwise */
110int __weak __init watchdog_nmi_probe(void)
111{
112 return hardlockup_detector_perf_init();
113}
114
115/**
116 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
117 *
118 * The reconfiguration steps are:
119 * watchdog_nmi_stop();
120 * update_variables();
121 * watchdog_nmi_start();
122 */
123void __weak watchdog_nmi_stop(void) { }
124
125/**
126 * watchdog_nmi_start - Start the watchdog after reconfiguration
127 *
128 * Counterpart to watchdog_nmi_stop().
129 *
130 * The following variables have been updated in update_variables() and
131 * contain the currently valid configuration:
132 * - watchdog_enabled
133 * - watchdog_thresh
134 * - watchdog_cpumask
135 */
136void __weak watchdog_nmi_start(void) { }
137
138/**
139 * lockup_detector_update_enable - Update the sysctl enable bit
140 *
141 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
142 * can't race with watchdog_nmi_disable().
143 */
144static void lockup_detector_update_enable(void)
145{
146 watchdog_enabled = 0;
147 if (!watchdog_user_enabled)
148 return;
149 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
150 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
151 if (soft_watchdog_user_enabled)
152 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
153}
154
155#ifdef CONFIG_SOFTLOCKUP_DETECTOR
156
157/*
158 * Delay the soflockup report when running a known slow code.
159 * It does _not_ affect the timestamp of the last successdul reschedule.
160 */
161#define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
162
163#ifdef CONFIG_SMP
164int __read_mostly sysctl_softlockup_all_cpu_backtrace;
165#endif
166
167static struct cpumask watchdog_allowed_mask __read_mostly;
168
169/* Global variables, exported for sysctl */
170unsigned int __read_mostly softlockup_panic =
171 IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
172
173static bool softlockup_initialized __read_mostly;
174static u64 __read_mostly sample_period;
175
176/* Timestamp taken after the last successful reschedule. */
177static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
178/* Timestamp of the last softlockup report. */
179static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
180static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
181static DEFINE_PER_CPU(bool, softlockup_touch_sync);
182static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
183static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
184static unsigned long soft_lockup_nmi_warn;
185
186static int __init nowatchdog_setup(char *str)
187{
188 watchdog_user_enabled = 0;
189 return 1;
190}
191__setup("nowatchdog", nowatchdog_setup);
192
193static int __init nosoftlockup_setup(char *str)
194{
195 soft_watchdog_user_enabled = 0;
196 return 1;
197}
198__setup("nosoftlockup", nosoftlockup_setup);
199
200static int __init watchdog_thresh_setup(char *str)
201{
202 get_option(&str, &watchdog_thresh);
203 return 1;
204}
205__setup("watchdog_thresh=", watchdog_thresh_setup);
206
207static void __lockup_detector_cleanup(void);
208
209/*
210 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
211 * lockups can have false positives under extreme conditions. So we generally
212 * want a higher threshold for soft lockups than for hard lockups. So we couple
213 * the thresholds with a factor: we make the soft threshold twice the amount of
214 * time the hard threshold is.
215 */
216static int get_softlockup_thresh(void)
217{
218 return watchdog_thresh * 2;
219}
220
221/*
222 * Returns seconds, approximately. We don't need nanosecond
223 * resolution, and we don't need to waste time with a big divide when
224 * 2^30ns == 1.074s.
225 */
226static unsigned long get_timestamp(void)
227{
228 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
229}
230
231static void set_sample_period(void)
232{
233 /*
234 * convert watchdog_thresh from seconds to ns
235 * the divide by 5 is to give hrtimer several chances (two
236 * or three with the current relation between the soft
237 * and hard thresholds) to increment before the
238 * hardlockup detector generates a warning
239 */
240 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
241 watchdog_update_hrtimer_threshold(sample_period);
242}
243
244static void update_report_ts(void)
245{
246 __this_cpu_write(watchdog_report_ts, get_timestamp());
247}
248
249/* Commands for resetting the watchdog */
250static void update_touch_ts(void)
251{
252 __this_cpu_write(watchdog_touch_ts, get_timestamp());
253 update_report_ts();
254}
255
256/**
257 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
258 *
259 * Call when the scheduler may have stalled for legitimate reasons
260 * preventing the watchdog task from executing - e.g. the scheduler
261 * entering idle state. This should only be used for scheduler events.
262 * Use touch_softlockup_watchdog() for everything else.
263 */
264notrace void touch_softlockup_watchdog_sched(void)
265{
266 /*
267 * Preemption can be enabled. It doesn't matter which CPU's watchdog
268 * report period gets restarted here, so use the raw_ operation.
269 */
270 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
271}
272
273notrace void touch_softlockup_watchdog(void)
274{
275 touch_softlockup_watchdog_sched();
276 wq_watchdog_touch(raw_smp_processor_id());
277}
278EXPORT_SYMBOL(touch_softlockup_watchdog);
279
280void touch_all_softlockup_watchdogs(void)
281{
282 int cpu;
283
284 /*
285 * watchdog_mutex cannpt be taken here, as this might be called
286 * from (soft)interrupt context, so the access to
287 * watchdog_allowed_cpumask might race with a concurrent update.
