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 struct cpumask watchdog_cpumask __read_mostly;
48 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
53 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
54 # endif /* CONFIG_SMP */
57 * Should we panic when a soft-lockup or hard-lockup occurs:
59 unsigned int __read_mostly hardlockup_panic =
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
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().
69 void __init hardlockup_detector_disable(void)
71 nmi_watchdog_user_enabled = 0;
74 static int __init hardlockup_panic_setup(char *str)
76 if (!strncmp(str, "panic", 5))
78 else if (!strncmp(str, "nopanic", 7))
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;
86 __setup("nmi_watchdog=", hardlockup_panic_setup);
88 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
91 * These functions can be overridden if an architecture implements its
92 * own hardlockup detector.
94 * watchdog_nmi_enable/disable can be implemented to start and stop when
95 * softlockup watchdog threads start and stop. The arch must select the
96 * SOFTLOCKUP_DETECTOR Kconfig.
98 int __weak watchdog_nmi_enable(unsigned int cpu)
100 hardlockup_detector_perf_enable();
104 void __weak watchdog_nmi_disable(unsigned int cpu)
106 hardlockup_detector_perf_disable();
109 /* Return 0, if a NMI watchdog is available. Error code otherwise */
110 int __weak __init watchdog_nmi_probe(void)
112 return hardlockup_detector_perf_init();
116 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
118 * The reconfiguration steps are:
119 * watchdog_nmi_stop();
120 * update_variables();
121 * watchdog_nmi_start();
123 void __weak watchdog_nmi_stop(void) { }
126 * watchdog_nmi_start - Start the watchdog after reconfiguration
128 * Counterpart to watchdog_nmi_stop().
130 * The following variables have been updated in update_variables() and
131 * contain the currently valid configuration:
136 void __weak watchdog_nmi_start(void) { }
139 * lockup_detector_update_enable - Update the sysctl enable bit
141 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
142 * can't race with watchdog_nmi_disable().
144 static void lockup_detector_update_enable(void)
146 watchdog_enabled = 0;
147 if (!watchdog_user_enabled)
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;
155 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
157 #define SOFTLOCKUP_RESET ULONG_MAX
160 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
163 static struct cpumask watchdog_allowed_mask __read_mostly;
165 /* Global variables, exported for sysctl */
166 unsigned int __read_mostly softlockup_panic =
167 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
169 static bool softlockup_initialized __read_mostly;
170 static u64 __read_mostly sample_period;
172 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
173 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
174 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
175 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
176 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
178 static unsigned long soft_lockup_nmi_warn;
180 static int __init nowatchdog_setup(char *str)
182 watchdog_user_enabled = 0;
185 __setup("nowatchdog", nowatchdog_setup);
187 static int __init nosoftlockup_setup(char *str)
189 soft_watchdog_user_enabled = 0;
192 __setup("nosoftlockup", nosoftlockup_setup);
194 static int __init watchdog_thresh_setup(char *str)
196 get_option(&str, &watchdog_thresh);
199 __setup("watchdog_thresh=", watchdog_thresh_setup);
201 static void __lockup_detector_cleanup(void);
204 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
205 * lockups can have false positives under extreme conditions. So we generally
206 * want a higher threshold for soft lockups than for hard lockups. So we couple
207 * the thresholds with a factor: we make the soft threshold twice the amount of
208 * time the hard threshold is.
210 static int get_softlockup_thresh(void)
212 return watchdog_thresh * 2;
216 * Returns seconds, approximately. We don't need nanosecond
217 * resolution, and we don't need to waste time with a big divide when
220 static unsigned long get_timestamp(void)
222 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
225 static void set_sample_period(void)
228 * convert watchdog_thresh from seconds to ns
229 * the divide by 5 is to give hrtimer several chances (two
230 * or three with the current relation between the soft
231 * and hard thresholds) to increment before the
232 * hardlockup detector generates a warning
234 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
235 watchdog_update_hrtimer_threshold(sample_period);
238 /* Commands for resetting the watchdog */
239 static void __touch_watchdog(void)
241 __this_cpu_write(watchdog_touch_ts, get_timestamp());
245 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
247 * Call when the scheduler may have stalled for legitimate reasons
248 * preventing the watchdog task from executing - e.g. the scheduler
249 * entering idle state. This should only be used for scheduler events.
