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>
30 static DEFINE_MUTEX(watchdog_proc_mutex);
32 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
33 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
35 unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
37 int __read_mostly nmi_watchdog_enabled;
38 int __read_mostly soft_watchdog_enabled;
39 int __read_mostly watchdog_user_enabled;
40 int __read_mostly watchdog_thresh = 10;
43 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
44 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
46 static struct cpumask watchdog_cpumask __read_mostly;
47 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
49 /* Helper for online, unparked cpus. */
50 #define for_each_watchdog_cpu(cpu) \
51 for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
54 * The 'watchdog_running' variable is set to 1 when the watchdog threads
55 * are registered/started and is set to 0 when the watchdog threads are
56 * unregistered/stopped, so it is an indicator whether the threads exist.
58 static int __read_mostly watchdog_running;
60 * If a subsystem has a need to deactivate the watchdog temporarily, it
61 * can use the suspend/resume interface to achieve this. The content of
62 * the 'watchdog_suspended' variable reflects this state. Existing threads
63 * are parked/unparked by the lockup_detector_{suspend|resume} functions
64 * (see comment blocks pertaining to those functions for further details).
66 * 'watchdog_suspended' also prevents threads from being registered/started
67 * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
68 * of 'watchdog_running' cannot change while the watchdog is deactivated
69 * temporarily (see related code in 'proc' handlers).
71 static int __read_mostly watchdog_suspended;
73 static u64 __read_mostly sample_period;
75 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
76 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
77 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
78 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
79 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
80 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
81 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
82 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
83 #ifdef CONFIG_HARDLOCKUP_DETECTOR
84 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
85 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
86 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
87 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
89 static unsigned long soft_lockup_nmi_warn;
93 * Should we panic when a soft-lockup or hard-lockup occurs:
95 #ifdef CONFIG_HARDLOCKUP_DETECTOR
96 unsigned int __read_mostly hardlockup_panic =
97 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
98 static unsigned long hardlockup_allcpu_dumped;
100 * We may not want to enable hard lockup detection by default in all cases,
101 * for example when running the kernel as a guest on a hypervisor. In these
102 * cases this function can be called to disable hard lockup detection. This
103 * function should only be executed once by the boot processor before the
104 * kernel command line parameters are parsed, because otherwise it is not
105 * possible to override this in hardlockup_panic_setup().
107 void hardlockup_detector_disable(void)
109 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
112 static int __init hardlockup_panic_setup(char *str)
114 if (!strncmp(str, "panic", 5))
115 hardlockup_panic = 1;
116 else if (!strncmp(str, "nopanic", 7))
117 hardlockup_panic = 0;
118 else if (!strncmp(str, "0", 1))
119 watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
120 else if (!strncmp(str, "1", 1))
121 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
124 __setup("nmi_watchdog=", hardlockup_panic_setup);
127 unsigned int __read_mostly softlockup_panic =
128 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
130 static int __init softlockup_panic_setup(char *str)
132 softlockup_panic = simple_strtoul(str, NULL, 0);
136 __setup("softlockup_panic=", softlockup_panic_setup);
138 static int __init nowatchdog_setup(char *str)
140 watchdog_enabled = 0;
143 __setup("nowatchdog", nowatchdog_setup);
145 static int __init nosoftlockup_setup(char *str)
147 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
150 __setup("nosoftlockup", nosoftlockup_setup);
153 static int __init softlockup_all_cpu_backtrace_setup(char *str)
155 sysctl_softlockup_all_cpu_backtrace =
156 !!simple_strtol(str, NULL, 0);
159 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
160 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
162 sysctl_hardlockup_all_cpu_backtrace =
163 !!simple_strtol(str, NULL, 0);
166 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
170 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
171 * lockups can have false positives under extreme conditions. So we generally
172 * want a higher threshold for soft lockups than for hard lockups. So we couple
173 * the thresholds with a factor: we make the soft threshold twice the amount of
174 * time the hard threshold is.