288 *
289 * The watchdog time stamp can race against a concurrent real
290 * update as well, the only side effect might be a cycle delay for
291 * the softlockup check.
292 */
293 for_each_cpu(cpu, &watchdog_allowed_mask) {
294 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
295 wq_watchdog_touch(cpu);
296 }
297}
298
299void touch_softlockup_watchdog_sync(void)
300{
301 __this_cpu_write(softlockup_touch_sync, true);
302 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
303}
304
305static int is_softlockup(unsigned long touch_ts,
306 unsigned long period_ts,
307 unsigned long now)
308{
309 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
310 /* Warn about unreasonable delays. */
311 if (time_after(now, period_ts + get_softlockup_thresh()))
312 return now - touch_ts;
313 }
314 return 0;
315}
316
317/* watchdog detector functions */
318bool is_hardlockup(void)
319{
320 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
321
322 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
323 return true;
324
325 __this_cpu_write(hrtimer_interrupts_saved, hrint);
326 return false;
327}
328
329static void watchdog_interrupt_count(void)
330{
331 __this_cpu_inc(hrtimer_interrupts);
332}
333
334static DEFINE_PER_CPU(struct completion, softlockup_completion);
335static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
336
337/*
338 * The watchdog feed function - touches the timestamp.
339 *
340 * It only runs once every sample_period seconds (4 seconds by
341 * default) to reset the softlockup timestamp. If this gets delayed
342 * for more than 2*watchdog_thresh seconds then the debug-printout
343 * triggers in watchdog_timer_fn().
344 */
345static int softlockup_fn(void *data)
346{
347 update_touch_ts();
348 complete(this_cpu_ptr(&softlockup_completion));
349
350 return 0;
351}
352
353/* watchdog kicker functions */
354static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
355{
356 unsigned long touch_ts, period_ts, now;
357 struct pt_regs *regs = get_irq_regs();
358 int duration;
359 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
360
361 if (!watchdog_enabled)
362 return HRTIMER_NORESTART;
363
364 /* kick the hardlockup detector */
365 watchdog_interrupt_count();
366
367 /* kick the softlockup detector */
368 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
369 reinit_completion(this_cpu_ptr(&softlockup_completion));
370 stop_one_cpu_nowait(smp_processor_id(),
371 softlockup_fn, NULL,
372 this_cpu_ptr(&softlockup_stop_work));
373 }
374
375 /* .. and repeat */
376 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
377
378 /*
379 * Read the current timestamp first. It might become invalid anytime
380 * when a virtual machine is stopped by the host or when the watchog
381 * is touched from NMI.
382 */
383 now = get_timestamp();
384 /*
385 * If a virtual machine is stopped by the host it can look to
386 * the watchdog like a soft lockup. This function touches the watchdog.
387 */
388 kvm_check_and_clear_guest_paused();
389 /*
390 * The stored timestamp is comparable with @now only when not touched.
391 * It might get touched anytime from NMI. Make sure that is_softlockup()
392 * uses the same (valid) value.
393 */
394 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
395
396 /* Reset the interval when touched by known problematic code. */
397 if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
398 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
399 /*
400 * If the time stamp was touched atomically
401 * make sure the scheduler tick is up to date.
402 */
403 __this_cpu_write(softlockup_touch_sync, false);
404 sched_clock_tick();
405 }
406
407 update_report_ts();
408 return HRTIMER_RESTART;
409 }
410
411 /* Check for a softlockup. */
412 touch_ts = __this_cpu_read(watchdog_touch_ts);
413 duration = is_softlockup(touch_ts, period_ts, now);
414 if (unlikely(duration)) {
415 /*
416 * Prevent multiple soft-lockup reports if one cpu is already
417 * engaged in dumping all cpu back traces.
418 */
419 if (softlockup_all_cpu_backtrace) {
420 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
421 return HRTIMER_RESTART;
422 }
423
424 /* Start period for the next softlockup warning. */
425 update_report_ts();
426
427 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
428 smp_processor_id(), duration,
429 current->comm, task_pid_nr(current));
430 print_modules();
431 print_irqtrace_events(current);
432 if (regs)
433 show_regs(regs);
434 else
435 dump_stack();
436
437 if (softlockup_all_cpu_backtrace) {
438 trigger_allbutself_cpu_backtrace();
439 clear_bit_unlock(0, &soft_lockup_nmi_warn);
440 }
441
442 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
443 if (softlockup_panic)
444 panic("softlockup: hung tasks");
445 }
446
447 return HRTIMER_RESTART;
448}
449
450static void watchdog_enable(unsigned int cpu)
451{
452 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
453 struct completion *done = this_cpu_ptr(&softlockup_completion);
454
455 WARN_ON_ONCE(cpu != smp_processor_id());
456
457 init_completion(done);
458 complete(done);
459
460 /*
461 * Start the timer first to prevent the NMI watchdog triggering
462 * before the timer has a chance to fire.