250 * Use touch_softlockup_watchdog() for everything else.
252 notrace void touch_softlockup_watchdog_sched(void)
255 * Preemption can be enabled. It doesn't matter which CPU's timestamp
256 * gets zeroed here, so use the raw_ operation.
258 raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
261 notrace void touch_softlockup_watchdog(void)
263 touch_softlockup_watchdog_sched();
264 wq_watchdog_touch(raw_smp_processor_id());
266 EXPORT_SYMBOL(touch_softlockup_watchdog);
268 void touch_all_softlockup_watchdogs(void)
273 * watchdog_mutex cannpt be taken here, as this might be called
274 * from (soft)interrupt context, so the access to
275 * watchdog_allowed_cpumask might race with a concurrent update.
277 * The watchdog time stamp can race against a concurrent real
278 * update as well, the only side effect might be a cycle delay for
279 * the softlockup check.
281 for_each_cpu(cpu, &watchdog_allowed_mask)
282 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
283 wq_watchdog_touch(-1);
286 void touch_softlockup_watchdog_sync(void)
288 __this_cpu_write(softlockup_touch_sync, true);
289 __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
292 static int is_softlockup(unsigned long touch_ts)
294 unsigned long now = get_timestamp();
296 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
297 /* Warn about unreasonable delays. */
298 if (time_after(now, touch_ts + get_softlockup_thresh()))
299 return now - touch_ts;
304 /* watchdog detector functions */
305 bool is_hardlockup(void)
307 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
309 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
312 __this_cpu_write(hrtimer_interrupts_saved, hrint);
316 static void watchdog_interrupt_count(void)
318 __this_cpu_inc(hrtimer_interrupts);
321 static DEFINE_PER_CPU(struct completion, softlockup_completion);
322 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
325 * The watchdog thread function - touches the timestamp.
327 * It only runs once every sample_period seconds (4 seconds by
328 * default) to reset the softlockup timestamp. If this gets delayed
329 * for more than 2*watchdog_thresh seconds then the debug-printout
330 * triggers in watchdog_timer_fn().
332 static int softlockup_fn(void *data)
335 complete(this_cpu_ptr(&softlockup_completion));
340 /* watchdog kicker functions */
341 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
343 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
344 struct pt_regs *regs = get_irq_regs();
346 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
348 if (!watchdog_enabled)
349 return HRTIMER_NORESTART;
351 /* kick the hardlockup detector */
352 watchdog_interrupt_count();
354 /* kick the softlockup detector */
355 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
356 reinit_completion(this_cpu_ptr(&softlockup_completion));
357 stop_one_cpu_nowait(smp_processor_id(),
359 this_cpu_ptr(&softlockup_stop_work));
363 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
365 if (touch_ts == SOFTLOCKUP_RESET) {
366 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
368 * If the time stamp was touched atomically
369 * make sure the scheduler tick is up to date.
