1 // SPDX-License-Identifier: GPL-2.0
3 * Detect hard and soft lockups on a system
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
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.
13 #define pr_fmt(fmt) "watchdog: " fmt
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>
27 #include <asm/irq_regs.h>
28 #include <linux/kvm_para.h>
30 static DEFINE_MUTEX(watchdog_mutex);
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
36 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37 # define NMI_WATCHDOG_DEFAULT 0
40 unsigned long __read_mostly watchdog_enabled;
41 int __read_mostly watchdog_user_enabled = 1;
42 int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43 int __read_mostly soft_watchdog_user_enabled = 1;
44 int __read_mostly watchdog_thresh = 10;
45 static int __read_mostly nmi_watchdog_available;
47 static struct cpumask watchdog_allowed_mask __read_mostly;
49 struct cpumask watchdog_cpumask __read_mostly;
50 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
52 #ifdef CONFIG_HARDLOCKUP_DETECTOR
54 * Should we panic when a soft-lockup or hard-lockup occurs:
56 unsigned int __read_mostly hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
59 * We may not want to enable hard lockup detection by default in all cases,
60 * for example when running the kernel as a guest on a hypervisor. In these
61 * cases this function can be called to disable hard lockup detection. This
62 * function should only be executed once by the boot processor before the
63 * kernel command line parameters are parsed, because otherwise it is not
64 * possible to override this in hardlockup_panic_setup().
66 void __init hardlockup_detector_disable(void)
68 nmi_watchdog_user_enabled = 0;
71 static int __init hardlockup_panic_setup(char *str)
73 if (!strncmp(str, "panic", 5))
75 else if (!strncmp(str, "nopanic", 7))
77 else if (!strncmp(str, "0", 1))
78 nmi_watchdog_user_enabled = 0;
79 else if (!strncmp(str, "1", 1))
80 nmi_watchdog_user_enabled = 1;
83 __setup("nmi_watchdog=", hardlockup_panic_setup);
86 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
88 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
90 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
93 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
94 # endif /* CONFIG_SMP */
95 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
98 * These functions can be overridden if an architecture implements its
99 * own hardlockup detector.
101 * watchdog_nmi_enable/disable can be implemented to start and stop when
102 * softlockup watchdog threads start and stop. The arch must select the
103 * SOFTLOCKUP_DETECTOR Kconfig.
105 int __weak watchdog_nmi_enable(unsigned int cpu)
107 hardlockup_detector_perf_enable();
111 void __weak watchdog_nmi_disable(unsigned int cpu)
113 hardlockup_detector_perf_disable();
116 /* Return 0, if a NMI watchdog is available. Error code otherwise */
117 int __weak __init watchdog_nmi_probe(void)
119 return hardlockup_detector_perf_init();
123 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
125 * The reconfiguration steps are:
126 * watchdog_nmi_stop();
127 * update_variables();
128 * watchdog_nmi_start();
130 void __weak watchdog_nmi_stop(void) { }
133 * watchdog_nmi_start - Start the watchdog after reconfiguration
135 * Counterpart to watchdog_nmi_stop().
137 * The following variables have been updated in update_variables() and
138 * contain the currently valid configuration:
143 void __weak watchdog_nmi_start(void) { }
146 * lockup_detector_update_enable - Update the sysctl enable bit
148 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
149 * can't race with watchdog_nmi_disable().
151 static void lockup_detector_update_enable(void)
153 watchdog_enabled = 0;
154 if (!watchdog_user_enabled)
156 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
157 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
158 if (soft_watchdog_user_enabled)
159 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
162 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
164 /* Global variables, exported for sysctl */
165 unsigned int __read_mostly softlockup_panic =
166 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
168 static bool softlockup_initialized __read_mostly;
169 static u64 __read_mostly sample_period;
171 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
172 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
173 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
174 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
175 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
176 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
177 static unsigned long soft_lockup_nmi_warn;
179 static int __init softlockup_panic_setup(char *str)
181 softlockup_panic = simple_strtoul(str, NULL, 0);
184 __setup("softlockup_panic=", softlockup_panic_setup);
186 static int __init nowatchdog_setup(char *str)
188 watchdog_user_enabled = 0;
191 __setup("nowatchdog", nowatchdog_setup);
193 static int __init nosoftlockup_setup(char *str)
195 soft_watchdog_user_enabled = 0;
198 __setup("nosoftlockup", nosoftlockup_setup);
200 static int __init watchdog_thresh_setup(char *str)
202 get_option(&str, &watchdog_thresh);
205 __setup("watchdog_thresh=", watchdog_thresh_setup);
208 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
210 static int __init softlockup_all_cpu_backtrace_setup(char *str)
212 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
215 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
218 static void __lockup_detector_cleanup(void);
221 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
222 * lockups can have false positives under extreme conditions. So we generally
223 * want a higher threshold for soft lockups than for hard lockups. So we couple
224 * the thresholds with a factor: we make the soft threshold twice the amount of
225 * time the hard threshold is.