463 */
464 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
465 hrtimer->function = watchdog_timer_fn;
466 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
467 HRTIMER_MODE_REL_PINNED_HARD);
468
469 /* Initialize timestamp */
470 update_touch_ts();
471 /* Enable the perf event */
472 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
473 watchdog_nmi_enable(cpu);
474}
475
476static void watchdog_disable(unsigned int cpu)
477{
478 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
479
480 WARN_ON_ONCE(cpu != smp_processor_id());
481
482 /*
483 * Disable the perf event first. That prevents that a large delay
484 * between disabling the timer and disabling the perf event causes
485 * the perf NMI to detect a false positive.
486 */
487 watchdog_nmi_disable(cpu);
488 hrtimer_cancel(hrtimer);
489 wait_for_completion(this_cpu_ptr(&softlockup_completion));
490}
491
492static int softlockup_stop_fn(void *data)
493{
494 watchdog_disable(smp_processor_id());
495 return 0;
496}
497
498static void softlockup_stop_all(void)
499{
500 int cpu;
501
502 if (!softlockup_initialized)
503 return;
504
505 for_each_cpu(cpu, &watchdog_allowed_mask)
506 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
507
508 cpumask_clear(&watchdog_allowed_mask);
509}
510
511static int softlockup_start_fn(void *data)
512{
513 watchdog_enable(smp_processor_id());
514 return 0;
515}
516
517static void softlockup_start_all(void)
518{
519 int cpu;
520
521 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
522 for_each_cpu(cpu, &watchdog_allowed_mask)
523 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
524}
525
526int lockup_detector_online_cpu(unsigned int cpu)
527{
528 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
529 watchdog_enable(cpu);
530 return 0;
531}
532
533int lockup_detector_offline_cpu(unsigned int cpu)
534{
535 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
536 watchdog_disable(cpu);
537 return 0;
538}
539
540static void __lockup_detector_reconfigure(void)
541{
542 cpus_read_lock();
543 watchdog_nmi_stop();
544
545 softlockup_stop_all();
546 set_sample_period();
547 lockup_detector_update_enable();
548 if (watchdog_enabled && watchdog_thresh)
549 softlockup_start_all();
550
551 watchdog_nmi_start();
552 cpus_read_unlock();
553 /*
554 * Must be called outside the cpus locked section to prevent
555 * recursive locking in the perf code.
556 */
557 __lockup_detector_cleanup();
558}
559
560void lockup_detector_reconfigure(void)
561{
562 mutex_lock(&watchdog_mutex);
563 __lockup_detector_reconfigure();
564 mutex_unlock(&watchdog_mutex);
565}
566
567/*
568 * Create the watchdog infrastructure and configure the detector(s).
569 */
570static __init void lockup_detector_setup(void)
571{
572 /*
573 * If sysctl is off and watchdog got disabled on the command line,
574 * nothing to do here.
575 */
576 lockup_detector_update_enable();
577
578 if (!IS_ENABLED(CONFIG_SYSCTL) &&
579 !(watchdog_enabled && watchdog_thresh))
580 return;
581
582 mutex_lock(&watchdog_mutex);
583 __lockup_detector_reconfigure();
584 softlockup_initialized = true;
585 mutex_unlock(&watchdog_mutex);
586}
587
588#else /* CONFIG_SOFTLOCKUP_DETECTOR */
589static void __lockup_detector_reconfigure(void)
590{
591 cpus_read_lock();
592 watchdog_nmi_stop();
593 lockup_detector_update_enable();
594 watchdog_nmi_start();
595 cpus_read_unlock();
596}
597void lockup_detector_reconfigure(void)
598{
599 __lockup_detector_reconfigure();
600}
601static inline void lockup_detector_setup(void)
602{
603 __lockup_detector_reconfigure();
604}
605#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
606
607static void __lockup_detector_cleanup(void)
608{
609 lockdep_assert_held(&watchdog_mutex);
610 hardlockup_detector_perf_cleanup();
611}
612
613/**
614 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
615 *
616 * Caller must not hold the cpu hotplug rwsem.
617 */
618void lockup_detector_cleanup(void)
619{
620 mutex_lock(&watchdog_mutex);
621 __lockup_detector_cleanup();
622 mutex_unlock(&watchdog_mutex);
623}
624
625/**
626 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
627 *
628 * Special interface for parisc. It prevents lockup detector warnings from
629 * the default pm_poweroff() function which busy loops forever.
