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) "NMI 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 <linux/tick.h>
23 #include <linux/workqueue.h>
25 #include <asm/irq_regs.h>
26 #include <linux/kvm_para.h>
27 #include <linux/perf_event.h>
28 #include <linux/kthread.h>
31 * The run state of the lockup detectors is controlled by the content of the
32 * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
33 * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
35 * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
36 * are variables that are only used as an 'interface' between the parameters
37 * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
38 * 'watchdog_thresh' variable is handled differently because its value is not
39 * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
42 #define NMI_WATCHDOG_ENABLED_BIT 0
43 #define SOFT_WATCHDOG_ENABLED_BIT 1
44 #define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT)
45 #define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT)
47 static DEFINE_MUTEX(watchdog_proc_mutex);
49 #ifdef CONFIG_HARDLOCKUP_DETECTOR
50 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
52 static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
54 int __read_mostly nmi_watchdog_enabled;
55 int __read_mostly soft_watchdog_enabled;
56 int __read_mostly watchdog_user_enabled;
57 int __read_mostly watchdog_thresh = 10;
60 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
61 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
63 #define sysctl_softlockup_all_cpu_backtrace 0
64 #define sysctl_hardlockup_all_cpu_backtrace 0
66 static struct cpumask watchdog_cpumask __read_mostly;
67 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
69 /* Helper for online, unparked cpus. */
70 #define for_each_watchdog_cpu(cpu) \
71 for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
74 * The 'watchdog_running' variable is set to 1 when the watchdog threads
75 * are registered/started and is set to 0 when the watchdog threads are
76 * unregistered/stopped, so it is an indicator whether the threads exist.
78 static int __read_mostly watchdog_running;
80 * If a subsystem has a need to deactivate the watchdog temporarily, it
81 * can use the suspend/resume interface to achieve this. The content of
82 * the 'watchdog_suspended' variable reflects this state. Existing threads
83 * are parked/unparked by the lockup_detector_{suspend|resume} functions
84 * (see comment blocks pertaining to those functions for further details).
86 * 'watchdog_suspended' also prevents threads from being registered/started
87 * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
88 * of 'watchdog_running' cannot change while the watchdog is deactivated
89 * temporarily (see related code in 'proc' handlers).
91 static int __read_mostly watchdog_suspended;
93 static u64 __read_mostly sample_period;
95 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
96 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
97 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
98 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
99 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
100 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
101 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
102 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
103 #ifdef CONFIG_HARDLOCKUP_DETECTOR
104 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
105 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
106 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
107 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
109 static unsigned long soft_lockup_nmi_warn;
113 * Should we panic when a soft-lockup or hard-lockup occurs:
115 #ifdef CONFIG_HARDLOCKUP_DETECTOR
116 unsigned int __read_mostly hardlockup_panic =
117 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
118 static unsigned long hardlockup_allcpu_dumped;
120 * We may not want to enable hard lockup detection by default in all cases,
121 * for example when running the kernel as a guest on a hypervisor. In these
122 * cases this function can be called to disable hard lockup detection. This
123 * function should only be executed once by the boot processor before the
124 * kernel command line parameters are parsed, because otherwise it is not
125 * possible to override this in hardlockup_panic_setup().
127 void hardlockup_detector_disable(void)
129 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
132 static int __init hardlockup_panic_setup(char *str)
134 if (!strncmp(str, "panic", 5))
135 hardlockup_panic = 1;
136 else if (!strncmp(str, "nopanic", 7))
137 hardlockup_panic = 0;
138 else if (!strncmp(str, "0", 1))
139 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
140 else if (!strncmp(str, "1", 1))
141 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
144 __setup("nmi_watchdog=", hardlockup_panic_setup);
147 unsigned int __read_mostly softlockup_panic =
148 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
150 static int __init softlockup_panic_setup(char *str)
152 softlockup_panic = simple_strtoul(str, NULL, 0);
156 __setup("softlockup_panic=", softlockup_panic_setup);
158 static int __init nowatchdog_setup(char *str)
160 watchdog_enabled = 0;
163 __setup("nowatchdog", nowatchdog_setup);
165 static int __init nosoftlockup_setup(char *str)
167 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
170 __setup("nosoftlockup", nosoftlockup_setup);
173 static int __init softlockup_all_cpu_backtrace_setup(char *str)
175 sysctl_softlockup_all_cpu_backtrace =
176 !!simple_strtol(str, NULL, 0);
179 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
180 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
182 sysctl_hardlockup_all_cpu_backtrace =
183 !!simple_strtol(str, NULL, 0);
186 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
190 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
191 * lockups can have false positives under extreme conditions. So we generally
192 * want a higher threshold for soft lockups than for hard lockups. So we couple
193 * the thresholds with a factor: we make the soft threshold twice the amount of
194 * time the hard threshold is.
