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 int __read_mostly nmi_watchdog_available;
47 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 #define SOFTLOCKUP_RESET ULONG_MAX
166 /* Global variables, exported for sysctl */
167 unsigned int __read_mostly softlockup_panic =
168 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
170 static bool softlockup_initialized __read_mostly;
171 static u64 __read_mostly sample_period;
173 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
174 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
175 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
176 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
178 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
179 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
180 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
181 static unsigned long soft_lockup_nmi_warn;
183 static int __init softlockup_panic_setup(char *str)
185 softlockup_panic = simple_strtoul(str, NULL, 0);
188 __setup("softlockup_panic=", softlockup_panic_setup);
190 static int __init nowatchdog_setup(char *str)
192 watchdog_user_enabled = 0;
195 __setup("nowatchdog", nowatchdog_setup);
197 static int __init nosoftlockup_setup(char *str)
199 soft_watchdog_user_enabled = 0;
202 __setup("nosoftlockup", nosoftlockup_setup);
205 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
207 static int __init softlockup_all_cpu_backtrace_setup(char *str)
209 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
212 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
215 static void __lockup_detector_cleanup(void);
218 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
219 * lockups can have false positives under extreme conditions. So we generally
220 * want a higher threshold for soft lockups than for hard lockups. So we couple
221 * the thresholds with a factor: we make the soft threshold twice the amount of
222 * time the hard threshold is.
224 static int get_softlockup_thresh(void)
226 return watchdog_thresh * 2;
230 * Returns seconds, approximately. We don't need nanosecond
231 * resolution, and we don't need to waste time with a big divide when
234 static unsigned long get_timestamp(void)
236 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
239 static void set_sample_period(void)
242 * convert watchdog_thresh from seconds to ns
243 * the divide by 5 is to give hrtimer several chances (two
244 * or three with the current relation between the soft
245 * and hard thresholds) to increment before the
246 * hardlockup detector generates a warning
248 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
249 watchdog_update_hrtimer_threshold(sample_period);
252 /* Commands for resetting the watchdog */
253 static void __touch_watchdog(void)
255 __this_cpu_write(watchdog_touch_ts, get_timestamp());
259 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
261 * Call when the scheduler may have stalled for legitimate reasons
262 * preventing the watchdog task from executing - e.g. the scheduler
263 * entering idle state. This should only be used for scheduler events.
264 * Use touch_softlockup_watchdog() for everything else.
266 notrace void touch_softlockup_watchdog_sched(void)
269 * Preemption can be enabled. It doesn't matter which CPU's timestamp
270 * gets zeroed here, so use the raw_ operation.
272 raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
275 notrace void touch_softlockup_watchdog(void)
277 touch_softlockup_watchdog_sched();
278 wq_watchdog_touch(raw_smp_processor_id());
280 EXPORT_SYMBOL(touch_softlockup_watchdog);
282 void touch_all_softlockup_watchdogs(void)
287 * watchdog_mutex cannpt be taken here, as this might be called
288 * from (soft)interrupt context, so the access to
289 * watchdog_allowed_cpumask might race with a concurrent update.
291 * The watchdog time stamp can race against a concurrent real
292 * update as well, the only side effect might be a cycle delay for
293 * the softlockup check.
295 for_each_cpu(cpu, &watchdog_allowed_mask)
296 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
297 wq_watchdog_touch(-1);
300 void touch_softlockup_watchdog_sync(void)
302 __this_cpu_write(softlockup_touch_sync, true);
303 __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
306 static int is_softlockup(unsigned long touch_ts)
308 unsigned long now = get_timestamp();
310 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
311 /* Warn about unreasonable delays. */
312 if (time_after(now, touch_ts + get_softlockup_thresh()))
313 return now - touch_ts;
318 /* watchdog detector functions */
319 bool is_hardlockup(void)
321 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
323 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
326 __this_cpu_write(hrtimer_interrupts_saved, hrint);
330 static void watchdog_interrupt_count(void)
332 __this_cpu_inc(hrtimer_interrupts);
335 static DEFINE_PER_CPU(struct completion, softlockup_completion);
336 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
339 * The watchdog thread function - touches the timestamp.
341 * It only runs once every sample_period seconds (4 seconds by
342 * default) to reset the softlockup timestamp. If this gets delayed
343 * for more than 2*watchdog_thresh seconds then the debug-printout
344 * triggers in watchdog_timer_fn().
346 static int softlockup_fn(void *data)
348 __this_cpu_write(soft_lockup_hrtimer_cnt,
349 __this_cpu_read(hrtimer_interrupts));
351 complete(this_cpu_ptr(&softlockup_completion));
356 /* watchdog kicker functions */
357 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
359 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
360 struct pt_regs *regs = get_irq_regs();
362 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
364 if (!watchdog_enabled)
365 return HRTIMER_NORESTART;
367 /* kick the hardlockup detector */
368 watchdog_interrupt_count();
370 /* kick the softlockup detector */
371 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
372 reinit_completion(this_cpu_ptr(&softlockup_completion));
373 stop_one_cpu_nowait(smp_processor_id(),
375 this_cpu_ptr(&softlockup_stop_work));
379 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
381 if (touch_ts == SOFTLOCKUP_RESET) {
382 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
384 * If the time stamp was touched atomically
385 * make sure the scheduler tick is up to date.