630 */
631void lockup_detector_soft_poweroff(void)
632{
633 watchdog_enabled = 0;
634}
635
636#ifdef CONFIG_SYSCTL
637
638/* Propagate any changes to the watchdog infrastructure */
639static void proc_watchdog_update(void)
640{
641 /* Remove impossible cpus to keep sysctl output clean. */
642 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
643 __lockup_detector_reconfigure();
644}
645
646/*
647 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
648 *
649 * caller | table->data points to | 'which'
650 * -------------------|----------------------------|--------------------------
651 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
652 * | | SOFT_WATCHDOG_ENABLED
653 * -------------------|----------------------------|--------------------------
654 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
655 * -------------------|----------------------------|--------------------------
656 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
657 */
658static int proc_watchdog_common(int which, struct ctl_table *table, int write,
659 void *buffer, size_t *lenp, loff_t *ppos)
660{
661 int err, old, *param = table->data;
662
663 mutex_lock(&watchdog_mutex);
664
665 if (!write) {
666 /*
667 * On read synchronize the userspace interface. This is a
668 * racy snapshot.
669 */
670 *param = (watchdog_enabled & which) != 0;
671 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
672 } else {
673 old = READ_ONCE(*param);
674 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
675 if (!err && old != READ_ONCE(*param))
676 proc_watchdog_update();
677 }
678 mutex_unlock(&watchdog_mutex);
679 return err;
680}
681
682/*
683 * /proc/sys/kernel/watchdog
684 */
685int proc_watchdog(struct ctl_table *table, int write,
686 void *buffer, size_t *lenp, loff_t *ppos)
687{
688 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
689 table, write, buffer, lenp, ppos);
690}
691
692/*
693 * /proc/sys/kernel/nmi_watchdog
694 */
695int proc_nmi_watchdog(struct ctl_table *table, int write,
696 void *buffer, size_t *lenp, loff_t *ppos)
697{
698 if (!nmi_watchdog_available && write)
699 return -ENOTSUPP;
700 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
701 table, write, buffer, lenp, ppos);
702}
703
704/*
705 * /proc/sys/kernel/soft_watchdog
706 */
707int proc_soft_watchdog(struct ctl_table *table, int write,
708 void *buffer, size_t *lenp, loff_t *ppos)
709{
710 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
711 table, write, buffer, lenp, ppos);
712}
713
714/*
715 * /proc/sys/kernel/watchdog_thresh
716 */
717int proc_watchdog_thresh(struct ctl_table *table, int write,
718 void *buffer, size_t *lenp, loff_t *ppos)
719{
720 int err, old;
721
722 mutex_lock(&watchdog_mutex);
723
724 old = READ_ONCE(watchdog_thresh);
725 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
726
727 if (!err && write && old != READ_ONCE(watchdog_thresh))
728 proc_watchdog_update();
729
730 mutex_unlock(&watchdog_mutex);
731 return err;
732}
733
734/*
735 * The cpumask is the mask of possible cpus that the watchdog can run
736 * on, not the mask of cpus it is actually running on. This allows the
737 * user to specify a mask that will include cpus that have not yet
738 * been brought online, if desired.
739 */
740int proc_watchdog_cpumask(struct ctl_table *table, int write,
741 void *buffer, size_t *lenp, loff_t *ppos)
742{
743 int err;
744
745 mutex_lock(&watchdog_mutex);
746
747 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
748 if (!err && write)
749 proc_watchdog_update();
750
751 mutex_unlock(&watchdog_mutex);
752 return err;
753}
754
755static const int sixty = 60;
756
757static struct ctl_table watchdog_sysctls[] = {
758 {
759 .procname = "watchdog",
760 .data = &watchdog_user_enabled,
761 .maxlen = sizeof(int),
762 .mode = 0644,
763 .proc_handler = proc_watchdog,
764 .extra1 = SYSCTL_ZERO,
765 .extra2 = SYSCTL_ONE,
766 },
767 {
768 .procname = "watchdog_thresh",
769 .data = &watchdog_thresh,
770 .maxlen = sizeof(int),
771 .mode = 0644,
772 .proc_handler = proc_watchdog_thresh,
773 .extra1 = SYSCTL_ZERO,
774 .extra2 = (void *)&sixty,
775 },
776 {
777 .procname = "nmi_watchdog",
778 .data = &nmi_watchdog_user_enabled,
779 .maxlen = sizeof(int),
780 .mode = NMI_WATCHDOG_SYSCTL_PERM,
781 .proc_handler = proc_nmi_watchdog,
782 .extra1 = SYSCTL_ZERO,
783 .extra2 = SYSCTL_ONE,
784 },
785 {
786 .procname = "watchdog_cpumask",
787 .data = &watchdog_cpumask_bits,
788 .maxlen = NR_CPUS,
789 .mode = 0644,
790 .proc_handler = proc_watchdog_cpumask,
791 },
792#ifdef CONFIG_SOFTLOCKUP_DETECTOR
793 {
794 .procname = "soft_watchdog",
795 .data = &soft_watchdog_user_enabled,