196 static int get_softlockup_thresh(void)
198 return watchdog_thresh * 2;
202 * Returns seconds, approximately. We don't need nanosecond
203 * resolution, and we don't need to waste time with a big divide when
206 static unsigned long get_timestamp(void)
208 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
211 static void set_sample_period(void)
214 * convert watchdog_thresh from seconds to ns
215 * the divide by 5 is to give hrtimer several chances (two
216 * or three with the current relation between the soft
217 * and hard thresholds) to increment before the
218 * hardlockup detector generates a warning
220 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
223 /* Commands for resetting the watchdog */
224 static void __touch_watchdog(void)
226 __this_cpu_write(watchdog_touch_ts, get_timestamp());
230 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
232 * Call when the scheduler may have stalled for legitimate reasons
233 * preventing the watchdog task from executing - e.g. the scheduler
234 * entering idle state. This should only be used for scheduler events.
235 * Use touch_softlockup_watchdog() for everything else.
237 void touch_softlockup_watchdog_sched(void)
240 * Preemption can be enabled. It doesn't matter which CPU's timestamp
241 * gets zeroed here, so use the raw_ operation.
243 raw_cpu_write(watchdog_touch_ts, 0);
246 void touch_softlockup_watchdog(void)
248 touch_softlockup_watchdog_sched();
249 wq_watchdog_touch(raw_smp_processor_id());
251 EXPORT_SYMBOL(touch_softlockup_watchdog);
253 void touch_all_softlockup_watchdogs(void)
258 * this is done lockless
259 * do we care if a 0 races with a timestamp?
260 * all it means is the softlock check starts one cycle later
262 for_each_watchdog_cpu(cpu)
263 per_cpu(watchdog_touch_ts, cpu) = 0;
264 wq_watchdog_touch(-1);
267 #ifdef CONFIG_HARDLOCKUP_DETECTOR
268 void touch_nmi_watchdog(void)
271 * Using __raw here because some code paths have
272 * preemption enabled. If preemption is enabled
273 * then interrupts should be enabled too, in which
274 * case we shouldn't have to worry about the watchdog
277 raw_cpu_write(watchdog_nmi_touch, true);
278 touch_softlockup_watchdog();
280 EXPORT_SYMBOL(touch_nmi_watchdog);
284 void touch_softlockup_watchdog_sync(void)
286 __this_cpu_write(softlockup_touch_sync, true);
287 __this_cpu_write(watchdog_touch_ts, 0);
290 #ifdef CONFIG_HARDLOCKUP_DETECTOR
291 /* watchdog detector functions */
292 static bool is_hardlockup(void)
294 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
296 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
299 __this_cpu_write(hrtimer_interrupts_saved, hrint);
304 static int is_softlockup(unsigned long touch_ts)
306 unsigned long now = get_timestamp();
308 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
309 /* Warn about unreasonable delays. */
310 if (time_after(now, touch_ts + get_softlockup_thresh()))
311 return now - touch_ts;
316 #ifdef CONFIG_HARDLOCKUP_DETECTOR
318 static DEFINE_RAW_SPINLOCK(watchdog_output_lock);
320 static struct perf_event_attr wd_hw_attr = {
321 .type = PERF_TYPE_HARDWARE,
322 .config = PERF_COUNT_HW_CPU_CYCLES,
323 .size = sizeof(struct perf_event_attr),
328 /* Callback function for perf event subsystem */
329 static void watchdog_overflow_callback(struct perf_event *event,
330 struct perf_sample_data *data,
331 struct pt_regs *regs)
333 /* Ensure the watchdog never gets throttled */
334 event->hw.interrupts = 0;
336 if (__this_cpu_read(watchdog_nmi_touch) == true) {
337 __this_cpu_write(watchdog_nmi_touch, false);
341 /* check for a hardlockup
342 * This is done by making sure our timer interrupt
343 * is incrementing. The timer interrupt should have
344 * fired multiple times before we overflow'd. If it hasn't
345 * then this is a good indication the cpu is stuck
347 if (is_hardlockup()) {
348 int this_cpu = smp_processor_id();
349 struct pt_regs *regs = get_irq_regs();
351 /* only print hardlockups once */
352 if (__this_cpu_read(hard_watchdog_warn) == true)
355 * If early-printk is enabled then make sure we do not
356 * lock up in printk() and kill console logging:
360 raw_spin_lock(&watchdog_output_lock);
362 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
364 print_irqtrace_events(current);
371 * Perform all-CPU dump only once to avoid multiple hardlockups
372 * generating interleaving traces
374 if (sysctl_hardlockup_all_cpu_backtrace &&
375 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
376 trigger_allbutself_cpu_backtrace();
378 raw_spin_unlock(&watchdog_output_lock);
379 if (hardlockup_panic)
380 nmi_panic(regs, "Hard LOCKUP");
382 __this_cpu_write(hard_watchdog_warn, true);
