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) "watchdog: " fmt
15 #include <linux/cpu.h>
16 #include <linux/nmi.h>
17 #include <linux/init.h>
18 #include <linux/module.h>
19 #include <linux/sysctl.h>
20 #include <linux/smpboot.h>
21 #include <linux/sched/rt.h>
22 #include <uapi/linux/sched/types.h>
23 #include <linux/tick.h>
24 #include <linux/workqueue.h>
25 #include <linux/sched/clock.h>
26 #include <linux/sched/debug.h>
28 #include <asm/irq_regs.h>
29 #include <linux/kvm_para.h>
30 #include <linux/kthread.h>
32 static DEFINE_MUTEX(watchdog_mutex);
34 int __read_mostly nmi_watchdog_enabled;
36 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
37 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED |
40 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
43 #ifdef CONFIG_HARDLOCKUP_DETECTOR
45 * Should we panic when a soft-lockup or hard-lockup occurs:
47 unsigned int __read_mostly hardlockup_panic =
48 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
50 * We may not want to enable hard lockup detection by default in all cases,
51 * for example when running the kernel as a guest on a hypervisor. In these
52 * cases this function can be called to disable hard lockup detection. This
53 * function should only be executed once by the boot processor before the
54 * kernel command line parameters are parsed, because otherwise it is not
55 * possible to override this in hardlockup_panic_setup().
57 void __init hardlockup_detector_disable(void)
59 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
62 static int __init hardlockup_panic_setup(char *str)
64 if (!strncmp(str, "panic", 5))
66 else if (!strncmp(str, "nopanic", 7))
68 else if (!strncmp(str, "0", 1))
69 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
70 else if (!strncmp(str, "1", 1))
71 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
74 __setup("nmi_watchdog=", hardlockup_panic_setup);
77 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
79 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
81 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
84 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
85 # endif /* CONFIG_SMP */
86 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
88 int __read_mostly watchdog_user_enabled;
89 int __read_mostly watchdog_thresh = 10;
91 struct cpumask watchdog_cpumask __read_mostly;
92 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
95 * These functions can be overridden if an architecture implements its
96 * own hardlockup detector.
98 * watchdog_nmi_enable/disable can be implemented to start and stop when
99 * softlockup watchdog threads start and stop. The arch must select the
100 * SOFTLOCKUP_DETECTOR Kconfig.
102 int __weak watchdog_nmi_enable(unsigned int cpu)
107 void __weak watchdog_nmi_disable(unsigned int cpu)
109 hardlockup_detector_perf_disable();
113 * watchdog_nmi_reconfigure can be implemented to be notified after any
114 * watchdog configuration change. The arch hardlockup watchdog should
115 * respond to the following variables:
116 * - nmi_watchdog_enabled
119 * - sysctl_hardlockup_all_cpu_backtrace
122 void __weak watchdog_nmi_reconfigure(void) { }
124 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
126 /* Global variables, exported for sysctl */
127 unsigned int __read_mostly softlockup_panic =
128 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
129 int __read_mostly soft_watchdog_enabled;
131 struct cpumask watchdog_allowed_mask __read_mostly;
132 static bool softlockup_threads_initialized __read_mostly;
134 static u64 __read_mostly sample_period;
136 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
137 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
138 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
139 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
140 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
141 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
142 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
143 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
144 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
145 static unsigned long soft_lockup_nmi_warn;
147 static int __init softlockup_panic_setup(char *str)
149 softlockup_panic = simple_strtoul(str, NULL, 0);
152 __setup("softlockup_panic=", softlockup_panic_setup);
154 static int __init nowatchdog_setup(char *str)
156 watchdog_enabled = 0;
159 __setup("nowatchdog", nowatchdog_setup);
161 static int __init nosoftlockup_setup(char *str)
163 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
166 __setup("nosoftlockup", nosoftlockup_setup);
169 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
171 static int __init softlockup_all_cpu_backtrace_setup(char *str)
173 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
176 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
179 static void __lockup_detector_cleanup(void);
182 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
183 * lockups can have false positives under extreme conditions. So we generally
184 * want a higher threshold for soft lockups than for hard lockups. So we couple
185 * the thresholds with a factor: we make the soft threshold twice the amount of
186 * time the hard threshold is.