227 static int get_softlockup_thresh(void)
229 return watchdog_thresh * 2;
233 * Returns seconds, approximately. We don't need nanosecond
234 * resolution, and we don't need to waste time with a big divide when
237 static unsigned long get_timestamp(void)
239 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
242 static void set_sample_period(void)
245 * convert watchdog_thresh from seconds to ns
246 * the divide by 5 is to give hrtimer several chances (two
247 * or three with the current relation between the soft
248 * and hard thresholds) to increment before the
249 * hardlockup detector generates a warning
251 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
252 watchdog_update_hrtimer_threshold(sample_period);
255 /* Commands for resetting the watchdog */
256 static void __touch_watchdog(void)
258 __this_cpu_write(watchdog_touch_ts, get_timestamp());
262 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
264 * Call when the scheduler may have stalled for legitimate reasons
265 * preventing the watchdog task from executing - e.g. the scheduler
266 * entering idle state. This should only be used for scheduler events.
267 * Use touch_softlockup_watchdog() for everything else.
269 notrace void touch_softlockup_watchdog_sched(void)
272 * Preemption can be enabled. It doesn't matter which CPU's timestamp
273 * gets zeroed here, so use the raw_ operation.
275 raw_cpu_write(watchdog_touch_ts, 0);
278 notrace void touch_softlockup_watchdog(void)
280 touch_softlockup_watchdog_sched();
281 wq_watchdog_touch(raw_smp_processor_id());
283 EXPORT_SYMBOL(touch_softlockup_watchdog);
285 void touch_all_softlockup_watchdogs(void)
290 * watchdog_mutex cannpt be taken here, as this might be called
291 * from (soft)interrupt context, so the access to
292 * watchdog_allowed_cpumask might race with a concurrent update.
294 * The watchdog time stamp can race against a concurrent real
295 * update as well, the only side effect might be a cycle delay for
296 * the softlockup check.
298 for_each_cpu(cpu, &watchdog_allowed_mask)
299 per_cpu(watchdog_touch_ts, cpu) = 0;
300 wq_watchdog_touch(-1);
303 void touch_softlockup_watchdog_sync(void)
305 __this_cpu_write(softlockup_touch_sync, true);
306 __this_cpu_write(watchdog_touch_ts, 0);
309 static int is_softlockup(unsigned long touch_ts)
311 unsigned long now = get_timestamp();
313 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
314 /* Warn about unreasonable delays. */
315 if (time_after(now, touch_ts + get_softlockup_thresh()))
316 return now - touch_ts;
321 /* watchdog detector functions */
322 bool is_hardlockup(void)
324 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
326 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
329 __this_cpu_write(hrtimer_interrupts_saved, hrint);
333 static void watchdog_interrupt_count(void)
335 __this_cpu_inc(hrtimer_interrupts);
338 static DEFINE_PER_CPU(struct completion, softlockup_completion);
339 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
342 * The watchdog thread function - touches the timestamp.
344 * It only runs once every sample_period seconds (4 seconds by
345 * default) to reset the softlockup timestamp. If this gets delayed
346 * for more than 2*watchdog_thresh seconds then the debug-printout
347 * triggers in watchdog_timer_fn().
349 static int softlockup_fn(void *data)
352 complete(this_cpu_ptr(&softlockup_completion));
357 /* watchdog kicker functions */
358 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
360 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
361 struct pt_regs *regs = get_irq_regs();
363 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
365 if (!watchdog_enabled)
366 return HRTIMER_NORESTART;
368 /* kick the hardlockup detector */
369 watchdog_interrupt_count();
371 /* kick the softlockup detector */
372 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
373 reinit_completion(this_cpu_ptr(&softlockup_completion));
374 stop_one_cpu_nowait(smp_processor_id(),
376 this_cpu_ptr(&softlockup_stop_work));
380 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
383 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
385 * If the time stamp was touched atomically
386 * make sure the scheduler tick is up to date.
