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_enabled = 1;
33 int __read_mostly watchdog_thresh = 10;
34 static int __read_mostly watchdog_disabled;
35 static u64 __read_mostly sample_period;
37 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
38 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
39 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
40 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
41 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
42 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
43 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
45 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
46 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
47 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
48 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
53 * Should we panic when a soft-lockup or hard-lockup occurs:
55 #ifdef CONFIG_HARDLOCKUP_DETECTOR
56 static int hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
59 static int __init hardlockup_panic_setup(char *str)
61 if (!strncmp(str, "panic", 5))
63 else if (!strncmp(str, "nopanic", 7))
65 else if (!strncmp(str, "0", 1))
69 __setup("nmi_watchdog=", hardlockup_panic_setup);
72 unsigned int __read_mostly softlockup_panic =
73 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
75 static int __init softlockup_panic_setup(char *str)
77 softlockup_panic = simple_strtoul(str, NULL, 0);
81 __setup("softlockup_panic=", softlockup_panic_setup);
83 static int __init nowatchdog_setup(char *str)
88 __setup("nowatchdog", nowatchdog_setup);
91 static int __init nosoftlockup_setup(char *str)
96 __setup("nosoftlockup", nosoftlockup_setup);
100 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
101 * lockups can have false positives under extreme conditions. So we generally
102 * want a higher threshold for soft lockups than for hard lockups. So we couple
103 * the thresholds with a factor: we make the soft threshold twice the amount of
104 * time the hard threshold is.
106 static int get_softlockup_thresh(void)
108 return watchdog_thresh * 2;
112 * Returns seconds, approximately. We don't need nanosecond
113 * resolution, and we don't need to waste time with a big divide when
116 static unsigned long get_timestamp(void)
118 return local_clock() >> 30LL; /* 2^30 ~= 10^9 */
121 static void set_sample_period(void)
124 * convert watchdog_thresh from seconds to ns
125 * the divide by 5 is to give hrtimer several chances (two
126 * or three with the current relation between the soft
127 * and hard thresholds) to increment before the
128 * hardlockup detector generates a warning
130 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
133 /* Commands for resetting the watchdog */
134 static void __touch_watchdog(void)
136 __this_cpu_write(watchdog_touch_ts, get_timestamp());
139 void touch_softlockup_watchdog(void)
141 __this_cpu_write(watchdog_touch_ts, 0);
143 EXPORT_SYMBOL(touch_softlockup_watchdog);
145 void touch_all_softlockup_watchdogs(void)
150 * this is done lockless
151 * do we care if a 0 races with a timestamp?
152 * all it means is the softlock check starts one cycle later
154 for_each_online_cpu(cpu)
155 per_cpu(watchdog_touch_ts, cpu) = 0;
158 #ifdef CONFIG_HARDLOCKUP_DETECTOR
159 void touch_nmi_watchdog(void)
161 if (watchdog_enabled) {
164 for_each_present_cpu(cpu) {
165 if (per_cpu(watchdog_nmi_touch, cpu) != true)
166 per_cpu(watchdog_nmi_touch, cpu) = true;
169 touch_softlockup_watchdog();
171 EXPORT_SYMBOL(touch_nmi_watchdog);
175 void touch_softlockup_watchdog_sync(void)
177 __raw_get_cpu_var(softlockup_touch_sync) = true;
178 __raw_get_cpu_var(watchdog_touch_ts) = 0;
181 #ifdef CONFIG_HARDLOCKUP_DETECTOR
182 /* watchdog detector functions */
183 static int is_hardlockup(void)
185 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
187 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
190 __this_cpu_write(hrtimer_interrupts_saved, hrint);
195 static int is_softlockup(unsigned long touch_ts)
197 unsigned long now = get_timestamp();
199 /* Warn about unreasonable delays: */
200 if (time_after(now, touch_ts + get_softlockup_thresh()))
201 return now - touch_ts;
206 #ifdef CONFIG_HARDLOCKUP_DETECTOR
208 static struct perf_event_attr wd_hw_attr = {
209 .type = PERF_TYPE_HARDWARE,
210 .config = PERF_COUNT_HW_CPU_CYCLES,