188 static int get_softlockup_thresh(void)
190 return watchdog_thresh * 2;
194 * Returns seconds, approximately. We don't need nanosecond
195 * resolution, and we don't need to waste time with a big divide when
198 static unsigned long get_timestamp(void)
200 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
203 static void set_sample_period(void)
206 * convert watchdog_thresh from seconds to ns
207 * the divide by 5 is to give hrtimer several chances (two
208 * or three with the current relation between the soft
209 * and hard thresholds) to increment before the
210 * hardlockup detector generates a warning
212 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
213 watchdog_update_hrtimer_threshold(sample_period);
216 /* Commands for resetting the watchdog */
217 static void __touch_watchdog(void)
219 __this_cpu_write(watchdog_touch_ts, get_timestamp());
223 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
225 * Call when the scheduler may have stalled for legitimate reasons
226 * preventing the watchdog task from executing - e.g. the scheduler
227 * entering idle state. This should only be used for scheduler events.
228 * Use touch_softlockup_watchdog() for everything else.
230 void touch_softlockup_watchdog_sched(void)
233 * Preemption can be enabled. It doesn't matter which CPU's timestamp
234 * gets zeroed here, so use the raw_ operation.
236 raw_cpu_write(watchdog_touch_ts, 0);
239 void touch_softlockup_watchdog(void)
241 touch_softlockup_watchdog_sched();
242 wq_watchdog_touch(raw_smp_processor_id());
244 EXPORT_SYMBOL(touch_softlockup_watchdog);
246 void touch_all_softlockup_watchdogs(void)
251 * watchdog_mutex cannpt be taken here, as this might be called
252 * from (soft)interrupt context, so the access to
253 * watchdog_allowed_cpumask might race with a concurrent update.
255 * The watchdog time stamp can race against a concurrent real
256 * update as well, the only side effect might be a cycle delay for
257 * the softlockup check.
259 for_each_cpu(cpu, &watchdog_allowed_mask)
260 per_cpu(watchdog_touch_ts, cpu) = 0;
261 wq_watchdog_touch(-1);
264 void touch_softlockup_watchdog_sync(void)
266 __this_cpu_write(softlockup_touch_sync, true);
267 __this_cpu_write(watchdog_touch_ts, 0);
270 static int is_softlockup(unsigned long touch_ts)
272 unsigned long now = get_timestamp();
274 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
275 /* Warn about unreasonable delays. */
276 if (time_after(now, touch_ts + get_softlockup_thresh()))
277 return now - touch_ts;
282 /* watchdog detector functions */
283 bool is_hardlockup(void)
285 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
287 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
290 __this_cpu_write(hrtimer_interrupts_saved, hrint);
294 static void watchdog_interrupt_count(void)
296 __this_cpu_inc(hrtimer_interrupts);
299 /* watchdog kicker functions */
300 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
302 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
303 struct pt_regs *regs = get_irq_regs();
305 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
307 if (!watchdog_enabled)
308 return HRTIMER_NORESTART;
310 /* kick the hardlockup detector */
311 watchdog_interrupt_count();
313 /* kick the softlockup detector */
314 wake_up_process(__this_cpu_read(softlockup_watchdog));
317 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
320 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
322 * If the time stamp was touched atomically
323 * make sure the scheduler tick is up to date.
325 __this_cpu_write(softlockup_touch_sync, false);
329 /* Clear the guest paused flag on watchdog reset */
330 kvm_check_and_clear_guest_paused();
332 return HRTIMER_RESTART;
335 /* check for a softlockup
336 * This is done by making sure a high priority task is
337 * being scheduled. The task touches the watchdog to
338 * indicate it is getting cpu time. If it hasn't then
339 * this is a good indication some task is hogging the cpu
341 duration = is_softlockup(touch_ts);
342 if (unlikely(duration)) {
344 * If a virtual machine is stopped by the host it can look to
345 * the watchdog like a soft lockup, check to see if the host
346 * stopped the vm before we issue the warning
348 if (kvm_check_and_clear_guest_paused())
349 return HRTIMER_RESTART;
352 if (__this_cpu_read(soft_watchdog_warn) == true) {
354 * When multiple processes are causing softlockups the
355 * softlockup detector only warns on the first one
356 * because the code relies on a full quiet cycle to
357 * re-arm. The second process prevents the quiet cycle
358 * and never gets reported. Use task pointers to detect
361 if (__this_cpu_read(softlockup_task_ptr_saved) !=
363 __this_cpu_write(soft_watchdog_warn, false);
366 return HRTIMER_RESTART;
369 if (softlockup_all_cpu_backtrace) {
370 /* Prevent multiple soft-lockup reports if one cpu is already
371 * engaged in dumping cpu back traces
373 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
374 /* Someone else will report us. Let's give up */
375 __this_cpu_write(soft_watchdog_warn, true);
376 return HRTIMER_RESTART;
380 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
381 smp_processor_id(), duration,
382 current->comm, task_pid_nr(current));
383 __this_cpu_write(softlockup_task_ptr_saved, current);
385 print_irqtrace_events(current);
391 if (softlockup_all_cpu_backtrace) {
392 /* Avoid generating two back traces for current
393 * given that one is already made above
395 trigger_allbutself_cpu_backtrace();
397 clear_bit(0, &soft_lockup_nmi_warn);
398 /* Barrier to sync with other cpus */
399 smp_mb__after_atomic();
402 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
403 if (softlockup_panic)
404 panic("softlockup: hung tasks");
405 __this_cpu_write(soft_watchdog_warn, true);
407 __this_cpu_write(soft_watchdog_warn, false);
409 return HRTIMER_RESTART;
412 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
414 struct sched_param param = { .sched_priority = prio };
416 sched_setscheduler(current, policy, ¶m);
419 static void watchdog_enable(unsigned int cpu)
421 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
424 * Start the timer first to prevent the NMI watchdog triggering
425 * before the timer has a chance to fire.
