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>
23 #include <asm/irq_regs.h>
24 #include <linux/kvm_para.h>
25 #include <linux/perf_event.h>
27 int watchdog_user_enabled = 1;
28 int __read_mostly watchdog_thresh = 10;
30 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
32 #define sysctl_softlockup_all_cpu_backtrace 0
35 static int __read_mostly watchdog_running;
36 static u64 __read_mostly sample_period;
38 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
39 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
40 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
41 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
42 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
43 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
44 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
45 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
46 #ifdef CONFIG_HARDLOCKUP_DETECTOR
47 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
48 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
49 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
50 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
52 static unsigned long soft_lockup_nmi_warn;
56 * Should we panic when a soft-lockup or hard-lockup occurs:
58 #ifdef CONFIG_HARDLOCKUP_DETECTOR
59 static int hardlockup_panic =
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
62 static bool hardlockup_detector_enabled = true;
64 * We may not want to enable hard lockup detection by default in all cases,
65 * for example when running the kernel as a guest on a hypervisor. In these
66 * cases this function can be called to disable hard lockup detection. This
67 * function should only be executed once by the boot processor before the
68 * kernel command line parameters are parsed, because otherwise it is not
69 * possible to override this in hardlockup_panic_setup().
71 void watchdog_enable_hardlockup_detector(bool val)
73 hardlockup_detector_enabled = val;
76 bool watchdog_hardlockup_detector_is_enabled(void)
78 return hardlockup_detector_enabled;
81 static int __init hardlockup_panic_setup(char *str)
83 if (!strncmp(str, "panic", 5))
85 else if (!strncmp(str, "nopanic", 7))
87 else if (!strncmp(str, "0", 1))
88 watchdog_user_enabled = 0;
89 else if (!strncmp(str, "1", 1) || !strncmp(str, "2", 1)) {
91 * Setting 'nmi_watchdog=1' or 'nmi_watchdog=2' (legacy option)
92 * has the same effect.
94 watchdog_user_enabled = 1;
95 watchdog_enable_hardlockup_detector(true);
99 __setup("nmi_watchdog=", hardlockup_panic_setup);
102 unsigned int __read_mostly softlockup_panic =
103 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
105 static int __init softlockup_panic_setup(char *str)
107 softlockup_panic = simple_strtoul(str, NULL, 0);
111 __setup("softlockup_panic=", softlockup_panic_setup);
113 static int __init nowatchdog_setup(char *str)
115 watchdog_user_enabled = 0;
118 __setup("nowatchdog", nowatchdog_setup);
121 static int __init nosoftlockup_setup(char *str)
123 watchdog_user_enabled = 0;
126 __setup("nosoftlockup", nosoftlockup_setup);
129 static int __init softlockup_all_cpu_backtrace_setup(char *str)
131 sysctl_softlockup_all_cpu_backtrace =
132 !!simple_strtol(str, NULL, 0);
135 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
139 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
140 * lockups can have false positives under extreme conditions. So we generally
141 * want a higher threshold for soft lockups than for hard lockups. So we couple
142 * the thresholds with a factor: we make the soft threshold twice the amount of
143 * time the hard threshold is.
145 static int get_softlockup_thresh(void)
147 return watchdog_thresh * 2;
151 * Returns seconds, approximately. We don't need nanosecond
152 * resolution, and we don't need to waste time with a big divide when
155 static unsigned long get_timestamp(void)
157 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
160 static void set_sample_period(void)
163 * convert watchdog_thresh from seconds to ns
164 * the divide by 5 is to give hrtimer several chances (two
165 * or three with the current relation between the soft
166 * and hard thresholds) to increment before the
167 * hardlockup detector generates a warning
169 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
172 /* Commands for resetting the watchdog */
173 static void __touch_watchdog(void)
175 __this_cpu_write(watchdog_touch_ts, get_timestamp());
178 void touch_softlockup_watchdog(void)
181 * Preemption can be enabled. It doesn't matter which CPU's timestamp
182 * gets zeroed here, so use the raw_ operation.
184 raw_cpu_write(watchdog_touch_ts, 0);
186 EXPORT_SYMBOL(touch_softlockup_watchdog);
188 void touch_all_softlockup_watchdogs(void)
193 * this is done lockless
194 * do we care if a 0 races with a timestamp?
