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/delay.h>
19 #include <linux/freezer.h>
20 #include <linux/kthread.h>
21 #include <linux/lockdep.h>
22 #include <linux/notifier.h>
23 #include <linux/module.h>
24 #include <linux/sysctl.h>
25 #include <linux/smpboot.h>
26 #include <linux/sched/rt.h>
28 #include <asm/irq_regs.h>
29 #include <linux/kvm_para.h>
30 #include <linux/perf_event.h>
32 int watchdog_user_enabled = 1;
33 int __read_mostly watchdog_thresh = 10;
35 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
37 #define sysctl_softlockup_all_cpu_backtrace 0
40 static int __read_mostly watchdog_running;
41 static u64 __read_mostly sample_period;
43 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
44 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
45 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
46 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
47 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
48 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
49 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
50 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
51 #ifdef CONFIG_HARDLOCKUP_DETECTOR
52 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
53 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
54 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
55 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
57 static unsigned long soft_lockup_nmi_warn;
61 * Should we panic when a soft-lockup or hard-lockup occurs:
63 #ifdef CONFIG_HARDLOCKUP_DETECTOR
64 static int hardlockup_panic =
65 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
67 static int __init hardlockup_panic_setup(char *str)
69 if (!strncmp(str, "panic", 5))
71 else if (!strncmp(str, "nopanic", 7))
73 else if (!strncmp(str, "0", 1))
74 watchdog_user_enabled = 0;
77 __setup("nmi_watchdog=", hardlockup_panic_setup);
80 unsigned int __read_mostly softlockup_panic =
81 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
83 static int __init softlockup_panic_setup(char *str)
85 softlockup_panic = simple_strtoul(str, NULL, 0);
89 __setup("softlockup_panic=", softlockup_panic_setup);
91 static int __init nowatchdog_setup(char *str)
93 watchdog_user_enabled = 0;
96 __setup("nowatchdog", nowatchdog_setup);
99 static int __init nosoftlockup_setup(char *str)
101 watchdog_user_enabled = 0;
104 __setup("nosoftlockup", nosoftlockup_setup);
107 static int __init softlockup_all_cpu_backtrace_setup(char *str)
109 sysctl_softlockup_all_cpu_backtrace =
110 !!simple_strtol(str, NULL, 0);
113 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
117 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
118 * lockups can have false positives under extreme conditions. So we generally
119 * want a higher threshold for soft lockups than for hard lockups. So we couple
120 * the thresholds with a factor: we make the soft threshold twice the amount of
121 * time the hard threshold is.
123 static int get_softlockup_thresh(void)
125 return watchdog_thresh * 2;
129 * Returns seconds, approximately. We don't need nanosecond
130 * resolution, and we don't need to waste time with a big divide when
133 static unsigned long get_timestamp(void)
135 return local_clock() >> 30LL; /* 2^30 ~= 10^9 */
138 static void set_sample_period(void)
141 * convert watchdog_thresh from seconds to ns
142 * the divide by 5 is to give hrtimer several chances (two
143 * or three with the current relation between the soft
144 * and hard thresholds) to increment before the
145 * hardlockup detector generates a warning
147 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
150 /* Commands for resetting the watchdog */
151 static void __touch_watchdog(void)
153 __this_cpu_write(watchdog_touch_ts, get_timestamp());
156 void touch_softlockup_watchdog(void)
159 * Preemption can be enabled. It doesn't matter which CPU's timestamp
160 * gets zeroed here, so use the raw_ operation.
162 raw_cpu_write(watchdog_touch_ts, 0);
164 EXPORT_SYMBOL(touch_softlockup_watchdog);
166 void touch_all_softlockup_watchdogs(void)
171 * this is done lockless
172 * do we care if a 0 races with a timestamp?