386 __this_cpu_write(hard_watchdog_warn, false);
389 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
391 static void watchdog_interrupt_count(void)
393 __this_cpu_inc(hrtimer_interrupts);
396 static int watchdog_nmi_enable(unsigned int cpu);
397 static void watchdog_nmi_disable(unsigned int cpu);
399 static int watchdog_enable_all_cpus(void);
400 static void watchdog_disable_all_cpus(void);
402 /* watchdog kicker functions */
403 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
405 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
406 struct pt_regs *regs = get_irq_regs();
408 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
410 /* kick the hardlockup detector */
411 watchdog_interrupt_count();
413 /* kick the softlockup detector */
414 wake_up_process(__this_cpu_read(softlockup_watchdog));
417 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
420 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
422 * If the time stamp was touched atomically
423 * make sure the scheduler tick is up to date.
425 __this_cpu_write(softlockup_touch_sync, false);
429 /* Clear the guest paused flag on watchdog reset */
430 kvm_check_and_clear_guest_paused();
432 return HRTIMER_RESTART;
435 /* check for a softlockup
436 * This is done by making sure a high priority task is
437 * being scheduled. The task touches the watchdog to
438 * indicate it is getting cpu time. If it hasn't then
439 * this is a good indication some task is hogging the cpu
441 duration = is_softlockup(touch_ts);
442 if (unlikely(duration)) {
444 * If a virtual machine is stopped by the host it can look to
445 * the watchdog like a soft lockup, check to see if the host
446 * stopped the vm before we issue the warning
448 if (kvm_check_and_clear_guest_paused())
449 return HRTIMER_RESTART;
452 if (__this_cpu_read(soft_watchdog_warn) == true) {
454 * When multiple processes are causing softlockups the
455 * softlockup detector only warns on the first one
456 * because the code relies on a full quiet cycle to
457 * re-arm. The second process prevents the quiet cycle
458 * and never gets reported. Use task pointers to detect
461 if (__this_cpu_read(softlockup_task_ptr_saved) !=
463 __this_cpu_write(soft_watchdog_warn, false);
466 return HRTIMER_RESTART;
469 if (softlockup_all_cpu_backtrace) {
470 /* Prevent multiple soft-lockup reports if one cpu is already
471 * engaged in dumping cpu back traces
473 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
474 /* Someone else will report us. Let's give up */
475 __this_cpu_write(soft_watchdog_warn, true);
476 return HRTIMER_RESTART;
480 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
481 smp_processor_id(), duration,
482 current->comm, task_pid_nr(current));
483 __this_cpu_write(softlockup_task_ptr_saved, current);
485 print_irqtrace_events(current);
491 if (softlockup_all_cpu_backtrace) {
492 /* Avoid generating two back traces for current
493 * given that one is already made above
495 trigger_allbutself_cpu_backtrace();
497 clear_bit(0, &soft_lockup_nmi_warn);
498 /* Barrier to sync with other cpus */
499 smp_mb__after_atomic();
502 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
503 if (softlockup_panic)
504 panic("softlockup: hung tasks");
505 __this_cpu_write(soft_watchdog_warn, true);
507 __this_cpu_write(soft_watchdog_warn, false);
509 return HRTIMER_RESTART;
512 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
514 struct sched_param param = { .sched_priority = prio };
516 sched_setscheduler(current, policy, ¶m);
519 static void watchdog_enable(unsigned int cpu)
521 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
523 /* kick off the timer for the hardlockup detector */
524 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
525 hrtimer->function = watchdog_timer_fn;
526 hrtimer->irqsafe = 1;
528 /* Enable the perf event */
529 watchdog_nmi_enable(cpu);
531 /* done here because hrtimer_start can only pin to smp_processor_id() */
532 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
533 HRTIMER_MODE_REL_PINNED);
535 /* initialize timestamp */
536 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
540 static void watchdog_disable(unsigned int cpu)
542 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
544 watchdog_set_prio(SCHED_NORMAL, 0);
545 hrtimer_cancel(hrtimer);
546 /* disable the perf event */
547 watchdog_nmi_disable(cpu);
550 static void watchdog_cleanup(unsigned int cpu, bool online)
552 watchdog_disable(cpu);
555 static int watchdog_should_run(unsigned int cpu)
557 return __this_cpu_read(hrtimer_interrupts) !=
558 __this_cpu_read(soft_lockup_hrtimer_cnt);
562 * The watchdog thread function - touches the timestamp.