195 * all it means is the softlock check starts one cycle later
197 for_each_online_cpu(cpu)
198 per_cpu(watchdog_touch_ts, cpu) = 0;
201 #ifdef CONFIG_HARDLOCKUP_DETECTOR
202 void touch_nmi_watchdog(void)
205 * Using __raw here because some code paths have
206 * preemption enabled. If preemption is enabled
207 * then interrupts should be enabled too, in which
208 * case we shouldn't have to worry about the watchdog
211 raw_cpu_write(watchdog_nmi_touch, true);
212 touch_softlockup_watchdog();
214 EXPORT_SYMBOL(touch_nmi_watchdog);
218 void touch_softlockup_watchdog_sync(void)
220 __this_cpu_write(softlockup_touch_sync, true);
221 __this_cpu_write(watchdog_touch_ts, 0);
224 #ifdef CONFIG_HARDLOCKUP_DETECTOR
225 /* watchdog detector functions */
226 static int is_hardlockup(void)
228 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
230 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
233 __this_cpu_write(hrtimer_interrupts_saved, hrint);
238 static int is_softlockup(unsigned long touch_ts)
240 unsigned long now = get_timestamp();
242 /* Warn about unreasonable delays: */
243 if (time_after(now, touch_ts + get_softlockup_thresh()))
244 return now - touch_ts;
249 #ifdef CONFIG_HARDLOCKUP_DETECTOR
251 static DEFINE_RAW_SPINLOCK(watchdog_output_lock);
253 static struct perf_event_attr wd_hw_attr = {
254 .type = PERF_TYPE_HARDWARE,
255 .config = PERF_COUNT_HW_CPU_CYCLES,
256 .size = sizeof(struct perf_event_attr),
261 /* Callback function for perf event subsystem */
262 static void watchdog_overflow_callback(struct perf_event *event,
263 struct perf_sample_data *data,
264 struct pt_regs *regs)
266 /* Ensure the watchdog never gets throttled */
267 event->hw.interrupts = 0;
269 if (__this_cpu_read(watchdog_nmi_touch) == true) {
270 __this_cpu_write(watchdog_nmi_touch, false);
274 /* check for a hardlockup
275 * This is done by making sure our timer interrupt
276 * is incrementing. The timer interrupt should have
277 * fired multiple times before we overflow'd. If it hasn't
278 * then this is a good indication the cpu is stuck
280 if (is_hardlockup()) {
281 int this_cpu = smp_processor_id();
283 /* only print hardlockups once */
284 if (__this_cpu_read(hard_watchdog_warn) == true)
287 * If early-printk is enabled then make sure we do not
288 * lock up in printk() and kill console logging:
292 if (hardlockup_panic) {
293 panic("Watchdog detected hard LOCKUP on cpu %d",
296 raw_spin_lock(&watchdog_output_lock);
297 WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
299 raw_spin_unlock(&watchdog_output_lock);
302 __this_cpu_write(hard_watchdog_warn, true);
306 __this_cpu_write(hard_watchdog_warn, false);
309 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
311 static void watchdog_interrupt_count(void)
313 __this_cpu_inc(hrtimer_interrupts);
316 static int watchdog_nmi_enable(unsigned int cpu);
317 static void watchdog_nmi_disable(unsigned int cpu);
319 /* watchdog kicker functions */
320 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
322 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
323 struct pt_regs *regs = get_irq_regs();
325 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
327 /* kick the hardlockup detector */
328 watchdog_interrupt_count();
330 /* kick the softlockup detector */
331 wake_up_process(__this_cpu_read(softlockup_watchdog));
334 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
337 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
339 * If the time stamp was touched atomically
340 * make sure the scheduler tick is up to date.