176 static int get_softlockup_thresh(void)
178 return watchdog_thresh * 2;
182 * Returns seconds, approximately. We don't need nanosecond
183 * resolution, and we don't need to waste time with a big divide when
186 static unsigned long get_timestamp(void)
188 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
191 static void set_sample_period(void)
194 * convert watchdog_thresh from seconds to ns
195 * the divide by 5 is to give hrtimer several chances (two
196 * or three with the current relation between the soft
197 * and hard thresholds) to increment before the
198 * hardlockup detector generates a warning
200 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
203 /* Commands for resetting the watchdog */
204 static void __touch_watchdog(void)
206 __this_cpu_write(watchdog_touch_ts, get_timestamp());
210 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
212 * Call when the scheduler may have stalled for legitimate reasons
213 * preventing the watchdog task from executing - e.g. the scheduler
214 * entering idle state. This should only be used for scheduler events.
215 * Use touch_softlockup_watchdog() for everything else.
217 void touch_softlockup_watchdog_sched(void)
220 * Preemption can be enabled. It doesn't matter which CPU's timestamp
221 * gets zeroed here, so use the raw_ operation.
223 raw_cpu_write(watchdog_touch_ts, 0);
226 void touch_softlockup_watchdog(void)
228 touch_softlockup_watchdog_sched();
229 wq_watchdog_touch(raw_smp_processor_id());
231 EXPORT_SYMBOL(touch_softlockup_watchdog);
233 void touch_all_softlockup_watchdogs(void)
238 * this is done lockless
239 * do we care if a 0 races with a timestamp?
240 * all it means is the softlock check starts one cycle later
242 for_each_watchdog_cpu(cpu)
243 per_cpu(watchdog_touch_ts, cpu) = 0;
244 wq_watchdog_touch(-1);
247 #ifdef CONFIG_HARDLOCKUP_DETECTOR
248 void touch_nmi_watchdog(void)
251 * Using __raw here because some code paths have
252 * preemption enabled. If preemption is enabled
253 * then interrupts should be enabled too, in which
254 * case we shouldn't have to worry about the watchdog
257 raw_cpu_write(watchdog_nmi_touch, true);
258 touch_softlockup_watchdog();
260 EXPORT_SYMBOL(touch_nmi_watchdog);
264 void touch_softlockup_watchdog_sync(void)
266 __this_cpu_write(softlockup_touch_sync, true);
267 __this_cpu_write(watchdog_touch_ts, 0);
270 #ifdef CONFIG_HARDLOCKUP_DETECTOR
271 /* watchdog detector functions */
272 bool is_hardlockup(void)
274 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
276 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
279 __this_cpu_write(hrtimer_interrupts_saved, hrint);
284 static int is_softlockup(unsigned long touch_ts)
286 unsigned long now = get_timestamp();
288 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
289 /* Warn about unreasonable delays. */
290 if (time_after(now, touch_ts + get_softlockup_thresh()))
291 return now - touch_ts;
296 #ifdef CONFIG_HARDLOCKUP_DETECTOR
298 static struct perf_event_attr wd_hw_attr = {
299 .type = PERF_TYPE_HARDWARE,
300 .config = PERF_COUNT_HW_CPU_CYCLES,
301 .size = sizeof(struct perf_event_attr),
306 /* Callback function for perf event subsystem */
307 static void watchdog_overflow_callback(struct perf_event *event,
308 struct perf_sample_data *data,
309 struct pt_regs *regs)
311 /* Ensure the watchdog never gets throttled */
312 event->hw.interrupts = 0;
314 if (__this_cpu_read(watchdog_nmi_touch) == true) {
315 __this_cpu_write(watchdog_nmi_touch, false);
319 /* check for a hardlockup
320 * This is done by making sure our timer interrupt
321 * is incrementing. The timer interrupt should have
322 * fired multiple times before we overflow'd. If it hasn't
323 * then this is a good indication the cpu is stuck
325 if (is_hardlockup()) {
326 int this_cpu = smp_processor_id();
328 /* only print hardlockups once */
329 if (__this_cpu_read(hard_watchdog_warn) == true)
332 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
334 print_irqtrace_events(current);
341 * Perform all-CPU dump only once to avoid multiple hardlockups
342 * generating interleaving traces
344 if (sysctl_hardlockup_all_cpu_backtrace &&
345 !test_and_set_bit(0, &hardlockup_allcpu_dumped))
346 trigger_allbutself_cpu_backtrace();
348 if (hardlockup_panic)
349 nmi_panic(regs, "Hard LOCKUP");
351 __this_cpu_write(hard_watchdog_warn, true);
355 __this_cpu_write(hard_watchdog_warn, false);
358 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
360 static void watchdog_interrupt_count(void)
362 __this_cpu_inc(hrtimer_interrupts);
365 static int watchdog_nmi_enable(unsigned int cpu);
366 static void watchdog_nmi_disable(unsigned int cpu);
368 static int watchdog_enable_all_cpus(void);
369 static void watchdog_disable_all_cpus(void);
371 /* watchdog kicker functions */
372 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
374 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
375 struct pt_regs *regs = get_irq_regs();
377 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
379 /* kick the hardlockup detector */
380 watchdog_interrupt_count();
382 /* kick the softlockup detector */
383 wake_up_process(__this_cpu_read(softlockup_watchdog));
386 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
389 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
391 * If the time stamp was touched atomically
392 * make sure the scheduler tick is up to date.