564 * It only runs once every sample_period seconds (4 seconds by
565 * default) to reset the softlockup timestamp. If this gets delayed
566 * for more than 2*watchdog_thresh seconds then the debug-printout
567 * triggers in watchdog_timer_fn().
569 static void watchdog(unsigned int cpu)
571 __this_cpu_write(soft_lockup_hrtimer_cnt,
572 __this_cpu_read(hrtimer_interrupts));
576 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
577 * failure path. Check for failures that can occur asynchronously -
578 * for example, when CPUs are on-lined - and shut down the hardware
579 * perf event on each CPU accordingly.
581 * The only non-obvious place this bit can be cleared is through
582 * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a
583 * pr_info here would be too noisy as it would result in a message
584 * every few seconds if the hardlockup was disabled but the softlockup
587 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
588 watchdog_nmi_disable(cpu);
591 #ifdef CONFIG_HARDLOCKUP_DETECTOR
593 * People like the simple clean cpu node info on boot.
594 * Reduce the watchdog noise by only printing messages
595 * that are different from what cpu0 displayed.
597 static unsigned long cpu0_err;
599 static int watchdog_nmi_enable(unsigned int cpu)
601 struct perf_event_attr *wd_attr;
602 struct perf_event *event = per_cpu(watchdog_ev, cpu);
604 /* nothing to do if the hard lockup detector is disabled */
605 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
608 /* is it already setup and enabled? */
609 if (event && event->state > PERF_EVENT_STATE_OFF)
612 /* it is setup but not enabled */
616 wd_attr = &wd_hw_attr;
617 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
619 /* Try to register using hardware perf events */
620 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
622 /* save cpu0 error for future comparision */
623 if (cpu == 0 && IS_ERR(event))
624 cpu0_err = PTR_ERR(event);
626 if (!IS_ERR(event)) {
627 /* only print for cpu0 or different than cpu0 */
628 if (cpu == 0 || cpu0_err)
629 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
634 * Disable the hard lockup detector if _any_ CPU fails to set up
635 * set up the hardware perf event. The watchdog() function checks
636 * the NMI_WATCHDOG_ENABLED bit periodically.
638 * The barriers are for syncing up watchdog_enabled across all the
639 * cpus, as clear_bit() does not use barriers.
