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(struct task_struct *, softlockup_task_ptr_saved);
177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
178 static unsigned long soft_lockup_nmi_warn;
180 static int __init softlockup_panic_setup(char *str)
182 softlockup_panic = simple_strtoul(str, NULL, 0);
185 __setup("softlockup_panic=", softlockup_panic_setup);
187 static int __init nowatchdog_setup(char *str)
189 watchdog_user_enabled = 0;
192 __setup("nowatchdog", nowatchdog_setup);
194 static int __init nosoftlockup_setup(char *str)
196 soft_watchdog_user_enabled = 0;
199 __setup("nosoftlockup", nosoftlockup_setup);
201 static int __init watchdog_thresh_setup(char *str)
203 get_option(&str, &watchdog_thresh);
206 __setup("watchdog_thresh=", watchdog_thresh_setup);
209 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
211 static int __init softlockup_all_cpu_backtrace_setup(char *str)
213 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
216 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
219 static void __lockup_detector_cleanup(void);
222 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
223 * lockups can have false positives under extreme conditions. So we generally
224 * want a higher threshold for soft lockups than for hard lockups. So we couple
225 * the thresholds with a factor: we make the soft threshold twice the amount of
226 * time the hard threshold is.
228 static int get_softlockup_thresh(void)
230 return watchdog_thresh * 2;
234 * Returns seconds, approximately. We don't need nanosecond
235 * resolution, and we don't need to waste time with a big divide when
238 static unsigned long get_timestamp(void)
240 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
243 static void set_sample_period(void)
246 * convert watchdog_thresh from seconds to ns
247 * the divide by 5 is to give hrtimer several chances (two
248 * or three with the current relation between the soft
249 * and hard thresholds) to increment before the
250 * hardlockup detector generates a warning
252 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
253 watchdog_update_hrtimer_threshold(sample_period);
256 /* Commands for resetting the watchdog */
257 static void __touch_watchdog(void)
259 __this_cpu_write(watchdog_touch_ts, get_timestamp());
263 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
265 * Call when the scheduler may have stalled for legitimate reasons
266 * preventing the watchdog task from executing - e.g. the scheduler
267 * entering idle state. This should only be used for scheduler events.
268 * Use touch_softlockup_watchdog() for everything else.
270 notrace void touch_softlockup_watchdog_sched(void)
273 * Preemption can be enabled. It doesn't matter which CPU's timestamp
274 * gets zeroed here, so use the raw_ operation.
276 raw_cpu_write(watchdog_touch_ts, 0);
279 notrace void touch_softlockup_watchdog(void)
281 touch_softlockup_watchdog_sched();
282 wq_watchdog_touch(raw_smp_processor_id());
284 EXPORT_SYMBOL(touch_softlockup_watchdog);
286 void touch_all_softlockup_watchdogs(void)
291 * watchdog_mutex cannpt be taken here, as this might be called
292 * from (soft)interrupt context, so the access to
293 * watchdog_allowed_cpumask might race with a concurrent update.
295 * The watchdog time stamp can race against a concurrent real
296 * update as well, the only side effect might be a cycle delay for
297 * the softlockup check.
299 for_each_cpu(cpu, &watchdog_allowed_mask)
300 per_cpu(watchdog_touch_ts, cpu) = 0;
301 wq_watchdog_touch(-1);
304 void touch_softlockup_watchdog_sync(void)
306 __this_cpu_write(softlockup_touch_sync, true);
307 __this_cpu_write(watchdog_touch_ts, 0);
310 static int is_softlockup(unsigned long touch_ts)
312 unsigned long now = get_timestamp();
314 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
315 /* Warn about unreasonable delays. */
316 if (time_after(now, touch_ts + get_softlockup_thresh()))
317 return now - touch_ts;
322 /* watchdog detector functions */
323 bool is_hardlockup(void)
325 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
327 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
330 __this_cpu_write(hrtimer_interrupts_saved, hrint);
334 static void watchdog_interrupt_count(void)
336 __this_cpu_inc(hrtimer_interrupts);
339 static DEFINE_PER_CPU(struct completion, softlockup_completion);
340 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
343 * The watchdog thread function - touches the timestamp.
345 * It only runs once every sample_period seconds (4 seconds by
346 * default) to reset the softlockup timestamp. If this gets delayed
347 * for more than 2*watchdog_thresh seconds then the debug-printout
348 * triggers in watchdog_timer_fn().
350 static int softlockup_fn(void *data)
353 complete(this_cpu_ptr(&softlockup_completion));
358 /* watchdog kicker functions */
359 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
361 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
362 struct pt_regs *regs = get_irq_regs();
364 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
366 if (!watchdog_enabled)
367 return HRTIMER_NORESTART;
369 /* kick the hardlockup detector */
370 watchdog_interrupt_count();
372 /* kick the softlockup detector */
373 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
374 reinit_completion(this_cpu_ptr(&softlockup_completion));
375 stop_one_cpu_nowait(smp_processor_id(),
377 this_cpu_ptr(&softlockup_stop_work));
381 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
384 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
386 * If the time stamp was touched atomically
387 * make sure the scheduler tick is up to date.