388 __this_cpu_write(softlockup_touch_sync, false);
392 /* Clear the guest paused flag on watchdog reset */
393 kvm_check_and_clear_guest_paused();
395 return HRTIMER_RESTART;
398 /* check for a softlockup
399 * This is done by making sure a high priority task is
400 * being scheduled. The task touches the watchdog to
401 * indicate it is getting cpu time. If it hasn't then
402 * this is a good indication some task is hogging the cpu
404 duration = is_softlockup(touch_ts);
405 if (unlikely(duration)) {
407 * If a virtual machine is stopped by the host it can look to
408 * the watchdog like a soft lockup, check to see if the host
409 * stopped the vm before we issue the warning
411 if (kvm_check_and_clear_guest_paused())
412 return HRTIMER_RESTART;
415 if (__this_cpu_read(soft_watchdog_warn) == true)
416 return HRTIMER_RESTART;
418 if (softlockup_all_cpu_backtrace) {
419 /* Prevent multiple soft-lockup reports if one cpu is already
420 * engaged in dumping cpu back traces
422 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
423 /* Someone else will report us. Let's give up */
424 __this_cpu_write(soft_watchdog_warn, true);
425 return HRTIMER_RESTART;
429 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
430 smp_processor_id(), duration,
431 current->comm, task_pid_nr(current));
433 print_irqtrace_events(current);
439 if (softlockup_all_cpu_backtrace) {
440 /* Avoid generating two back traces for current
441 * given that one is already made above
443 trigger_allbutself_cpu_backtrace();
445 clear_bit(0, &soft_lockup_nmi_warn);
446 /* Barrier to sync with other cpus */
447 smp_mb__after_atomic();
450 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
451 if (softlockup_panic)
452 panic("softlockup: hung tasks");
453 __this_cpu_write(soft_watchdog_warn, true);
455 __this_cpu_write(soft_watchdog_warn, false);
457 return HRTIMER_RESTART;
460 static void watchdog_enable(unsigned int cpu)
462 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
463 struct completion *done = this_cpu_ptr(&softlockup_completion);
465 WARN_ON_ONCE(cpu != smp_processor_id());
467 init_completion(done);
471 * Start the timer first to prevent the NMI watchdog triggering
472 * before the timer has a chance to fire.
474 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
475 hrtimer->function = watchdog_timer_fn;
476 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
477 HRTIMER_MODE_REL_PINNED_HARD);
479 /* Initialize timestamp */
481 /* Enable the perf event */
482 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
483 watchdog_nmi_enable(cpu);
486 static void watchdog_disable(unsigned int cpu)
488 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
490 WARN_ON_ONCE(cpu != smp_processor_id());
493 * Disable the perf event first. That prevents that a large delay
494 * between disabling the timer and disabling the perf event causes
495 * the perf NMI to detect a false positive.
497 watchdog_nmi_disable(cpu);
498 hrtimer_cancel(hrtimer);
499 wait_for_completion(this_cpu_ptr(&softlockup_completion));
502 static int softlockup_stop_fn(void *data)
504 watchdog_disable(smp_processor_id());
508 static void softlockup_stop_all(void)
512 if (!softlockup_initialized)
515 for_each_cpu(cpu, &watchdog_allowed_mask)
516 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
518 cpumask_clear(&watchdog_allowed_mask);
521 static int softlockup_start_fn(void *data)
523 watchdog_enable(smp_processor_id());
527 static void softlockup_start_all(void)
531 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
532 for_each_cpu(cpu, &watchdog_allowed_mask)
533 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
536 int lockup_detector_online_cpu(unsigned int cpu)
538 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
539 watchdog_enable(cpu);
543 int lockup_detector_offline_cpu(unsigned int cpu)
545 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
546 watchdog_disable(cpu);
550 static void lockup_detector_reconfigure(void)
555 softlockup_stop_all();
557 lockup_detector_update_enable();
558 if (watchdog_enabled && watchdog_thresh)
559 softlockup_start_all();
561 watchdog_nmi_start();
564 * Must be called outside the cpus locked section to prevent
565 * recursive locking in the perf code.
567 __lockup_detector_cleanup();
571 * Create the watchdog thread infrastructure and configure the detector(s).
573 * The threads are not unparked as watchdog_allowed_mask is empty. When
574 * the threads are successfully initialized, take the proper locks and
575 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
577 static __init void lockup_detector_setup(void)
580 * If sysctl is off and watchdog got disabled on the command line,
581 * nothing to do here.