641 smp_mb__before_atomic();
642 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
643 smp_mb__after_atomic();
645 /* skip displaying the same error again */
646 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
647 return PTR_ERR(event);
649 /* vary the KERN level based on the returned errno */
650 if (PTR_ERR(event) == -EOPNOTSUPP)
651 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
652 else if (PTR_ERR(event) == -ENOENT)
653 pr_warn("disabled (cpu%i): hardware events not enabled\n",
656 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
657 cpu, PTR_ERR(event));
659 pr_info("Shutting down hard lockup detector on all cpus\n");
661 return PTR_ERR(event);
665 per_cpu(watchdog_ev, cpu) = event;
667 perf_event_enable(per_cpu(watchdog_ev, cpu));
672 static void watchdog_nmi_disable(unsigned int cpu)
674 struct perf_event *event = per_cpu(watchdog_ev, cpu);
677 perf_event_disable(event);
678 per_cpu(watchdog_ev, cpu) = NULL;
680 /* should be in cleanup, but blocks oprofile */
681 perf_event_release_kernel(event);
684 /* watchdog_nmi_enable() expects this to be zero initially. */
690 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
691 static void watchdog_nmi_disable(unsigned int cpu) { return; }
692 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
694 static struct smp_hotplug_thread watchdog_threads = {
695 .store = &softlockup_watchdog,
696 .thread_should_run = watchdog_should_run,
697 .thread_fn = watchdog,
698 .thread_comm = "watchdog/%u",
699 .setup = watchdog_enable,
700 .cleanup = watchdog_cleanup,
701 .park = watchdog_disable,
702 .unpark = watchdog_enable,
706 * park all watchdog threads that are specified in 'watchdog_cpumask'
708 * This function returns an error if kthread_park() of a watchdog thread
709 * fails. In this situation, the watchdog threads of some CPUs can already
710 * be parked and the watchdog threads of other CPUs can still be runnable.
711 * Callers are expected to handle this special condition as appropriate in
714 * This function may only be called in a context that is protected against
715 * races with CPU hotplug - for example, via get_online_cpus().
717 static int watchdog_park_threads(void)
721 for_each_watchdog_cpu(cpu) {
722 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
731 * unpark all watchdog threads that are specified in 'watchdog_cpumask'
733 * This function may only be called in a context that is protected against
734 * races with CPU hotplug - for example, via get_online_cpus().
736 static void watchdog_unpark_threads(void)
740 for_each_watchdog_cpu(cpu)
741 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
745 * Suspend the hard and soft lockup detector by parking the watchdog threads.
747 int lockup_detector_suspend(void)
752 mutex_lock(&watchdog_proc_mutex);
754 * Multiple suspend requests can be active in parallel (counted by
755 * the 'watchdog_suspended' variable). If the watchdog threads are
756 * running, the first caller takes care that they will be parked.
757 * The state of 'watchdog_running' cannot change while a suspend
758 * request is active (see related code in 'proc' handlers).
760 if (watchdog_running && !watchdog_suspended)
761 ret = watchdog_park_threads();
764 watchdog_suspended++;
766 watchdog_disable_all_cpus();
767 pr_err("Failed to suspend lockup detectors, disabled\n");
768 watchdog_enabled = 0;
771 mutex_unlock(&watchdog_proc_mutex);
777 * Resume the hard and soft lockup detector by unparking the watchdog threads.
779 void lockup_detector_resume(void)
781 mutex_lock(&watchdog_proc_mutex);
783 watchdog_suspended--;
785 * The watchdog threads are unparked if they were previously running
786 * and if there is no more active suspend request.
788 if (watchdog_running && !watchdog_suspended)
789 watchdog_unpark_threads();
791 mutex_unlock(&watchdog_proc_mutex);
795 static int update_watchdog_all_cpus(void)
799 ret = watchdog_park_threads();
803 watchdog_unpark_threads();
808 static int watchdog_enable_all_cpus(void)
812 if (!watchdog_running) {
813 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
816 pr_err("Failed to create watchdog threads, disabled\n");
818 watchdog_running = 1;
821 * Enable/disable the lockup detectors or
822 * change the sample period 'on the fly'.
824 err = update_watchdog_all_cpus();
827 watchdog_disable_all_cpus();
828 pr_err("Failed to update lockup detectors, disabled\n");
833 watchdog_enabled = 0;
838 static void watchdog_disable_all_cpus(void)
840 if (watchdog_running) {
841 watchdog_running = 0;
842 smpboot_unregister_percpu_thread(&watchdog_threads);
849 * Update the run state of the lockup detectors.
851 static int proc_watchdog_update(void)
856 * Watchdog threads won't be started if they are already active.
857 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
858 * care of this. If those threads are already active, the sample
859 * period will be updated and the lockup detectors will be enabled
860 * or disabled 'on the fly'.