427 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
428 hrtimer->function = watchdog_timer_fn;
429 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
430 HRTIMER_MODE_REL_PINNED);
432 /* Initialize timestamp */
434 /* Enable the perf event */
435 watchdog_nmi_enable(cpu);
437 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
440 static void watchdog_disable(unsigned int cpu)
442 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
444 watchdog_set_prio(SCHED_NORMAL, 0);
446 * Disable the perf event first. That prevents that a large delay
447 * between disabling the timer and disabling the perf event causes
448 * the perf NMI to detect a false positive.
450 watchdog_nmi_disable(cpu);
451 hrtimer_cancel(hrtimer);
454 static void watchdog_cleanup(unsigned int cpu, bool online)
456 watchdog_disable(cpu);
459 static int watchdog_should_run(unsigned int cpu)
461 return __this_cpu_read(hrtimer_interrupts) !=
462 __this_cpu_read(soft_lockup_hrtimer_cnt);
466 * The watchdog thread function - touches the timestamp.
468 * It only runs once every sample_period seconds (4 seconds by
469 * default) to reset the softlockup timestamp. If this gets delayed
470 * for more than 2*watchdog_thresh seconds then the debug-printout
471 * triggers in watchdog_timer_fn().
473 static void watchdog(unsigned int cpu)
475 __this_cpu_write(soft_lockup_hrtimer_cnt,
476 __this_cpu_read(hrtimer_interrupts));
480 static struct smp_hotplug_thread watchdog_threads = {
481 .store = &softlockup_watchdog,
482 .thread_should_run = watchdog_should_run,
483 .thread_fn = watchdog,
484 .thread_comm = "watchdog/%u",
485 .setup = watchdog_enable,
486 .cleanup = watchdog_cleanup,
487 .park = watchdog_disable,
488 .unpark = watchdog_enable,
491 static void softlockup_update_smpboot_threads(void)
493 lockdep_assert_held(&watchdog_mutex);
495 if (!softlockup_threads_initialized)
498 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
499 &watchdog_allowed_mask);
500 __lockup_detector_cleanup();
503 /* Temporarily park all watchdog threads */
504 static void softlockup_park_all_threads(void)
506 cpumask_clear(&watchdog_allowed_mask);
507 softlockup_update_smpboot_threads();
511 * Park threads which are not longer enabled and unpark threads which have
512 * been newly enabled.
514 static void softlockup_update_threads(void)
516 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
517 softlockup_update_smpboot_threads();
520 static void softlockup_reconfigure_threads(bool enabled)
522 softlockup_park_all_threads();
525 softlockup_update_threads();
529 * Create the watchdog thread infrastructure.
531 * The threads are not unparked as watchdog_allowed_mask is empty. When
532 * the threads are sucessfully initialized, take the proper locks and
533 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
535 static __init void softlockup_init_threads(void)
540 * If sysctl is off and watchdog got disabled on the command line,
541 * nothing to do here.
543 if (!IS_ENABLED(CONFIG_SYSCTL) &&
544 !(watchdog_enabled && watchdog_thresh))
547 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
548 &watchdog_allowed_mask);
550 pr_err("Failed to initialize soft lockup detector threads\n");
554 mutex_lock(&watchdog_mutex);
555 softlockup_threads_initialized = true;
556 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
557 mutex_unlock(&watchdog_mutex);
560 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
561 static inline int watchdog_park_threads(void) { return 0; }
562 static inline void watchdog_unpark_threads(void) { }
563 static inline int watchdog_enable_all_cpus(void) { return 0; }
564 static inline void watchdog_disable_all_cpus(void) { }
565 static inline void softlockup_init_threads(void) { }
566 static inline void softlockup_update_threads(void) { }
567 static inline void softlockup_reconfigure_threads(bool enabled) { }
568 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
570 static void __lockup_detector_cleanup(void)
572 lockdep_assert_held(&watchdog_mutex);
573 hardlockup_detector_perf_cleanup();
577 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
579 * Caller must not hold the cpu hotplug rwsem.