342 __this_cpu_write(softlockup_touch_sync, false);
346 /* Clear the guest paused flag on watchdog reset */
347 kvm_check_and_clear_guest_paused();
349 return HRTIMER_RESTART;
352 /* check for a softlockup
353 * This is done by making sure a high priority task is
354 * being scheduled. The task touches the watchdog to
355 * indicate it is getting cpu time. If it hasn't then
356 * this is a good indication some task is hogging the cpu
358 duration = is_softlockup(touch_ts);
359 if (unlikely(duration)) {
361 * If a virtual machine is stopped by the host it can look to
362 * the watchdog like a soft lockup, check to see if the host
363 * stopped the vm before we issue the warning
365 if (kvm_check_and_clear_guest_paused())
366 return HRTIMER_RESTART;
369 if (__this_cpu_read(soft_watchdog_warn) == true) {
371 * When multiple processes are causing softlockups the
372 * softlockup detector only warns on the first one
373 * because the code relies on a full quiet cycle to
374 * re-arm. The second process prevents the quiet cycle
375 * and never gets reported. Use task pointers to detect
378 if (__this_cpu_read(softlockup_task_ptr_saved) !=
380 __this_cpu_write(soft_watchdog_warn, false);
383 return HRTIMER_RESTART;
386 if (softlockup_all_cpu_backtrace) {
387 /* Prevent multiple soft-lockup reports if one cpu is already
388 * engaged in dumping cpu back traces
390 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
391 /* Someone else will report us. Let's give up */
392 __this_cpu_write(soft_watchdog_warn, true);
393 return HRTIMER_RESTART;
397 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
398 smp_processor_id(), duration,
399 current->comm, task_pid_nr(current));
400 __this_cpu_write(softlockup_task_ptr_saved, current);
402 print_irqtrace_events(current);
408 if (softlockup_all_cpu_backtrace) {
409 /* Avoid generating two back traces for current
410 * given that one is already made above
412 trigger_allbutself_cpu_backtrace();
414 clear_bit(0, &soft_lockup_nmi_warn);
415 /* Barrier to sync with other cpus */
416 smp_mb__after_atomic();
419 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
420 if (softlockup_panic)
421 panic("softlockup: hung tasks");
422 __this_cpu_write(soft_watchdog_warn, true);
424 __this_cpu_write(soft_watchdog_warn, false);
426 return HRTIMER_RESTART;
429 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
431 struct sched_param param = { .sched_priority = prio };
433 sched_setscheduler(current, policy, ¶m);
436 static void watchdog_enable(unsigned int cpu)
438 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
440 /* kick off the timer for the hardlockup detector */
441 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
442 hrtimer->function = watchdog_timer_fn;
443 hrtimer->irqsafe = 1;
445 /* Enable the perf event */
446 watchdog_nmi_enable(cpu);
448 /* done here because hrtimer_start can only pin to smp_processor_id() */
449 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
450 HRTIMER_MODE_REL_PINNED);
452 /* initialize timestamp */
453 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
457 static void watchdog_disable(unsigned int cpu)
459 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
461 watchdog_set_prio(SCHED_NORMAL, 0);
462 hrtimer_cancel(hrtimer);
463 /* disable the perf event */
464 watchdog_nmi_disable(cpu);
467 static void watchdog_cleanup(unsigned int cpu, bool online)
469 watchdog_disable(cpu);
472 static int watchdog_should_run(unsigned int cpu)
474 return __this_cpu_read(hrtimer_interrupts) !=
475 __this_cpu_read(soft_lockup_hrtimer_cnt);
479 * The watchdog thread function - touches the timestamp.
481 * It only runs once every sample_period seconds (4 seconds by
482 * default) to reset the softlockup timestamp. If this gets delayed
483 * for more than 2*watchdog_thresh seconds then the debug-printout
484 * triggers in watchdog_timer_fn().
486 static void watchdog(unsigned int cpu)
488 __this_cpu_write(soft_lockup_hrtimer_cnt,
489 __this_cpu_read(hrtimer_interrupts));
493 #ifdef CONFIG_HARDLOCKUP_DETECTOR
495 * People like the simple clean cpu node info on boot.
496 * Reduce the watchdog noise by only printing messages
497 * that are different from what cpu0 displayed.
499 static unsigned long cpu0_err;
501 static int watchdog_nmi_enable(unsigned int cpu)
503 struct perf_event_attr *wd_attr;
504 struct perf_event *event = per_cpu(watchdog_ev, cpu);
507 * Some kernels need to default hard lockup detection to
508 * 'disabled', for example a guest on a hypervisor.