394 __this_cpu_write(softlockup_touch_sync, false);
398 /* Clear the guest paused flag on watchdog reset */
399 kvm_check_and_clear_guest_paused();
401 return HRTIMER_RESTART;
404 /* check for a softlockup
405 * This is done by making sure a high priority task is
406 * being scheduled. The task touches the watchdog to
407 * indicate it is getting cpu time. If it hasn't then
408 * this is a good indication some task is hogging the cpu
410 duration = is_softlockup(touch_ts);
411 if (unlikely(duration)) {
413 * If a virtual machine is stopped by the host it can look to
414 * the watchdog like a soft lockup, check to see if the host
415 * stopped the vm before we issue the warning
417 if (kvm_check_and_clear_guest_paused())
418 return HRTIMER_RESTART;
421 if (__this_cpu_read(soft_watchdog_warn) == true) {
423 * When multiple processes are causing softlockups the
424 * softlockup detector only warns on the first one
425 * because the code relies on a full quiet cycle to
426 * re-arm. The second process prevents the quiet cycle
427 * and never gets reported. Use task pointers to detect
430 if (__this_cpu_read(softlockup_task_ptr_saved) !=
432 __this_cpu_write(soft_watchdog_warn, false);
435 return HRTIMER_RESTART;
438 if (softlockup_all_cpu_backtrace) {
439 /* Prevent multiple soft-lockup reports if one cpu is already
440 * engaged in dumping cpu back traces
442 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
443 /* Someone else will report us. Let's give up */
444 __this_cpu_write(soft_watchdog_warn, true);
445 return HRTIMER_RESTART;
449 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
450 smp_processor_id(), duration,
451 current->comm, task_pid_nr(current));
452 __this_cpu_write(softlockup_task_ptr_saved, current);
454 print_irqtrace_events(current);
460 if (softlockup_all_cpu_backtrace) {
461 /* Avoid generating two back traces for current
462 * given that one is already made above
464 trigger_allbutself_cpu_backtrace();
466 clear_bit(0, &soft_lockup_nmi_warn);
467 /* Barrier to sync with other cpus */
468 smp_mb__after_atomic();
471 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
472 if (softlockup_panic)
473 panic("softlockup: hung tasks");
474 __this_cpu_write(soft_watchdog_warn, true);
476 __this_cpu_write(soft_watchdog_warn, false);
478 return HRTIMER_RESTART;
481 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
483 struct sched_param param = { .sched_priority = prio };
485 sched_setscheduler(current, policy, ¶m);
488 static void watchdog_enable(unsigned int cpu)
490 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
492 /* kick off the timer for the hardlockup detector */
493 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
494 hrtimer->function = watchdog_timer_fn;
496 /* Enable the perf event */
497 watchdog_nmi_enable(cpu);
499 /* done here because hrtimer_start can only pin to smp_processor_id() */
500 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
501 HRTIMER_MODE_REL_PINNED);
503 /* initialize timestamp */
504 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
508 static void watchdog_disable(unsigned int cpu)
510 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
512 watchdog_set_prio(SCHED_NORMAL, 0);
513 hrtimer_cancel(hrtimer);
514 /* disable the perf event */
515 watchdog_nmi_disable(cpu);
518 static void watchdog_cleanup(unsigned int cpu, bool online)
520 watchdog_disable(cpu);
523 static int watchdog_should_run(unsigned int cpu)
525 return __this_cpu_read(hrtimer_interrupts) !=
526 __this_cpu_read(soft_lockup_hrtimer_cnt);
530 * The watchdog thread function - touches the timestamp.