796 .maxlen = sizeof(int),
797 .mode = 0644,
798 .proc_handler = proc_soft_watchdog,
799 .extra1 = SYSCTL_ZERO,
800 .extra2 = SYSCTL_ONE,
801 },
802 {
803 .procname = "softlockup_panic",
804 .data = &softlockup_panic,
805 .maxlen = sizeof(int),
806 .mode = 0644,
807 .proc_handler = proc_dointvec_minmax,
808 .extra1 = SYSCTL_ZERO,
809 .extra2 = SYSCTL_ONE,
810 },
811#ifdef CONFIG_SMP
812 {
813 .procname = "softlockup_all_cpu_backtrace",
814 .data = &sysctl_softlockup_all_cpu_backtrace,
815 .maxlen = sizeof(int),
816 .mode = 0644,
817 .proc_handler = proc_dointvec_minmax,
818 .extra1 = SYSCTL_ZERO,
819 .extra2 = SYSCTL_ONE,
820 },
821#endif /* CONFIG_SMP */
822#endif
823#ifdef CONFIG_HARDLOCKUP_DETECTOR
824 {
825 .procname = "hardlockup_panic",
826 .data = &hardlockup_panic,
827 .maxlen = sizeof(int),
828 .mode = 0644,
829 .proc_handler = proc_dointvec_minmax,
830 .extra1 = SYSCTL_ZERO,
831 .extra2 = SYSCTL_ONE,
832 },
833#ifdef CONFIG_SMP
834 {
835 .procname = "hardlockup_all_cpu_backtrace",
836 .data = &sysctl_hardlockup_all_cpu_backtrace,
837 .maxlen = sizeof(int),
838 .mode = 0644,
839 .proc_handler = proc_dointvec_minmax,
840 .extra1 = SYSCTL_ZERO,
841 .extra2 = SYSCTL_ONE,
842 },
843#endif /* CONFIG_SMP */
844#endif
845 {}
846};
847
848static void __init watchdog_sysctl_init(void)
849{
850 register_sysctl_init("kernel", watchdog_sysctls);
851}
852#else
853#define watchdog_sysctl_init() do { } while (0)
854#endif /* CONFIG_SYSCTL */
855
856void __init lockup_detector_init(void)
857{
858 if (tick_nohz_full_enabled())
859 pr_info("Disabling watchdog on nohz_full cores by default\n");
860
861 cpumask_copy(&watchdog_cpumask,
862 housekeeping_cpumask(HK_TYPE_TIMER));
863
864 if (!watchdog_nmi_probe())
865 nmi_watchdog_available = true;
866 lockup_detector_setup();
867 watchdog_sysctl_init();
868}
1/*
2 * Detect hard and soft lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
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.
10 */
11
12#define pr_fmt(fmt) "NMI watchdog: " fmt
13
14#include <linux/mm.h>
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
26#include <asm/irq_regs.h>
27#include <linux/kvm_para.h>
28#include <linux/perf_event.h>
29
30int watchdog_enabled = 1;
31int __read_mostly watchdog_thresh = 10;
32
33static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
34static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
35static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
36static DEFINE_PER_CPU(bool, softlockup_touch_sync);
37static DEFINE_PER_CPU(bool, soft_watchdog_warn);
38#ifdef CONFIG_HARDLOCKUP_DETECTOR
39static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
41static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
42static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
43static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
44#endif
45
46/* boot commands */
47/*
48 * Should we panic when a soft-lockup or hard-lockup occurs:
49 */
50#ifdef CONFIG_HARDLOCKUP_DETECTOR
51static int hardlockup_panic =
52 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
53
54static int __init hardlockup_panic_setup(char *str)
55{
56 if (!strncmp(str, "panic", 5))
57 hardlockup_panic = 1;
58 else if (!strncmp(str, "nopanic", 7))
59 hardlockup_panic = 0;
60 else if (!strncmp(str, "0", 1))
61 watchdog_enabled = 0;
62 return 1;
63}
64__setup("nmi_watchdog=", hardlockup_panic_setup);
65#endif
66
67unsigned int __read_mostly softlockup_panic =
68 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
69
70static int __init softlockup_panic_setup(char *str)
71{
72 softlockup_panic = simple_strtoul(str, NULL, 0);
73
74 return 1;
75}
76__setup("softlockup_panic=", softlockup_panic_setup);
77
78static int __init nowatchdog_setup(char *str)
79{
80 watchdog_enabled = 0;
81 return 1;
82}
83__setup("nowatchdog", nowatchdog_setup);
84
85/* deprecated */
86static int __init nosoftlockup_setup(char *str)
87{
88 watchdog_enabled = 0;
89 return 1;
90}
91__setup("nosoftlockup", nosoftlockup_setup);
92/* */
93
94/*
95 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
96 * lockups can have false positives under extreme conditions. So we generally
97 * want a higher threshold for soft lockups than for hard lockups. So we couple
98 * the thresholds with a factor: we make the soft threshold twice the amount of
99 * time the hard threshold is.
100 */
101static int get_softlockup_thresh(void)
102{
103 return watchdog_thresh * 2;
104}
105
106/*
107 * Returns seconds, approximately. We don't need nanosecond
108 * resolution, and we don't need to waste time with a big divide when
109 * 2^30ns == 1.074s.