581 void lockup_detector_cleanup(void)
583 mutex_lock(&watchdog_mutex);
584 __lockup_detector_cleanup();
585 mutex_unlock(&watchdog_mutex);
589 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
591 * Special interface for parisc. It prevents lockup detector warnings from
592 * the default pm_poweroff() function which busy loops forever.
594 void lockup_detector_soft_poweroff(void)
596 watchdog_enabled = 0;
602 * Update the run state of the lockup detectors.
604 static void proc_watchdog_update(void)
606 softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh);
607 watchdog_nmi_reconfigure();
611 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
613 * caller | table->data points to | 'which' contains the flag(s)
614 * -------------------|-----------------------|-----------------------------
615 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
616 * | | with SOFT_WATCHDOG_ENABLED
617 * -------------------|-----------------------|-----------------------------
618 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
619 * -------------------|-----------------------|-----------------------------
620 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
622 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
623 void __user *buffer, size_t *lenp, loff_t *ppos)
626 int *watchdog_param = (int *)table->data;
628 cpu_hotplug_disable();
629 mutex_lock(&watchdog_mutex);
632 * If the parameter is being read return the state of the corresponding
633 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
634 * run state of the lockup detectors.
637 *watchdog_param = (watchdog_enabled & which) != 0;
638 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
640 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
645 * There is a race window between fetching the current value
646 * from 'watchdog_enabled' and storing the new value. During
647 * this race window, watchdog_nmi_enable() can sneak in and
648 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
649 * The 'cmpxchg' detects this race and the loop retries.
652 old = watchdog_enabled;
654 * If the parameter value is not zero set the
655 * corresponding bit(s), else clear it(them).
661 } while (cmpxchg(&watchdog_enabled, old, new) != old);
664 proc_watchdog_update();
667 mutex_unlock(&watchdog_mutex);
668 cpu_hotplug_enable();
673 * /proc/sys/kernel/watchdog
675 int proc_watchdog(struct ctl_table *table, int write,
676 void __user *buffer, size_t *lenp, loff_t *ppos)
678 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
679 table, write, buffer, lenp, ppos);
683 * /proc/sys/kernel/nmi_watchdog
685 int proc_nmi_watchdog(struct ctl_table *table, int write,
686 void __user *buffer, size_t *lenp, loff_t *ppos)
688 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
689 table, write, buffer, lenp, ppos);
693 * /proc/sys/kernel/soft_watchdog
695 int proc_soft_watchdog(struct ctl_table *table, int write,
696 void __user *buffer, size_t *lenp, loff_t *ppos)
698 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
699 table, write, buffer, lenp, ppos);
703 * /proc/sys/kernel/watchdog_thresh
705 int proc_watchdog_thresh(struct ctl_table *table, int write,
706 void __user *buffer, size_t *lenp, loff_t *ppos)
710 cpu_hotplug_disable();
711 mutex_lock(&watchdog_mutex);
713 old = READ_ONCE(watchdog_thresh);
714 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
716 if (!err && write && old != READ_ONCE(watchdog_thresh))
717 proc_watchdog_update();
719 mutex_unlock(&watchdog_mutex);
720 cpu_hotplug_enable();
724 static void proc_watchdog_cpumask_update(void)
726 /* Remove impossible cpus to keep sysctl output clean. */
727 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
729 softlockup_update_threads();
730 watchdog_nmi_reconfigure();
734 * The cpumask is the mask of possible cpus that the watchdog can run
735 * on, not the mask of cpus it is actually running on. This allows the
736 * user to specify a mask that will include cpus that have not yet
737 * been brought online, if desired.
739 int proc_watchdog_cpumask(struct ctl_table *table, int write,
740 void __user *buffer, size_t *lenp, loff_t *ppos)
744 cpu_hotplug_disable();
745 mutex_lock(&watchdog_mutex);
747 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
749 proc_watchdog_cpumask_update();
751 mutex_unlock(&watchdog_mutex);
752 cpu_hotplug_enable();
755 #endif /* CONFIG_SYSCTL */
757 void __init lockup_detector_init(void)
759 #ifdef CONFIG_NO_HZ_FULL
760 if (tick_nohz_full_enabled()) {
761 pr_info("Disabling watchdog on nohz_full cores by default\n");
762 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
764 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
766 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
769 softlockup_init_threads();