389 __this_cpu_write(softlockup_touch_sync, false);
393 /* Clear the guest paused flag on watchdog reset */
394 kvm_check_and_clear_guest_paused();
396 return HRTIMER_RESTART;
399 /* check for a softlockup
400 * This is done by making sure a high priority task is
401 * being scheduled. The task touches the watchdog to
402 * indicate it is getting cpu time. If it hasn't then
403 * this is a good indication some task is hogging the cpu
405 duration = is_softlockup(touch_ts);
406 if (unlikely(duration)) {
408 * If a virtual machine is stopped by the host it can look to
409 * the watchdog like a soft lockup, check to see if the host
410 * stopped the vm before we issue the warning
412 if (kvm_check_and_clear_guest_paused())
413 return HRTIMER_RESTART;
416 if (__this_cpu_read(soft_watchdog_warn) == true) {
418 * When multiple processes are causing softlockups the
419 * softlockup detector only warns on the first one
420 * because the code relies on a full quiet cycle to
421 * re-arm. The second process prevents the quiet cycle
422 * and never gets reported. Use task pointers to detect
425 if (__this_cpu_read(softlockup_task_ptr_saved) !=
427 __this_cpu_write(soft_watchdog_warn, false);
430 return HRTIMER_RESTART;
433 if (softlockup_all_cpu_backtrace) {
434 /* Prevent multiple soft-lockup reports if one cpu is already
435 * engaged in dumping cpu back traces
437 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
438 /* Someone else will report us. Let's give up */
439 __this_cpu_write(soft_watchdog_warn, true);
440 return HRTIMER_RESTART;
444 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
445 smp_processor_id(), duration,
446 current->comm, task_pid_nr(current));
447 __this_cpu_write(softlockup_task_ptr_saved, current);
449 print_irqtrace_events(current);
455 if (softlockup_all_cpu_backtrace) {
456 /* Avoid generating two back traces for current
457 * given that one is already made above
459 trigger_allbutself_cpu_backtrace();
461 clear_bit(0, &soft_lockup_nmi_warn);
462 /* Barrier to sync with other cpus */
463 smp_mb__after_atomic();
466 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
467 if (softlockup_panic)
468 panic("softlockup: hung tasks");
469 __this_cpu_write(soft_watchdog_warn, true);
471 __this_cpu_write(soft_watchdog_warn, false);
473 return HRTIMER_RESTART;
476 static void watchdog_enable(unsigned int cpu)
478 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
479 struct completion *done = this_cpu_ptr(&softlockup_completion);
481 WARN_ON_ONCE(cpu != smp_processor_id());
483 init_completion(done);
487 * Start the timer first to prevent the NMI watchdog triggering
488 * before the timer has a chance to fire.
490 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
491 hrtimer->function = watchdog_timer_fn;
492 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
493 HRTIMER_MODE_REL_PINNED_HARD);
495 /* Initialize timestamp */
497 /* Enable the perf event */
498 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
499 watchdog_nmi_enable(cpu);
502 static void watchdog_disable(unsigned int cpu)
504 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
506 WARN_ON_ONCE(cpu != smp_processor_id());
509 * Disable the perf event first. That prevents that a large delay
510 * between disabling the timer and disabling the perf event causes
511 * the perf NMI to detect a false positive.
513 watchdog_nmi_disable(cpu);
514 hrtimer_cancel(hrtimer);
515 wait_for_completion(this_cpu_ptr(&softlockup_completion));
518 static int softlockup_stop_fn(void *data)
520 watchdog_disable(smp_processor_id());
524 static void softlockup_stop_all(void)
528 if (!softlockup_initialized)
531 for_each_cpu(cpu, &watchdog_allowed_mask)
532 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
534 cpumask_clear(&watchdog_allowed_mask);
537 static int softlockup_start_fn(void *data)
539 watchdog_enable(smp_processor_id());
543 static void softlockup_start_all(void)
547 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
548 for_each_cpu(cpu, &watchdog_allowed_mask)
549 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
552 int lockup_detector_online_cpu(unsigned int cpu)
554 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
555 watchdog_enable(cpu);
559 int lockup_detector_offline_cpu(unsigned int cpu)
561 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
562 watchdog_disable(cpu);
566 static void lockup_detector_reconfigure(void)
571 softlockup_stop_all();
573 lockup_detector_update_enable();
574 if (watchdog_enabled && watchdog_thresh)
575 softlockup_start_all();
577 watchdog_nmi_start();
580 * Must be called outside the cpus locked section to prevent
581 * recursive locking in the perf code.