532 * It only runs once every sample_period seconds (4 seconds by
533 * default) to reset the softlockup timestamp. If this gets delayed
534 * for more than 2*watchdog_thresh seconds then the debug-printout
535 * triggers in watchdog_timer_fn().
537 static void watchdog(unsigned int cpu)
539 __this_cpu_write(soft_lockup_hrtimer_cnt,
540 __this_cpu_read(hrtimer_interrupts));
544 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
545 * failure path. Check for failures that can occur asynchronously -
546 * for example, when CPUs are on-lined - and shut down the hardware
547 * perf event on each CPU accordingly.
549 * The only non-obvious place this bit can be cleared is through
550 * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a
551 * pr_info here would be too noisy as it would result in a message
552 * every few seconds if the hardlockup was disabled but the softlockup
555 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
556 watchdog_nmi_disable(cpu);
559 #ifdef CONFIG_HARDLOCKUP_DETECTOR
561 * People like the simple clean cpu node info on boot.
562 * Reduce the watchdog noise by only printing messages
563 * that are different from what cpu0 displayed.
565 static unsigned long cpu0_err;
567 static int watchdog_nmi_enable(unsigned int cpu)
569 struct perf_event_attr *wd_attr;
570 struct perf_event *event = per_cpu(watchdog_ev, cpu);
572 /* nothing to do if the hard lockup detector is disabled */
573 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
576 /* is it already setup and enabled? */
577 if (event && event->state > PERF_EVENT_STATE_OFF)
580 /* it is setup but not enabled */
584 wd_attr = &wd_hw_attr;
585 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
587 /* Try to register using hardware perf events */
588 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
590 /* save cpu0 error for future comparision */
591 if (cpu == 0 && IS_ERR(event))
592 cpu0_err = PTR_ERR(event);
594 if (!IS_ERR(event)) {
595 /* only print for cpu0 or different than cpu0 */
596 if (cpu == 0 || cpu0_err)
597 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
602 * Disable the hard lockup detector if _any_ CPU fails to set up
603 * set up the hardware perf event. The watchdog() function checks
604 * the NMI_WATCHDOG_ENABLED bit periodically.
606 * The barriers are for syncing up watchdog_enabled across all the
607 * cpus, as clear_bit() does not use barriers.
609 smp_mb__before_atomic();
610 clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
611 smp_mb__after_atomic();
613 /* skip displaying the same error again */
614 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
615 return PTR_ERR(event);
617 /* vary the KERN level based on the returned errno */
618 if (PTR_ERR(event) == -EOPNOTSUPP)
619 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
620 else if (PTR_ERR(event) == -ENOENT)
621 pr_warn("disabled (cpu%i): hardware events not enabled\n",
624 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
625 cpu, PTR_ERR(event));
627 pr_info("Shutting down hard lockup detector on all cpus\n");
629 return PTR_ERR(event);
633 per_cpu(watchdog_ev, cpu) = event;
635 perf_event_enable(per_cpu(watchdog_ev, cpu));
640 static void watchdog_nmi_disable(unsigned int cpu)
642 struct perf_event *event = per_cpu(watchdog_ev, cpu);
645 perf_event_disable(event);
646 per_cpu(watchdog_ev, cpu) = NULL;
648 /* should be in cleanup, but blocks oprofile */
649 perf_event_release_kernel(event);
652 /* watchdog_nmi_enable() expects this to be zero initially. */
658 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
659 static void watchdog_nmi_disable(unsigned int cpu) { return; }
660 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
662 static struct smp_hotplug_thread watchdog_threads = {
663 .store = &softlockup_watchdog,
664 .thread_should_run = watchdog_should_run,
665 .thread_fn = watchdog,
666 .thread_comm = "watchdog/%u",
667 .setup = watchdog_enable,
668 .cleanup = watchdog_cleanup,
669 .park = watchdog_disable,
670 .unpark = watchdog_enable,
674 * park all watchdog threads that are specified in 'watchdog_cpumask'
676 * This function returns an error if kthread_park() of a watchdog thread
677 * fails. In this situation, the watchdog threads of some CPUs can already
678 * be parked and the watchdog threads of other CPUs can still be runnable.