173 * all it means is the softlock check starts one cycle later
175 for_each_online_cpu(cpu)
176 per_cpu(watchdog_touch_ts, cpu) = 0;
179 #ifdef CONFIG_HARDLOCKUP_DETECTOR
180 void touch_nmi_watchdog(void)
183 * Using __raw here because some code paths have
184 * preemption enabled. If preemption is enabled
185 * then interrupts should be enabled too, in which
186 * case we shouldn't have to worry about the watchdog
189 __raw_get_cpu_var(watchdog_nmi_touch) = true;
190 touch_softlockup_watchdog();
192 EXPORT_SYMBOL(touch_nmi_watchdog);
196 void touch_softlockup_watchdog_sync(void)
198 __raw_get_cpu_var(softlockup_touch_sync) = true;
199 __raw_get_cpu_var(watchdog_touch_ts) = 0;
202 #ifdef CONFIG_HARDLOCKUP_DETECTOR
203 /* watchdog detector functions */
204 static int is_hardlockup(void)
206 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
208 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
211 __this_cpu_write(hrtimer_interrupts_saved, hrint);
216 static int is_softlockup(unsigned long touch_ts)
218 unsigned long now = get_timestamp();
220 /* Warn about unreasonable delays: */
221 if (time_after(now, touch_ts + get_softlockup_thresh()))
222 return now - touch_ts;
227 #ifdef CONFIG_HARDLOCKUP_DETECTOR
229 static struct perf_event_attr wd_hw_attr = {
230 .type = PERF_TYPE_HARDWARE,
231 .config = PERF_COUNT_HW_CPU_CYCLES,
232 .size = sizeof(struct perf_event_attr),
237 /* Callback function for perf event subsystem */
238 static void watchdog_overflow_callback(struct perf_event *event,
239 struct perf_sample_data *data,
240 struct pt_regs *regs)
242 /* Ensure the watchdog never gets throttled */
243 event->hw.interrupts = 0;
245 if (__this_cpu_read(watchdog_nmi_touch) == true) {
246 __this_cpu_write(watchdog_nmi_touch, false);
250 /* check for a hardlockup
251 * This is done by making sure our timer interrupt
252 * is incrementing. The timer interrupt should have
253 * fired multiple times before we overflow'd. If it hasn't
254 * then this is a good indication the cpu is stuck
256 if (is_hardlockup()) {
257 int this_cpu = smp_processor_id();
259 /* only print hardlockups once */
260 if (__this_cpu_read(hard_watchdog_warn) == true)
263 if (hardlockup_panic)
264 panic("Watchdog detected hard LOCKUP on cpu %d",
267 WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
270 __this_cpu_write(hard_watchdog_warn, true);
274 __this_cpu_write(hard_watchdog_warn, false);
277 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
279 static void watchdog_interrupt_count(void)
281 __this_cpu_inc(hrtimer_interrupts);
284 static int watchdog_nmi_enable(unsigned int cpu);
285 static void watchdog_nmi_disable(unsigned int cpu);
287 /* watchdog kicker functions */
288 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
290 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
291 struct pt_regs *regs = get_irq_regs();
293 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
295 /* kick the hardlockup detector */
296 watchdog_interrupt_count();
298 /* kick the softlockup detector */
299 wake_up_process(__this_cpu_read(softlockup_watchdog));
302 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
305 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
307 * If the time stamp was touched atomically
308 * make sure the scheduler tick is up to date.