110 */
111static unsigned long get_timestamp(int this_cpu)
112{
113 return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */
114}
115
116static unsigned long get_sample_period(void)
117{
118 /*
119 * convert watchdog_thresh from seconds to ns
120 * the divide by 5 is to give hrtimer several chances (two
121 * or three with the current relation between the soft
122 * and hard thresholds) to increment before the
123 * hardlockup detector generates a warning
124 */
125 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
126}
127
128/* Commands for resetting the watchdog */
129static void __touch_watchdog(void)
130{
131 int this_cpu = smp_processor_id();
132
133 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
134}
135
136void touch_softlockup_watchdog(void)
137{
138 __this_cpu_write(watchdog_touch_ts, 0);
139}
140EXPORT_SYMBOL(touch_softlockup_watchdog);
141
142void touch_all_softlockup_watchdogs(void)
143{
144 int cpu;
145
146 /*
147 * this is done lockless
148 * do we care if a 0 races with a timestamp?
149 * all it means is the softlock check starts one cycle later
150 */
151 for_each_online_cpu(cpu)
152 per_cpu(watchdog_touch_ts, cpu) = 0;
153}
154
155#ifdef CONFIG_HARDLOCKUP_DETECTOR
156void touch_nmi_watchdog(void)
157{
158 if (watchdog_enabled) {
159 unsigned cpu;
160
161 for_each_present_cpu(cpu) {
162 if (per_cpu(watchdog_nmi_touch, cpu) != true)
163 per_cpu(watchdog_nmi_touch, cpu) = true;
164 }
165 }
166 touch_softlockup_watchdog();
167}
168EXPORT_SYMBOL(touch_nmi_watchdog);
169
170#endif
171
172void touch_softlockup_watchdog_sync(void)
173{
174 __raw_get_cpu_var(softlockup_touch_sync) = true;
175 __raw_get_cpu_var(watchdog_touch_ts) = 0;
176}
177
178#ifdef CONFIG_HARDLOCKUP_DETECTOR
179/* watchdog detector functions */
180static int is_hardlockup(void)
181{
182 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
183
184 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
185 return 1;
186
187 __this_cpu_write(hrtimer_interrupts_saved, hrint);
188 return 0;
189}
190#endif
191
192static int is_softlockup(unsigned long touch_ts)
193{
194 unsigned long now = get_timestamp(smp_processor_id());
195
196 /* Warn about unreasonable delays: */
197 if (time_after(now, touch_ts + get_softlockup_thresh()))
198 return now - touch_ts;
199
200 return 0;
201}
202
203#ifdef CONFIG_HARDLOCKUP_DETECTOR
204
205static struct perf_event_attr wd_hw_attr = {
206 .type = PERF_TYPE_HARDWARE,
207 .config = PERF_COUNT_HW_CPU_CYCLES,
208 .size = sizeof(struct perf_event_attr),
209 .pinned = 1,
210 .disabled = 1,
211};
212
213/* Callback function for perf event subsystem */
214static void watchdog_overflow_callback(struct perf_event *event,
215 struct perf_sample_data *data,
216 struct pt_regs *regs)
217{
218 /* Ensure the watchdog never gets throttled */
219 event->hw.interrupts = 0;
220
221 if (__this_cpu_read(watchdog_nmi_touch) == true) {
222 __this_cpu_write(watchdog_nmi_touch, false);
223 return;
224 }
225
226 /* check for a hardlockup
227 * This is done by making sure our timer interrupt
228 * is incrementing. The timer interrupt should have
229 * fired multiple times before we overflow'd. If it hasn't
230 * then this is a good indication the cpu is stuck
231 */
232 if (is_hardlockup()) {
233 int this_cpu = smp_processor_id();
234
235 /* only print hardlockups once */
236 if (__this_cpu_read(hard_watchdog_warn) == true)
237 return;
238
239 if (hardlockup_panic)
240 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
241 else
242 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
243
244 __this_cpu_write(hard_watchdog_warn, true);
245 return;
246 }
247
248 __this_cpu_write(hard_watchdog_warn, false);
249 return;
250}
251static void watchdog_interrupt_count(void)
252{
253 __this_cpu_inc(hrtimer_interrupts);
254}
255#else
256static inline void watchdog_interrupt_count(void) { return; }
257#endif /* CONFIG_HARDLOCKUP_DETECTOR */
258
259/* watchdog kicker functions */
260static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
261{
262 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
263 struct pt_regs *regs = get_irq_regs();
264 int duration;
265
266 /* kick the hardlockup detector */
267 watchdog_interrupt_count();
268
269 /* kick the softlockup detector */
270 wake_up_process(__this_cpu_read(softlockup_watchdog));
271
272 /* .. and repeat */
273 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
274
275 if (touch_ts == 0) {
276 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
277 /*
278 * If the time stamp was touched atomically
279 * make sure the scheduler tick is up to date.