510 if (!watchdog_hardlockup_detector_is_enabled()) {
511 event = ERR_PTR(-ENOENT);
515 /* is it already setup and enabled? */
516 if (event && event->state > PERF_EVENT_STATE_OFF)
519 /* it is setup but not enabled */
523 wd_attr = &wd_hw_attr;
524 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
526 /* Try to register using hardware perf events */
527 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
530 /* save cpu0 error for future comparision */
531 if (cpu == 0 && IS_ERR(event))
532 cpu0_err = PTR_ERR(event);
534 if (!IS_ERR(event)) {
535 /* only print for cpu0 or different than cpu0 */
536 if (cpu == 0 || cpu0_err)
537 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
541 /* skip displaying the same error again */
542 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
543 return PTR_ERR(event);
545 /* vary the KERN level based on the returned errno */
546 if (PTR_ERR(event) == -EOPNOTSUPP)
547 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
548 else if (PTR_ERR(event) == -ENOENT)
549 pr_warn("disabled (cpu%i): hardware events not enabled\n",
552 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
553 cpu, PTR_ERR(event));
554 return PTR_ERR(event);
558 per_cpu(watchdog_ev, cpu) = event;
560 perf_event_enable(per_cpu(watchdog_ev, cpu));
565 static void watchdog_nmi_disable(unsigned int cpu)
567 struct perf_event *event = per_cpu(watchdog_ev, cpu);
570 perf_event_disable(event);
571 per_cpu(watchdog_ev, cpu) = NULL;
573 /* should be in cleanup, but blocks oprofile */
574 perf_event_release_kernel(event);
577 /* watchdog_nmi_enable() expects this to be zero initially. */
582 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
583 static void watchdog_nmi_disable(unsigned int cpu) { return; }
584 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
586 static struct smp_hotplug_thread watchdog_threads = {
587 .store = &softlockup_watchdog,
588 .thread_should_run = watchdog_should_run,
589 .thread_fn = watchdog,
590 .thread_comm = "watchdog/%u",
591 .setup = watchdog_enable,
592 .cleanup = watchdog_cleanup,
593 .park = watchdog_disable,
594 .unpark = watchdog_enable,
597 static void restart_watchdog_hrtimer(void *info)
599 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
603 * No need to cancel and restart hrtimer if it is currently executing
604 * because it will reprogram itself with the new period now.
605 * We should never see it unqueued here because we are running per-cpu
606 * with interrupts disabled.
608 ret = hrtimer_try_to_cancel(hrtimer);
610 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
611 HRTIMER_MODE_REL_PINNED);
614 static void update_timers(int cpu)
617 * Make sure that perf event counter will adopt to a new
618 * sampling period. Updating the sampling period directly would
619 * be much nicer but we do not have an API for that now so
620 * let's use a big hammer.
621 * Hrtimer will adopt the new period on the next tick but this
622 * might be late already so we have to restart the timer as well.
624 watchdog_nmi_disable(cpu);
625 smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
626 watchdog_nmi_enable(cpu);
629 static void update_timers_all_cpus(void)
634 for_each_online_cpu(cpu)
639 static int watchdog_enable_all_cpus(bool sample_period_changed)
643 if (!watchdog_running) {
644 err = smpboot_register_percpu_thread(&watchdog_threads);
646 pr_err("Failed to create watchdog threads, disabled\n");
648 watchdog_running = 1;
649 } else if (sample_period_changed) {
650 update_timers_all_cpus();
656 /* prepare/enable/disable routines */
657 /* sysctl functions */
659 static void watchdog_disable_all_cpus(void)
661 if (watchdog_running) {
662 watchdog_running = 0;
663 smpboot_unregister_percpu_thread(&watchdog_threads);
668 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
671 int proc_dowatchdog(struct ctl_table *table, int write,
672 void __user *buffer, size_t *lenp, loff_t *ppos)
674 int err, old_thresh, old_enabled;
676 static DEFINE_MUTEX(watchdog_proc_mutex);
678 mutex_lock(&watchdog_proc_mutex);
679 old_thresh = ACCESS_ONCE(watchdog_thresh);
680 old_enabled = ACCESS_ONCE(watchdog_user_enabled);
681 old_hardlockup = watchdog_hardlockup_detector_is_enabled();
683 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
689 * Watchdog threads shouldn't be enabled if they are
690 * disabled. The 'watchdog_running' variable check in
691 * watchdog_*_all_cpus() function takes care of this.
693 if (watchdog_user_enabled && watchdog_thresh) {
695 * Prevent a change in watchdog_thresh accidentally overriding
696 * the enablement of the hardlockup detector.
698 if (watchdog_user_enabled != old_enabled)
699 watchdog_enable_hardlockup_detector(true);
700 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
702 watchdog_disable_all_cpus();
704 /* Restore old values on failure */
706 watchdog_thresh = old_thresh;
707 watchdog_user_enabled = old_enabled;
708 watchdog_enable_hardlockup_detector(old_hardlockup);
711 mutex_unlock(&watchdog_proc_mutex);
714 #endif /* CONFIG_SYSCTL */
716 void __init lockup_detector_init(void)
720 if (watchdog_user_enabled)
721 watchdog_enable_all_cpus(false);