310 __this_cpu_write(softlockup_touch_sync, false);
314 /* Clear the guest paused flag on watchdog reset */
315 kvm_check_and_clear_guest_paused();
317 return HRTIMER_RESTART;
320 /* check for a softlockup
321 * This is done by making sure a high priority task is
322 * being scheduled. The task touches the watchdog to
323 * indicate it is getting cpu time. If it hasn't then
324 * this is a good indication some task is hogging the cpu
326 duration = is_softlockup(touch_ts);
327 if (unlikely(duration)) {
329 * If a virtual machine is stopped by the host it can look to
330 * the watchdog like a soft lockup, check to see if the host
331 * stopped the vm before we issue the warning
333 if (kvm_check_and_clear_guest_paused())
334 return HRTIMER_RESTART;
337 if (__this_cpu_read(soft_watchdog_warn) == true) {
339 * When multiple processes are causing softlockups the
340 * softlockup detector only warns on the first one
341 * because the code relies on a full quiet cycle to
342 * re-arm. The second process prevents the quiet cycle
343 * and never gets reported. Use task pointers to detect
346 if (__this_cpu_read(softlockup_task_ptr_saved) !=
348 __this_cpu_write(soft_watchdog_warn, false);
351 return HRTIMER_RESTART;
354 if (softlockup_all_cpu_backtrace) {
355 /* Prevent multiple soft-lockup reports if one cpu is already
356 * engaged in dumping cpu back traces
358 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
359 /* Someone else will report us. Let's give up */
360 __this_cpu_write(soft_watchdog_warn, true);
361 return HRTIMER_RESTART;
365 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
366 smp_processor_id(), duration,
367 current->comm, task_pid_nr(current));
368 __this_cpu_write(softlockup_task_ptr_saved, current);
370 print_irqtrace_events(current);
376 if (softlockup_all_cpu_backtrace) {
377 /* Avoid generating two back traces for current
378 * given that one is already made above
380 trigger_allbutself_cpu_backtrace();
382 clear_bit(0, &soft_lockup_nmi_warn);
383 /* Barrier to sync with other cpus */
384 smp_mb__after_atomic();
387 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
388 if (softlockup_panic)
389 panic("softlockup: hung tasks");
390 __this_cpu_write(soft_watchdog_warn, true);
392 __this_cpu_write(soft_watchdog_warn, false);
394 return HRTIMER_RESTART;
397 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
399 struct sched_param param = { .sched_priority = prio };
401 sched_setscheduler(current, policy, ¶m);
404 static void watchdog_enable(unsigned int cpu)
406 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
408 /* kick off the timer for the hardlockup detector */
409 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
410 hrtimer->function = watchdog_timer_fn;
412 /* Enable the perf event */
413 watchdog_nmi_enable(cpu);
415 /* done here because hrtimer_start can only pin to smp_processor_id() */
416 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
417 HRTIMER_MODE_REL_PINNED);
419 /* initialize timestamp */
420 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
424 static void watchdog_disable(unsigned int cpu)
426 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
428 watchdog_set_prio(SCHED_NORMAL, 0);
429 hrtimer_cancel(hrtimer);
430 /* disable the perf event */
431 watchdog_nmi_disable(cpu);
434 static void watchdog_cleanup(unsigned int cpu, bool online)
436 watchdog_disable(cpu);
439 static int watchdog_should_run(unsigned int cpu)
441 return __this_cpu_read(hrtimer_interrupts) !=
442 __this_cpu_read(soft_lockup_hrtimer_cnt);
446 * The watchdog thread function - touches the timestamp.
448 * It only runs once every sample_period seconds (4 seconds by
449 * default) to reset the softlockup timestamp. If this gets delayed
450 * for more than 2*watchdog_thresh seconds then the debug-printout
451 * triggers in watchdog_timer_fn().
453 static void watchdog(unsigned int cpu)
455 __this_cpu_write(soft_lockup_hrtimer_cnt,
456 __this_cpu_read(hrtimer_interrupts));
460 #ifdef CONFIG_HARDLOCKUP_DETECTOR
462 * People like the simple clean cpu node info on boot.
463 * Reduce the watchdog noise by only printing messages
464 * that are different from what cpu0 displayed.