280 */
281 __this_cpu_write(softlockup_touch_sync, false);
282 sched_clock_tick();
283 }
284
285 /* Clear the guest paused flag on watchdog reset */
286 kvm_check_and_clear_guest_paused();
287 __touch_watchdog();
288 return HRTIMER_RESTART;
289 }
290
291 /* check for a softlockup
292 * This is done by making sure a high priority task is
293 * being scheduled. The task touches the watchdog to
294 * indicate it is getting cpu time. If it hasn't then
295 * this is a good indication some task is hogging the cpu
296 */
297 duration = is_softlockup(touch_ts);
298 if (unlikely(duration)) {
299 /*
300 * If a virtual machine is stopped by the host it can look to
301 * the watchdog like a soft lockup, check to see if the host
302 * stopped the vm before we issue the warning
303 */
304 if (kvm_check_and_clear_guest_paused())
305 return HRTIMER_RESTART;
306
307 /* only warn once */
308 if (__this_cpu_read(soft_watchdog_warn) == true)
309 return HRTIMER_RESTART;
310
311 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
312 smp_processor_id(), duration,
313 current->comm, task_pid_nr(current));
314 print_modules();
315 print_irqtrace_events(current);
316 if (regs)
317 show_regs(regs);
318 else
319 dump_stack();
320
321 if (softlockup_panic)
322 panic("softlockup: hung tasks");
323 __this_cpu_write(soft_watchdog_warn, true);
324 } else
325 __this_cpu_write(soft_watchdog_warn, false);
326
327 return HRTIMER_RESTART;
328}
329
330
331/*
332 * The watchdog thread - touches the timestamp.
333 */
334static int watchdog(void *unused)
335{
336 struct sched_param param = { .sched_priority = 0 };
337 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
338
339 /* initialize timestamp */
340 __touch_watchdog();
341
342 /* kick off the timer for the hardlockup detector */
343 /* done here because hrtimer_start can only pin to smp_processor_id() */
344 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
345 HRTIMER_MODE_REL_PINNED);
346
347 set_current_state(TASK_INTERRUPTIBLE);
348 /*
349 * Run briefly (kicked by the hrtimer callback function) once every
350 * get_sample_period() seconds (4 seconds by default) to reset the
351 * softlockup timestamp. If this gets delayed for more than
352 * 2*watchdog_thresh seconds then the debug-printout triggers in
353 * watchdog_timer_fn().
354 */
355 while (!kthread_should_stop()) {
356 __touch_watchdog();
357 schedule();
358
359 if (kthread_should_stop())
360 break;
361
362 set_current_state(TASK_INTERRUPTIBLE);
363 }
364 /*
365 * Drop the policy/priority elevation during thread exit to avoid a
366 * scheduling latency spike.
367 */
368 __set_current_state(TASK_RUNNING);
369 sched_setscheduler(current, SCHED_NORMAL, ¶m);
370 return 0;
371}
372
373
374#ifdef CONFIG_HARDLOCKUP_DETECTOR
375/*
376 * People like the simple clean cpu node info on boot.
377 * Reduce the watchdog noise by only printing messages
378 * that are different from what cpu0 displayed.
379 */
380static unsigned long cpu0_err;
381
382static int watchdog_nmi_enable(int cpu)
383{
384 struct perf_event_attr *wd_attr;
385 struct perf_event *event = per_cpu(watchdog_ev, cpu);
386
387 /* is it already setup and enabled? */
388 if (event && event->state > PERF_EVENT_STATE_OFF)
389 goto out;
390
391 /* it is setup but not enabled */
392 if (event != NULL)
393 goto out_enable;
394
395 wd_attr = &wd_hw_attr;
396 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
397
398 /* Try to register using hardware perf events */
399 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
400
401 /* save cpu0 error for future comparision */
402 if (cpu == 0 && IS_ERR(event))
403 cpu0_err = PTR_ERR(event);
404
405 if (!IS_ERR(event)) {
406 /* only print for cpu0 or different than cpu0 */
407 if (cpu == 0 || cpu0_err)
408 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
409 goto out_save;
410 }
411
412 /* skip displaying the same error again */
413 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
414 return PTR_ERR(event);
415
416 /* vary the KERN level based on the returned errno */
417 if (PTR_ERR(event) == -EOPNOTSUPP)
418 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
419 else if (PTR_ERR(event) == -ENOENT)
420 pr_warning("disabled (cpu%i): hardware events not enabled\n",
421 cpu);
422 else
423 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
424 cpu, PTR_ERR(event));
425 return PTR_ERR(event);
426
427 /* success path */
428out_save:
429 per_cpu(watchdog_ev, cpu) = event;
430out_enable:
431 perf_event_enable(per_cpu(watchdog_ev, cpu));
432out:
433 return 0;
434}
435
436static void watchdog_nmi_disable(int cpu)
437{