679 * Callers are expected to handle this special condition as appropriate in
682 * This function may only be called in a context that is protected against
683 * races with CPU hotplug - for example, via get_online_cpus().
685 static int watchdog_park_threads(void)
689 for_each_watchdog_cpu(cpu) {
690 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
699 * unpark all watchdog threads that are specified in 'watchdog_cpumask'
701 * This function may only be called in a context that is protected against
702 * races with CPU hotplug - for example, via get_online_cpus().
704 static void watchdog_unpark_threads(void)
708 for_each_watchdog_cpu(cpu)
709 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
713 * Suspend the hard and soft lockup detector by parking the watchdog threads.
715 int lockup_detector_suspend(void)
720 mutex_lock(&watchdog_proc_mutex);
722 * Multiple suspend requests can be active in parallel (counted by
723 * the 'watchdog_suspended' variable). If the watchdog threads are
724 * running, the first caller takes care that they will be parked.
725 * The state of 'watchdog_running' cannot change while a suspend
726 * request is active (see related code in 'proc' handlers).
728 if (watchdog_running && !watchdog_suspended)
729 ret = watchdog_park_threads();
732 watchdog_suspended++;
734 watchdog_disable_all_cpus();
735 pr_err("Failed to suspend lockup detectors, disabled\n");
736 watchdog_enabled = 0;
739 mutex_unlock(&watchdog_proc_mutex);
745 * Resume the hard and soft lockup detector by unparking the watchdog threads.
747 void lockup_detector_resume(void)
749 mutex_lock(&watchdog_proc_mutex);
751 watchdog_suspended--;
753 * The watchdog threads are unparked if they were previously running
754 * and if there is no more active suspend request.
756 if (watchdog_running && !watchdog_suspended)
757 watchdog_unpark_threads();
759 mutex_unlock(&watchdog_proc_mutex);
763 static int update_watchdog_all_cpus(void)
767 ret = watchdog_park_threads();
771 watchdog_unpark_threads();
776 static int watchdog_enable_all_cpus(void)
780 if (!watchdog_running) {
781 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
784 pr_err("Failed to create watchdog threads, disabled\n");
786 watchdog_running = 1;
789 * Enable/disable the lockup detectors or
790 * change the sample period 'on the fly'.
792 err = update_watchdog_all_cpus();
795 watchdog_disable_all_cpus();
796 pr_err("Failed to update lockup detectors, disabled\n");
801 watchdog_enabled = 0;
806 static void watchdog_disable_all_cpus(void)
808 if (watchdog_running) {
809 watchdog_running = 0;
810 smpboot_unregister_percpu_thread(&watchdog_threads);
817 * Update the run state of the lockup detectors.
819 static int proc_watchdog_update(void)
824 * Watchdog threads won't be started if they are already active.
825 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
826 * care of this. If those threads are already active, the sample
827 * period will be updated and the lockup detectors will be enabled
828 * or disabled 'on the fly'.
830 if (watchdog_enabled && watchdog_thresh)
831 err = watchdog_enable_all_cpus();
833 watchdog_disable_all_cpus();
840 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
842 * caller | table->data points to | 'which' contains the flag(s)
843 * -------------------|-----------------------|-----------------------------
844 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
845 * | | with SOFT_WATCHDOG_ENABLED
846 * -------------------|-----------------------|-----------------------------
847 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
848 * -------------------|-----------------------|-----------------------------
849 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
851 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
852 void __user *buffer, size_t *lenp, loff_t *ppos)
855 int *watchdog_param = (int *)table->data;
858 mutex_lock(&watchdog_proc_mutex);
860 if (watchdog_suspended) {
861 /* no parameter changes allowed while watchdog is suspended */
867 * If the parameter is being read return the state of the corresponding
868 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
869 * run state of the lockup detectors.