387 __this_cpu_write(softlockup_touch_sync, false);
391 /* Clear the guest paused flag on watchdog reset */
392 kvm_check_and_clear_guest_paused();
394 return HRTIMER_RESTART;
397 /* check for a softlockup
398 * This is done by making sure a high priority task is
399 * being scheduled. The task touches the watchdog to
400 * indicate it is getting cpu time. If it hasn't then
401 * this is a good indication some task is hogging the cpu
403 duration = is_softlockup(touch_ts);
404 if (unlikely(duration)) {
406 * If a virtual machine is stopped by the host it can look to
407 * the watchdog like a soft lockup, check to see if the host
408 * stopped the vm before we issue the warning
410 if (kvm_check_and_clear_guest_paused())
411 return HRTIMER_RESTART;
414 if (__this_cpu_read(soft_watchdog_warn) == true) {
416 * When multiple processes are causing softlockups the
417 * softlockup detector only warns on the first one
418 * because the code relies on a full quiet cycle to
419 * re-arm. The second process prevents the quiet cycle
420 * and never gets reported. Use task pointers to detect
423 if (__this_cpu_read(softlockup_task_ptr_saved) !=
425 __this_cpu_write(soft_watchdog_warn, false);
428 return HRTIMER_RESTART;
431 if (softlockup_all_cpu_backtrace) {
432 /* Prevent multiple soft-lockup reports if one cpu is already
433 * engaged in dumping cpu back traces
435 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
436 /* Someone else will report us. Let's give up */
437 __this_cpu_write(soft_watchdog_warn, true);
438 return HRTIMER_RESTART;
442 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
443 smp_processor_id(), duration,
444 current->comm, task_pid_nr(current));
445 __this_cpu_write(softlockup_task_ptr_saved, current);
447 print_irqtrace_events(current);
453 if (softlockup_all_cpu_backtrace) {
454 /* Avoid generating two back traces for current
455 * given that one is already made above
457 trigger_allbutself_cpu_backtrace();
459 clear_bit(0, &soft_lockup_nmi_warn);
460 /* Barrier to sync with other cpus */
461 smp_mb__after_atomic();
464 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
465 if (softlockup_panic)
466 panic("softlockup: hung tasks");
467 __this_cpu_write(soft_watchdog_warn, true);
469 __this_cpu_write(soft_watchdog_warn, false);
471 return HRTIMER_RESTART;
474 static void watchdog_enable(unsigned int cpu)
476 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
477 struct completion *done = this_cpu_ptr(&softlockup_completion);
479 WARN_ON_ONCE(cpu != smp_processor_id());
481 init_completion(done);
485 * Start the timer first to prevent the NMI watchdog triggering
486 * before the timer has a chance to fire.
488 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
489 hrtimer->function = watchdog_timer_fn;
490 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
491 HRTIMER_MODE_REL_PINNED);
493 /* Initialize timestamp */
495 /* Enable the perf event */
496 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
497 watchdog_nmi_enable(cpu);
500 static void watchdog_disable(unsigned int cpu)
502 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
504 WARN_ON_ONCE(cpu != smp_processor_id());
507 * Disable the perf event first. That prevents that a large delay
508 * between disabling the timer and disabling the perf event causes
509 * the perf NMI to detect a false positive.
511 watchdog_nmi_disable(cpu);
512 hrtimer_cancel(hrtimer);
513 wait_for_completion(this_cpu_ptr(&softlockup_completion));
516 static int softlockup_stop_fn(void *data)
518 watchdog_disable(smp_processor_id());
522 static void softlockup_stop_all(void)
526 if (!softlockup_initialized)
529 for_each_cpu(cpu, &watchdog_allowed_mask)
530 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
532 cpumask_clear(&watchdog_allowed_mask);
535 static int softlockup_start_fn(void *data)
537 watchdog_enable(smp_processor_id());
541 static void softlockup_start_all(void)
545 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
546 for_each_cpu(cpu, &watchdog_allowed_mask)
547 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
550 int lockup_detector_online_cpu(unsigned int cpu)
552 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
553 watchdog_enable(cpu);
557 int lockup_detector_offline_cpu(unsigned int cpu)
559 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
560 watchdog_disable(cpu);
564 static void lockup_detector_reconfigure(void)
569 softlockup_stop_all();
571 lockup_detector_update_enable();
572 if (watchdog_enabled && watchdog_thresh)
573 softlockup_start_all();
575 watchdog_nmi_start();
578 * Must be called outside the cpus locked section to prevent
579 * recursive locking in the perf code.