438 struct perf_event *event = per_cpu(watchdog_ev, cpu);
439
440 if (event) {
441 perf_event_disable(event);
442 per_cpu(watchdog_ev, cpu) = NULL;
443
444 /* should be in cleanup, but blocks oprofile */
445 perf_event_release_kernel(event);
446 }
447 return;
448}
449#else
450static int watchdog_nmi_enable(int cpu) { return 0; }
451static void watchdog_nmi_disable(int cpu) { return; }
452#endif /* CONFIG_HARDLOCKUP_DETECTOR */
453
454/* prepare/enable/disable routines */
455static void watchdog_prepare_cpu(int cpu)
456{
457 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
458
459 WARN_ON(per_cpu(softlockup_watchdog, cpu));
460 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
461 hrtimer->function = watchdog_timer_fn;
462}
463
464static int watchdog_enable(int cpu)
465{
466 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
467 int err = 0;
468
469 /* enable the perf event */
470 err = watchdog_nmi_enable(cpu);
471
472 /* Regardless of err above, fall through and start softlockup */
473
474 /* create the watchdog thread */
475 if (!p) {
476 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
477 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
478 if (IS_ERR(p)) {
479 pr_err("softlockup watchdog for %i failed\n", cpu);
480 if (!err) {
481 /* if hardlockup hasn't already set this */
482 err = PTR_ERR(p);
483 /* and disable the perf event */
484 watchdog_nmi_disable(cpu);
485 }
486 goto out;
487 }
488 sched_setscheduler(p, SCHED_FIFO, ¶m);
489 kthread_bind(p, cpu);
490 per_cpu(watchdog_touch_ts, cpu) = 0;
491 per_cpu(softlockup_watchdog, cpu) = p;
492 wake_up_process(p);
493 }
494
495out:
496 return err;
497}
498
499static void watchdog_disable(int cpu)
500{
501 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
502 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
503
504 /*
505 * cancel the timer first to stop incrementing the stats
506 * and waking up the kthread
507 */
508 hrtimer_cancel(hrtimer);
509
510 /* disable the perf event */
511 watchdog_nmi_disable(cpu);
512
513 /* stop the watchdog thread */
514 if (p) {
515 per_cpu(softlockup_watchdog, cpu) = NULL;
516 kthread_stop(p);
517 }
518}
519
520/* sysctl functions */
521#ifdef CONFIG_SYSCTL
522static void watchdog_enable_all_cpus(void)
523{
524 int cpu;
525
526 watchdog_enabled = 0;
527
528 for_each_online_cpu(cpu)
529 if (!watchdog_enable(cpu))
530 /* if any cpu succeeds, watchdog is considered
531 enabled for the system */
532 watchdog_enabled = 1;
533
534 if (!watchdog_enabled)
535 pr_err("failed to be enabled on some cpus\n");
536
537}
538
539static void watchdog_disable_all_cpus(void)
540{
541 int cpu;
542
543 for_each_online_cpu(cpu)
544 watchdog_disable(cpu);
545
546 /* if all watchdogs are disabled, then they are disabled for the system */
547 watchdog_enabled = 0;
548}
549
550
551/*
552 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
553 */
554
555int proc_dowatchdog(struct ctl_table *table, int write,
556 void __user *buffer, size_t *lenp, loff_t *ppos)
557{
558 int ret;
559
560 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
561 if (ret || !write)
562 goto out;
563
564 if (watchdog_enabled && watchdog_thresh)
565 watchdog_enable_all_cpus();
566 else
567 watchdog_disable_all_cpus();
568
569out:
570 return ret;
571}
572#endif /* CONFIG_SYSCTL */
573
574
575/*
576 * Create/destroy watchdog threads as CPUs come and go:
577 */
578static int __cpuinit
579cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
580{
581 int hotcpu = (unsigned long)hcpu;
582
583 switch (action) {
584 case CPU_UP_PREPARE:
585 case CPU_UP_PREPARE_FROZEN:
586 watchdog_prepare_cpu(hotcpu);
587 break;
588 case CPU_ONLINE:
589 case CPU_ONLINE_FROZEN:
590 if (watchdog_enabled)
591 watchdog_enable(hotcpu);
592 break;
593#ifdef CONFIG_HOTPLUG_CPU
594 case CPU_UP_CANCELED:
595 case CPU_UP_CANCELED_FROZEN:
596 watchdog_disable(hotcpu);
597 break;
598 case CPU_DEAD:
599 case CPU_DEAD_FROZEN:
600 watchdog_disable(hotcpu);
601 break;
602#endif /* CONFIG_HOTPLUG_CPU */
603 }
604
605 /*
606 * hardlockup and softlockup are not important enough
607 * to block cpu bring up. Just always succeed and
608 * rely on printk output to flag problems.
609 */
610 return NOTIFY_OK;
611}
612
613static struct notifier_block __cpuinitdata cpu_nfb = {
614 .notifier_call = cpu_callback
615};
616
617void __init lockup_detector_init(void)
618{
619 void *cpu = (void *)(long)smp_processor_id();
620 int err;
621
622 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
623 WARN_ON(notifier_to_errno(err));
624
625 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
626 register_cpu_notifier(&cpu_nfb);
627
628 return;
629}