872 *watchdog_param = (watchdog_enabled & which) != 0;
873 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
875 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
880 * There is a race window between fetching the current value
881 * from 'watchdog_enabled' and storing the new value. During
882 * this race window, watchdog_nmi_enable() can sneak in and
883 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
884 * The 'cmpxchg' detects this race and the loop retries.
887 old = watchdog_enabled;
889 * If the parameter value is not zero set the
890 * corresponding bit(s), else clear it(them).
896 } while (cmpxchg(&watchdog_enabled, old, new) != old);
899 * Update the run state of the lockup detectors. There is _no_
900 * need to check the value returned by proc_watchdog_update()
901 * and to restore the previous value of 'watchdog_enabled' as
902 * both lockup detectors are disabled if proc_watchdog_update()
908 err = proc_watchdog_update();
911 mutex_unlock(&watchdog_proc_mutex);
917 * /proc/sys/kernel/watchdog
919 int proc_watchdog(struct ctl_table *table, int write,
920 void __user *buffer, size_t *lenp, loff_t *ppos)
922 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
923 table, write, buffer, lenp, ppos);
927 * /proc/sys/kernel/nmi_watchdog
929 int proc_nmi_watchdog(struct ctl_table *table, int write,
930 void __user *buffer, size_t *lenp, loff_t *ppos)
932 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
933 table, write, buffer, lenp, ppos);
937 * /proc/sys/kernel/soft_watchdog
939 int proc_soft_watchdog(struct ctl_table *table, int write,
940 void __user *buffer, size_t *lenp, loff_t *ppos)
942 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
943 table, write, buffer, lenp, ppos);
947 * /proc/sys/kernel/watchdog_thresh
949 int proc_watchdog_thresh(struct ctl_table *table, int write,
950 void __user *buffer, size_t *lenp, loff_t *ppos)
955 mutex_lock(&watchdog_proc_mutex);
957 if (watchdog_suspended) {
958 /* no parameter changes allowed while watchdog is suspended */
963 old = ACCESS_ONCE(watchdog_thresh);
964 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
970 * Update the sample period. Restore on failure.
972 new = ACCESS_ONCE(watchdog_thresh);
977 err = proc_watchdog_update();
979 watchdog_thresh = old;
983 mutex_unlock(&watchdog_proc_mutex);
989 * The cpumask is the mask of possible cpus that the watchdog can run
990 * on, not the mask of cpus it is actually running on. This allows the
991 * user to specify a mask that will include cpus that have not yet
992 * been brought online, if desired.
994 int proc_watchdog_cpumask(struct ctl_table *table, int write,
995 void __user *buffer, size_t *lenp, loff_t *ppos)
1000 mutex_lock(&watchdog_proc_mutex);
1002 if (watchdog_suspended) {
1003 /* no parameter changes allowed while watchdog is suspended */
1008 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1009 if (!err && write) {
1010 /* Remove impossible cpus to keep sysctl output cleaner. */
1011 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
1014 if (watchdog_running) {
1016 * Failure would be due to being unable to allocate
1017 * a temporary cpumask, so we are likely not in a
1018 * position to do much else to make things better.
1020 if (smpboot_update_cpumask_percpu_thread(
1021 &watchdog_threads, &watchdog_cpumask) != 0)
1022 pr_err("cpumask update failed\n");
1026 mutex_unlock(&watchdog_proc_mutex);
1031 #endif /* CONFIG_SYSCTL */
1033 void __init lockup_detector_init(void)
1035 set_sample_period();
1037 #ifdef CONFIG_NO_HZ_FULL
1038 if (tick_nohz_full_enabled()) {
1039 pr_info("Disabling watchdog on nohz_full cores by default\n");
1040 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
1042 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1044 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
1047 if (watchdog_enabled)
1048 watchdog_enable_all_cpus();