371 __this_cpu_write(softlockup_touch_sync, false);
375 /* Clear the guest paused flag on watchdog reset */
376 kvm_check_and_clear_guest_paused();
378 return HRTIMER_RESTART;
381 /* check for a softlockup
382 * This is done by making sure a high priority task is
383 * being scheduled. The task touches the watchdog to
384 * indicate it is getting cpu time. If it hasn't then
385 * this is a good indication some task is hogging the cpu
387 duration = is_softlockup(touch_ts);
388 if (unlikely(duration)) {
390 * If a virtual machine is stopped by the host it can look to
391 * the watchdog like a soft lockup, check to see if the host
392 * stopped the vm before we issue the warning
394 if (kvm_check_and_clear_guest_paused())
395 return HRTIMER_RESTART;
398 if (__this_cpu_read(soft_watchdog_warn) == true)
399 return HRTIMER_RESTART;
401 if (softlockup_all_cpu_backtrace) {
402 /* Prevent multiple soft-lockup reports if one cpu is already
403 * engaged in dumping cpu back traces
405 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
406 /* Someone else will report us. Let's give up */
407 __this_cpu_write(soft_watchdog_warn, true);
408 return HRTIMER_RESTART;
412 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
413 smp_processor_id(), duration,
414 current->comm, task_pid_nr(current));
416 print_irqtrace_events(current);
422 if (softlockup_all_cpu_backtrace) {
423 /* Avoid generating two back traces for current
424 * given that one is already made above
426 trigger_allbutself_cpu_backtrace();
428 clear_bit(0, &soft_lockup_nmi_warn);
429 /* Barrier to sync with other cpus */
430 smp_mb__after_atomic();
433 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
434 if (softlockup_panic)
435 panic("softlockup: hung tasks");
436 __this_cpu_write(soft_watchdog_warn, true);
438 __this_cpu_write(soft_watchdog_warn, false);
440 return HRTIMER_RESTART;
443 static void watchdog_enable(unsigned int cpu)
445 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
446 struct completion *done = this_cpu_ptr(&softlockup_completion);
448 WARN_ON_ONCE(cpu != smp_processor_id());
450 init_completion(done);
454 * Start the timer first to prevent the NMI watchdog triggering
455 * before the timer has a chance to fire.
457 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
458 hrtimer->function = watchdog_timer_fn;
459 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
460 HRTIMER_MODE_REL_PINNED_HARD);
462 /* Initialize timestamp */
464 /* Enable the perf event */
465 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
466 watchdog_nmi_enable(cpu);
469 static void watchdog_disable(unsigned int cpu)
471 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
473 WARN_ON_ONCE(cpu != smp_processor_id());
476 * Disable the perf event first. That prevents that a large delay
477 * between disabling the timer and disabling the perf event causes
478 * the perf NMI to detect a false positive.
480 watchdog_nmi_disable(cpu);
481 hrtimer_cancel(hrtimer);
482 wait_for_completion(this_cpu_ptr(&softlockup_completion));
485 static int softlockup_stop_fn(void *data)
487 watchdog_disable(smp_processor_id());
491 static void softlockup_stop_all(void)
495 if (!softlockup_initialized)
498 for_each_cpu(cpu, &watchdog_allowed_mask)
499 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
501 cpumask_clear(&watchdog_allowed_mask);
504 static int softlockup_start_fn(void *data)
506 watchdog_enable(smp_processor_id());
510 static void softlockup_start_all(void)
514 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
515 for_each_cpu(cpu, &watchdog_allowed_mask)
516 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
519 int lockup_detector_online_cpu(unsigned int cpu)
521 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
522 watchdog_enable(cpu);
526 int lockup_detector_offline_cpu(unsigned int cpu)
528 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
529 watchdog_disable(cpu);
533 static void lockup_detector_reconfigure(void)
538 softlockup_stop_all();
540 lockup_detector_update_enable();
541 if (watchdog_enabled && watchdog_thresh)
542 softlockup_start_all();
544 watchdog_nmi_start();
547 * Must be called outside the cpus locked section to prevent
548 * recursive locking in the perf code.
550 __lockup_detector_cleanup();
554 * Create the watchdog thread infrastructure and configure the detector(s).
556 * The threads are not unparked as watchdog_allowed_mask is empty. When
557 * the threads are successfully initialized, take the proper locks and
558 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
560 static __init void lockup_detector_setup(void)
563 * If sysctl is off and watchdog got disabled on the command line,
564 * nothing to do here.