583 lockup_detector_update_enable();
585 if (!IS_ENABLED(CONFIG_SYSCTL) &&
586 !(watchdog_enabled && watchdog_thresh))
589 mutex_lock(&watchdog_mutex);
590 lockup_detector_reconfigure();
591 softlockup_initialized = true;
592 mutex_unlock(&watchdog_mutex);
595 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
596 static void lockup_detector_reconfigure(void)
600 lockup_detector_update_enable();
601 watchdog_nmi_start();
604 static inline void lockup_detector_setup(void)
606 lockup_detector_reconfigure();
608 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
610 static void __lockup_detector_cleanup(void)
612 lockdep_assert_held(&watchdog_mutex);
613 hardlockup_detector_perf_cleanup();
617 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
619 * Caller must not hold the cpu hotplug rwsem.
621 void lockup_detector_cleanup(void)
623 mutex_lock(&watchdog_mutex);
624 __lockup_detector_cleanup();
625 mutex_unlock(&watchdog_mutex);
629 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
631 * Special interface for parisc. It prevents lockup detector warnings from
632 * the default pm_poweroff() function which busy loops forever.
634 void lockup_detector_soft_poweroff(void)
636 watchdog_enabled = 0;
641 /* Propagate any changes to the watchdog threads */
642 static void proc_watchdog_update(void)
644 /* Remove impossible cpus to keep sysctl output clean. */
645 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
646 lockup_detector_reconfigure();
650 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
652 * caller | table->data points to | 'which'
653 * -------------------|----------------------------|--------------------------
654 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
655 * | | SOFT_WATCHDOG_ENABLED
656 * -------------------|----------------------------|--------------------------
657 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
658 * -------------------|----------------------------|--------------------------
659 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
661 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
662 void __user *buffer, size_t *lenp, loff_t *ppos)
664 int err, old, *param = table->data;
666 mutex_lock(&watchdog_mutex);
670 * On read synchronize the userspace interface. This is a
673 *param = (watchdog_enabled & which) != 0;
674 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
676 old = READ_ONCE(*param);
677 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
678 if (!err && old != READ_ONCE(*param))
679 proc_watchdog_update();
681 mutex_unlock(&watchdog_mutex);
686 * /proc/sys/kernel/watchdog
688 int proc_watchdog(struct ctl_table *table, int write,
689 void __user *buffer, size_t *lenp, loff_t *ppos)
691 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
692 table, write, buffer, lenp, ppos);
696 * /proc/sys/kernel/nmi_watchdog
698 int proc_nmi_watchdog(struct ctl_table *table, int write,
699 void __user *buffer, size_t *lenp, loff_t *ppos)
701 if (!nmi_watchdog_available && write)
703 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
704 table, write, buffer, lenp, ppos);
708 * /proc/sys/kernel/soft_watchdog
710 int proc_soft_watchdog(struct ctl_table *table, int write,
711 void __user *buffer, size_t *lenp, loff_t *ppos)
713 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
714 table, write, buffer, lenp, ppos);
718 * /proc/sys/kernel/watchdog_thresh
720 int proc_watchdog_thresh(struct ctl_table *table, int write,
721 void __user *buffer, size_t *lenp, loff_t *ppos)
725 mutex_lock(&watchdog_mutex);
727 old = READ_ONCE(watchdog_thresh);
728 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
730 if (!err && write && old != READ_ONCE(watchdog_thresh))
731 proc_watchdog_update();
733 mutex_unlock(&watchdog_mutex);
738 * The cpumask is the mask of possible cpus that the watchdog can run
739 * on, not the mask of cpus it is actually running on. This allows the
740 * user to specify a mask that will include cpus that have not yet
741 * been brought online, if desired.
743 int proc_watchdog_cpumask(struct ctl_table *table, int write,
744 void __user *buffer, size_t *lenp, loff_t *ppos)
748 mutex_lock(&watchdog_mutex);
750 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
752 proc_watchdog_update();
754 mutex_unlock(&watchdog_mutex);
757 #endif /* CONFIG_SYSCTL */
759 void __init lockup_detector_init(void)
761 if (tick_nohz_full_enabled())
762 pr_info("Disabling watchdog on nohz_full cores by default\n");
764 cpumask_copy(&watchdog_cpumask,
765 housekeeping_cpumask(HK_FLAG_TIMER));
767 if (!watchdog_nmi_probe())
768 nmi_watchdog_available = true;
769 lockup_detector_setup();