862 if (watchdog_enabled && watchdog_thresh)
863 err = watchdog_enable_all_cpus();
865 watchdog_disable_all_cpus();
872 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
874 * caller | table->data points to | 'which' contains the flag(s)
875 * -------------------|-----------------------|-----------------------------
876 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
877 * | | with SOFT_WATCHDOG_ENABLED
878 * -------------------|-----------------------|-----------------------------
879 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
880 * -------------------|-----------------------|-----------------------------
881 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
883 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
884 void __user *buffer, size_t *lenp, loff_t *ppos)
887 int *watchdog_param = (int *)table->data;
890 mutex_lock(&watchdog_proc_mutex);
892 if (watchdog_suspended) {
893 /* no parameter changes allowed while watchdog is suspended */
899 * If the parameter is being read return the state of the corresponding
900 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
901 * run state of the lockup detectors.
904 *watchdog_param = (watchdog_enabled & which) != 0;
905 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
907 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
912 * There is a race window between fetching the current value
913 * from 'watchdog_enabled' and storing the new value. During
914 * this race window, watchdog_nmi_enable() can sneak in and
915 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
916 * The 'cmpxchg' detects this race and the loop retries.
919 old = watchdog_enabled;
921 * If the parameter value is not zero set the
922 * corresponding bit(s), else clear it(them).
928 } while (cmpxchg(&watchdog_enabled, old, new) != old);
931 * Update the run state of the lockup detectors. There is _no_
932 * need to check the value returned by proc_watchdog_update()
933 * and to restore the previous value of 'watchdog_enabled' as
934 * both lockup detectors are disabled if proc_watchdog_update()
940 err = proc_watchdog_update();
943 mutex_unlock(&watchdog_proc_mutex);
949 * /proc/sys/kernel/watchdog
951 int proc_watchdog(struct ctl_table *table, int write,
952 void __user *buffer, size_t *lenp, loff_t *ppos)
954 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
955 table, write, buffer, lenp, ppos);
959 * /proc/sys/kernel/nmi_watchdog
961 int proc_nmi_watchdog(struct ctl_table *table, int write,
962 void __user *buffer, size_t *lenp, loff_t *ppos)
964 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
965 table, write, buffer, lenp, ppos);
969 * /proc/sys/kernel/soft_watchdog
971 int proc_soft_watchdog(struct ctl_table *table, int write,
972 void __user *buffer, size_t *lenp, loff_t *ppos)
974 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
975 table, write, buffer, lenp, ppos);
979 * /proc/sys/kernel/watchdog_thresh
981 int proc_watchdog_thresh(struct ctl_table *table, int write,
982 void __user *buffer, size_t *lenp, loff_t *ppos)
987 mutex_lock(&watchdog_proc_mutex);
989 if (watchdog_suspended) {
990 /* no parameter changes allowed while watchdog is suspended */
995 old = ACCESS_ONCE(watchdog_thresh);
996 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1002 * Update the sample period. Restore on failure.
1004 new = ACCESS_ONCE(watchdog_thresh);
1008 set_sample_period();
1009 err = proc_watchdog_update();
1011 watchdog_thresh = old;
1012 set_sample_period();
1015 mutex_unlock(&watchdog_proc_mutex);
1021 * The cpumask is the mask of possible cpus that the watchdog can run
1022 * on, not the mask of cpus it is actually running on. This allows the
1023 * user to specify a mask that will include cpus that have not yet
1024 * been brought online, if desired.
1026 int proc_watchdog_cpumask(struct ctl_table *table, int write,
1027 void __user *buffer, size_t *lenp, loff_t *ppos)
1032 mutex_lock(&watchdog_proc_mutex);
1034 if (watchdog_suspended) {
1035 /* no parameter changes allowed while watchdog is suspended */
1040 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1041 if (!err && write) {
1042 /* Remove impossible cpus to keep sysctl output cleaner. */
1043 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1046 if (watchdog_running) {
1048 * Failure would be due to being unable to allocate
1049 * a temporary cpumask, so we are likely not in a
1050 * position to do much else to make things better.
1052 if (smpboot_update_cpumask_percpu_thread(
1053 &watchdog_threads, &watchdog_cpumask) != 0)
1054 pr_err("cpumask update failed\n");
1058 mutex_unlock(&watchdog_proc_mutex);
1063 #endif /* CONFIG_SYSCTL */
1065 void __init lockup_detector_init(void)
1067 set_sample_period();
1069 #ifdef CONFIG_NO_HZ_FULL
1070 if (tick_nohz_full_enabled()) {
1071 pr_info("Disabling watchdog on nohz_full cores by default\n");
1072 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1074 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1076 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1079 if (watchdog_enabled)
1080 watchdog_enable_all_cpus();