581 __lockup_detector_cleanup();
585 * Create the watchdog thread infrastructure and configure the detector(s).
587 * The threads are not unparked as watchdog_allowed_mask is empty. When
588 * the threads are sucessfully initialized, take the proper locks and
589 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
591 static __init void lockup_detector_setup(void)
594 * If sysctl is off and watchdog got disabled on the command line,
595 * nothing to do here.
597 lockup_detector_update_enable();
599 if (!IS_ENABLED(CONFIG_SYSCTL) &&
600 !(watchdog_enabled && watchdog_thresh))
603 mutex_lock(&watchdog_mutex);
604 lockup_detector_reconfigure();
605 softlockup_initialized = true;
606 mutex_unlock(&watchdog_mutex);
609 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
610 static void lockup_detector_reconfigure(void)
614 lockup_detector_update_enable();
615 watchdog_nmi_start();
618 static inline void lockup_detector_setup(void)
620 lockup_detector_reconfigure();
622 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
624 static void __lockup_detector_cleanup(void)
626 lockdep_assert_held(&watchdog_mutex);
627 hardlockup_detector_perf_cleanup();
631 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
633 * Caller must not hold the cpu hotplug rwsem.
635 void lockup_detector_cleanup(void)
637 mutex_lock(&watchdog_mutex);
638 __lockup_detector_cleanup();
639 mutex_unlock(&watchdog_mutex);
643 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
645 * Special interface for parisc. It prevents lockup detector warnings from
646 * the default pm_poweroff() function which busy loops forever.
648 void lockup_detector_soft_poweroff(void)
650 watchdog_enabled = 0;
655 /* Propagate any changes to the watchdog threads */
656 static void proc_watchdog_update(void)
658 /* Remove impossible cpus to keep sysctl output clean. */
659 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
660 lockup_detector_reconfigure();
664 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
666 * caller | table->data points to | 'which'
667 * -------------------|----------------------------|--------------------------
668 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
669 * | | SOFT_WATCHDOG_ENABLED
670 * -------------------|----------------------------|--------------------------
671 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
672 * -------------------|----------------------------|--------------------------
673 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
675 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
676 void __user *buffer, size_t *lenp, loff_t *ppos)
678 int err, old, *param = table->data;
680 mutex_lock(&watchdog_mutex);
684 * On read synchronize the userspace interface. This is a
687 *param = (watchdog_enabled & which) != 0;
688 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
690 old = READ_ONCE(*param);
691 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
692 if (!err && old != READ_ONCE(*param))
693 proc_watchdog_update();
695 mutex_unlock(&watchdog_mutex);
700 * /proc/sys/kernel/watchdog
702 int proc_watchdog(struct ctl_table *table, int write,
703 void __user *buffer, size_t *lenp, loff_t *ppos)
705 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
706 table, write, buffer, lenp, ppos);
710 * /proc/sys/kernel/nmi_watchdog
712 int proc_nmi_watchdog(struct ctl_table *table, int write,
713 void __user *buffer, size_t *lenp, loff_t *ppos)
715 if (!nmi_watchdog_available && write)
717 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
718 table, write, buffer, lenp, ppos);
722 * /proc/sys/kernel/soft_watchdog
724 int proc_soft_watchdog(struct ctl_table *table, int write,
725 void __user *buffer, size_t *lenp, loff_t *ppos)
727 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
728 table, write, buffer, lenp, ppos);
732 * /proc/sys/kernel/watchdog_thresh
734 int proc_watchdog_thresh(struct ctl_table *table, int write,
735 void __user *buffer, size_t *lenp, loff_t *ppos)
739 mutex_lock(&watchdog_mutex);
741 old = READ_ONCE(watchdog_thresh);
742 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
744 if (!err && write && old != READ_ONCE(watchdog_thresh))
745 proc_watchdog_update();
747 mutex_unlock(&watchdog_mutex);
752 * The cpumask is the mask of possible cpus that the watchdog can run
753 * on, not the mask of cpus it is actually running on. This allows the
754 * user to specify a mask that will include cpus that have not yet
755 * been brought online, if desired.
757 int proc_watchdog_cpumask(struct ctl_table *table, int write,
758 void __user *buffer, size_t *lenp, loff_t *ppos)
762 mutex_lock(&watchdog_mutex);
764 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
766 proc_watchdog_update();
768 mutex_unlock(&watchdog_mutex);
771 #endif /* CONFIG_SYSCTL */
773 void __init lockup_detector_init(void)
775 if (tick_nohz_full_enabled())
776 pr_info("Disabling watchdog on nohz_full cores by default\n");
778 cpumask_copy(&watchdog_cpumask,
779 housekeeping_cpumask(HK_FLAG_TIMER));
781 if (!watchdog_nmi_probe())
782 nmi_watchdog_available = true;
783 lockup_detector_setup();