466 static unsigned long cpu0_err;
468 static int watchdog_nmi_enable(unsigned int cpu)
470 struct perf_event_attr *wd_attr;
471 struct perf_event *event = per_cpu(watchdog_ev, cpu);
473 /* is it already setup and enabled? */
474 if (event && event->state > PERF_EVENT_STATE_OFF)
477 /* it is setup but not enabled */
481 wd_attr = &wd_hw_attr;
482 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
484 /* Try to register using hardware perf events */
485 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
487 /* save cpu0 error for future comparision */
488 if (cpu == 0 && IS_ERR(event))
489 cpu0_err = PTR_ERR(event);
491 if (!IS_ERR(event)) {
492 /* only print for cpu0 or different than cpu0 */
493 if (cpu == 0 || cpu0_err)
494 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
498 /* skip displaying the same error again */
499 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
500 return PTR_ERR(event);
502 /* vary the KERN level based on the returned errno */
503 if (PTR_ERR(event) == -EOPNOTSUPP)
504 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
505 else if (PTR_ERR(event) == -ENOENT)
506 pr_warn("disabled (cpu%i): hardware events not enabled\n",
509 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
510 cpu, PTR_ERR(event));
511 return PTR_ERR(event);
515 per_cpu(watchdog_ev, cpu) = event;
517 perf_event_enable(per_cpu(watchdog_ev, cpu));
522 static void watchdog_nmi_disable(unsigned int cpu)
524 struct perf_event *event = per_cpu(watchdog_ev, cpu);
527 perf_event_disable(event);
528 per_cpu(watchdog_ev, cpu) = NULL;
530 /* should be in cleanup, but blocks oprofile */
531 perf_event_release_kernel(event);
536 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
537 static void watchdog_nmi_disable(unsigned int cpu) { return; }
538 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
540 static struct smp_hotplug_thread watchdog_threads = {
541 .store = &softlockup_watchdog,
542 .thread_should_run = watchdog_should_run,
543 .thread_fn = watchdog,
544 .thread_comm = "watchdog/%u",
545 .setup = watchdog_enable,
546 .cleanup = watchdog_cleanup,
547 .park = watchdog_disable,
548 .unpark = watchdog_enable,
551 static void restart_watchdog_hrtimer(void *info)
553 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
557 * No need to cancel and restart hrtimer if it is currently executing
558 * because it will reprogram itself with the new period now.
559 * We should never see it unqueued here because we are running per-cpu
560 * with interrupts disabled.
562 ret = hrtimer_try_to_cancel(hrtimer);
564 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
565 HRTIMER_MODE_REL_PINNED);
568 static void update_timers(int cpu)
571 * Make sure that perf event counter will adopt to a new
572 * sampling period. Updating the sampling period directly would
573 * be much nicer but we do not have an API for that now so
574 * let's use a big hammer.
575 * Hrtimer will adopt the new period on the next tick but this
576 * might be late already so we have to restart the timer as well.
578 watchdog_nmi_disable(cpu);
579 smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
580 watchdog_nmi_enable(cpu);
583 static void update_timers_all_cpus(void)
588 for_each_online_cpu(cpu)
593 static int watchdog_enable_all_cpus(bool sample_period_changed)
597 if (!watchdog_running) {
598 err = smpboot_register_percpu_thread(&watchdog_threads);
600 pr_err("Failed to create watchdog threads, disabled\n");
602 watchdog_running = 1;
603 } else if (sample_period_changed) {
604 update_timers_all_cpus();
610 /* prepare/enable/disable routines */
611 /* sysctl functions */
613 static void watchdog_disable_all_cpus(void)
615 if (watchdog_running) {
616 watchdog_running = 0;
617 smpboot_unregister_percpu_thread(&watchdog_threads);
622 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
625 int proc_dowatchdog(struct ctl_table *table, int write,
626 void __user *buffer, size_t *lenp, loff_t *ppos)
628 int err, old_thresh, old_enabled;
629 static DEFINE_MUTEX(watchdog_proc_mutex);
631 mutex_lock(&watchdog_proc_mutex);
632 old_thresh = ACCESS_ONCE(watchdog_thresh);
633 old_enabled = ACCESS_ONCE(watchdog_user_enabled);
635 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
641 * Watchdog threads shouldn't be enabled if they are
642 * disabled. The 'watchdog_running' variable check in
643 * watchdog_*_all_cpus() function takes care of this.
645 if (watchdog_user_enabled && watchdog_thresh)
646 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
648 watchdog_disable_all_cpus();
650 /* Restore old values on failure */
652 watchdog_thresh = old_thresh;
653 watchdog_user_enabled = old_enabled;
656 mutex_unlock(&watchdog_proc_mutex);
659 #endif /* CONFIG_SYSCTL */
661 void __init lockup_detector_init(void)
665 if (watchdog_user_enabled)
666 watchdog_enable_all_cpus(false);