583 __lockup_detector_cleanup();
587 * Create the watchdog thread infrastructure and configure the detector(s).
589 * The threads are not unparked as watchdog_allowed_mask is empty. When
590 * the threads are successfully initialized, take the proper locks and
591 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
593 static __init void lockup_detector_setup(void)
596 * If sysctl is off and watchdog got disabled on the command line,
597 * nothing to do here.
599 lockup_detector_update_enable();
601 if (!IS_ENABLED(CONFIG_SYSCTL) &&
602 !(watchdog_enabled && watchdog_thresh))
605 mutex_lock(&watchdog_mutex);
606 lockup_detector_reconfigure();
607 softlockup_initialized = true;
608 mutex_unlock(&watchdog_mutex);
611 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
612 static void lockup_detector_reconfigure(void)
616 lockup_detector_update_enable();
617 watchdog_nmi_start();
620 static inline void lockup_detector_setup(void)
622 lockup_detector_reconfigure();
624 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
626 static void __lockup_detector_cleanup(void)
628 lockdep_assert_held(&watchdog_mutex);
629 hardlockup_detector_perf_cleanup();
633 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
635 * Caller must not hold the cpu hotplug rwsem.
637 void lockup_detector_cleanup(void)
639 mutex_lock(&watchdog_mutex);
640 __lockup_detector_cleanup();
641 mutex_unlock(&watchdog_mutex);
645 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
647 * Special interface for parisc. It prevents lockup detector warnings from
648 * the default pm_poweroff() function which busy loops forever.
650 void lockup_detector_soft_poweroff(void)
652 watchdog_enabled = 0;
657 /* Propagate any changes to the watchdog threads */
658 static void proc_watchdog_update(void)
660 /* Remove impossible cpus to keep sysctl output clean. */
661 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
662 lockup_detector_reconfigure();
666 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
668 * caller | table->data points to | 'which'
669 * -------------------|----------------------------|--------------------------
670 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
671 * | | SOFT_WATCHDOG_ENABLED
672 * -------------------|----------------------------|--------------------------
673 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
674 * -------------------|----------------------------|--------------------------
675 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
677 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
678 void __user *buffer, size_t *lenp, loff_t *ppos)
680 int err, old, *param = table->data;
682 mutex_lock(&watchdog_mutex);
686 * On read synchronize the userspace interface. This is a
689 *param = (watchdog_enabled & which) != 0;
690 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
692 old = READ_ONCE(*param);
693 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
694 if (!err && old != READ_ONCE(*param))
695 proc_watchdog_update();
697 mutex_unlock(&watchdog_mutex);
702 * /proc/sys/kernel/watchdog
704 int proc_watchdog(struct ctl_table *table, int write,
705 void __user *buffer, size_t *lenp, loff_t *ppos)
707 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
708 table, write, buffer, lenp, ppos);
712 * /proc/sys/kernel/nmi_watchdog
714 int proc_nmi_watchdog(struct ctl_table *table, int write,
715 void __user *buffer, size_t *lenp, loff_t *ppos)
717 if (!nmi_watchdog_available && write)
719 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
720 table, write, buffer, lenp, ppos);
724 * /proc/sys/kernel/soft_watchdog
726 int proc_soft_watchdog(struct ctl_table *table, int write,
727 void __user *buffer, size_t *lenp, loff_t *ppos)
729 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
730 table, write, buffer, lenp, ppos);
734 * /proc/sys/kernel/watchdog_thresh
736 int proc_watchdog_thresh(struct ctl_table *table, int write,
737 void __user *buffer, size_t *lenp, loff_t *ppos)
741 mutex_lock(&watchdog_mutex);
743 old = READ_ONCE(watchdog_thresh);
744 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
746 if (!err && write && old != READ_ONCE(watchdog_thresh))
747 proc_watchdog_update();
749 mutex_unlock(&watchdog_mutex);
754 * The cpumask is the mask of possible cpus that the watchdog can run
755 * on, not the mask of cpus it is actually running on. This allows the
756 * user to specify a mask that will include cpus that have not yet
757 * been brought online, if desired.
759 int proc_watchdog_cpumask(struct ctl_table *table, int write,
760 void __user *buffer, size_t *lenp, loff_t *ppos)
764 mutex_lock(&watchdog_mutex);
766 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
768 proc_watchdog_update();
770 mutex_unlock(&watchdog_mutex);
773 #endif /* CONFIG_SYSCTL */
775 void __init lockup_detector_init(void)
777 if (tick_nohz_full_enabled())
778 pr_info("Disabling watchdog on nohz_full cores by default\n");
780 cpumask_copy(&watchdog_cpumask,
781 housekeeping_cpumask(HK_FLAG_TIMER));
783 if (!watchdog_nmi_probe())
784 nmi_watchdog_available = true;
785 lockup_detector_setup();