566 lockup_detector_update_enable();
568 if (!IS_ENABLED(CONFIG_SYSCTL) &&
569 !(watchdog_enabled && watchdog_thresh))
572 mutex_lock(&watchdog_mutex);
573 lockup_detector_reconfigure();
574 softlockup_initialized = true;
575 mutex_unlock(&watchdog_mutex);
578 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
579 static void lockup_detector_reconfigure(void)
583 lockup_detector_update_enable();
584 watchdog_nmi_start();
587 static inline void lockup_detector_setup(void)
589 lockup_detector_reconfigure();
591 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
593 static void __lockup_detector_cleanup(void)
595 lockdep_assert_held(&watchdog_mutex);
596 hardlockup_detector_perf_cleanup();
600 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
602 * Caller must not hold the cpu hotplug rwsem.
604 void lockup_detector_cleanup(void)
606 mutex_lock(&watchdog_mutex);
607 __lockup_detector_cleanup();
608 mutex_unlock(&watchdog_mutex);
612 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
614 * Special interface for parisc. It prevents lockup detector warnings from
615 * the default pm_poweroff() function which busy loops forever.
617 void lockup_detector_soft_poweroff(void)
619 watchdog_enabled = 0;
624 /* Propagate any changes to the watchdog threads */
625 static void proc_watchdog_update(void)
627 /* Remove impossible cpus to keep sysctl output clean. */
628 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
629 lockup_detector_reconfigure();
633 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
635 * caller | table->data points to | 'which'
636 * -------------------|----------------------------|--------------------------
637 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
638 * | | SOFT_WATCHDOG_ENABLED
639 * -------------------|----------------------------|--------------------------
640 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
641 * -------------------|----------------------------|--------------------------
642 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
644 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
645 void *buffer, size_t *lenp, loff_t *ppos)
647 int err, old, *param = table->data;
649 mutex_lock(&watchdog_mutex);
653 * On read synchronize the userspace interface. This is a
656 *param = (watchdog_enabled & which) != 0;
657 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
659 old = READ_ONCE(*param);
660 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
661 if (!err && old != READ_ONCE(*param))
662 proc_watchdog_update();
664 mutex_unlock(&watchdog_mutex);
669 * /proc/sys/kernel/watchdog
671 int proc_watchdog(struct ctl_table *table, int write,
672 void *buffer, size_t *lenp, loff_t *ppos)
674 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
675 table, write, buffer, lenp, ppos);
679 * /proc/sys/kernel/nmi_watchdog
681 int proc_nmi_watchdog(struct ctl_table *table, int write,
682 void *buffer, size_t *lenp, loff_t *ppos)
684 if (!nmi_watchdog_available && write)
686 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
687 table, write, buffer, lenp, ppos);
691 * /proc/sys/kernel/soft_watchdog
693 int proc_soft_watchdog(struct ctl_table *table, int write,
694 void *buffer, size_t *lenp, loff_t *ppos)
696 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
697 table, write, buffer, lenp, ppos);
701 * /proc/sys/kernel/watchdog_thresh
703 int proc_watchdog_thresh(struct ctl_table *table, int write,
704 void *buffer, size_t *lenp, loff_t *ppos)
708 mutex_lock(&watchdog_mutex);
710 old = READ_ONCE(watchdog_thresh);
711 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
713 if (!err && write && old != READ_ONCE(watchdog_thresh))
714 proc_watchdog_update();
716 mutex_unlock(&watchdog_mutex);
721 * The cpumask is the mask of possible cpus that the watchdog can run
722 * on, not the mask of cpus it is actually running on. This allows the
723 * user to specify a mask that will include cpus that have not yet
724 * been brought online, if desired.
726 int proc_watchdog_cpumask(struct ctl_table *table, int write,
727 void *buffer, size_t *lenp, loff_t *ppos)
731 mutex_lock(&watchdog_mutex);
733 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
735 proc_watchdog_update();
737 mutex_unlock(&watchdog_mutex);
740 #endif /* CONFIG_SYSCTL */
742 void __init lockup_detector_init(void)
744 if (tick_nohz_full_enabled())
745 pr_info("Disabling watchdog on nohz_full cores by default\n");
747 cpumask_copy(&watchdog_cpumask,
748 housekeeping_cpumask(HK_FLAG_TIMER));
750 if (!watchdog_nmi_probe())
751 nmi_watchdog_available = true;
752 lockup_detector_setup();