211 .size = sizeof(struct perf_event_attr),
216 /* Callback function for perf event subsystem */
217 static void watchdog_overflow_callback(struct perf_event *event,
218 struct perf_sample_data *data,
219 struct pt_regs *regs)
221 /* Ensure the watchdog never gets throttled */
222 event->hw.interrupts = 0;
224 if (__this_cpu_read(watchdog_nmi_touch) == true) {
225 __this_cpu_write(watchdog_nmi_touch, false);
229 /* check for a hardlockup
230 * This is done by making sure our timer interrupt
231 * is incrementing. The timer interrupt should have
232 * fired multiple times before we overflow'd. If it hasn't
233 * then this is a good indication the cpu is stuck
235 if (is_hardlockup()) {
236 int this_cpu = smp_processor_id();
238 /* only print hardlockups once */
239 if (__this_cpu_read(hard_watchdog_warn) == true)
242 if (hardlockup_panic) {
243 trigger_all_cpu_backtrace();
244 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
246 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
249 __this_cpu_write(hard_watchdog_warn, true);
253 __this_cpu_write(hard_watchdog_warn, false);
256 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
258 static void watchdog_interrupt_count(void)
260 __this_cpu_inc(hrtimer_interrupts);
263 static int watchdog_nmi_enable(unsigned int cpu);
264 static void watchdog_nmi_disable(unsigned int cpu);
266 /* watchdog kicker functions */
267 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
269 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
270 struct pt_regs *regs = get_irq_regs();
273 /* kick the hardlockup detector */
274 watchdog_interrupt_count();
276 /* kick the softlockup detector */
277 wake_up_process(__this_cpu_read(softlockup_watchdog));
280 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
283 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
285 * If the time stamp was touched atomically
286 * make sure the scheduler tick is up to date.
288 __this_cpu_write(softlockup_touch_sync, false);
292 /* Clear the guest paused flag on watchdog reset */
293 kvm_check_and_clear_guest_paused();
295 return HRTIMER_RESTART;
298 /* check for a softlockup
299 * This is done by making sure a high priority task is
300 * being scheduled. The task touches the watchdog to
301 * indicate it is getting cpu time. If it hasn't then
302 * this is a good indication some task is hogging the cpu
304 duration = is_softlockup(touch_ts);
305 if (unlikely(duration)) {
307 * If a virtual machine is stopped by the host it can look to
308 * the watchdog like a soft lockup, check to see if the host
309 * stopped the vm before we issue the warning
311 if (kvm_check_and_clear_guest_paused())
312 return HRTIMER_RESTART;
315 if (__this_cpu_read(soft_watchdog_warn) == true)
316 return HRTIMER_RESTART;
318 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
319 smp_processor_id(), duration,
320 current->comm, task_pid_nr(current));
322 print_irqtrace_events(current);
328 if (softlockup_panic) {
329 trigger_all_cpu_backtrace();
330 panic("softlockup: hung tasks");
332 __this_cpu_write(soft_watchdog_warn, true);
334 __this_cpu_write(soft_watchdog_warn, false);
336 return HRTIMER_RESTART;
339 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
341 struct sched_param param = { .sched_priority = prio };
343 sched_setscheduler(current, policy, ¶m);
346 static void watchdog_enable(unsigned int cpu)
348 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
350 /* kick off the timer for the hardlockup detector */
351 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
352 hrtimer->function = watchdog_timer_fn;
354 if (!watchdog_enabled) {
355 kthread_park(current);
359 /* Enable the perf event */
360 watchdog_nmi_enable(cpu);
362 /* done here because hrtimer_start can only pin to smp_processor_id() */
363 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
364 HRTIMER_MODE_REL_PINNED);
366 /* initialize timestamp */
367 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
371 static void watchdog_disable(unsigned int cpu)
373 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
375 watchdog_set_prio(SCHED_NORMAL, 0);
376 hrtimer_cancel(hrtimer);
377 /* disable the perf event */
378 watchdog_nmi_disable(cpu);
381 static int watchdog_should_run(unsigned int cpu)
383 return __this_cpu_read(hrtimer_interrupts) !=
384 __this_cpu_read(soft_lockup_hrtimer_cnt);
388 * The watchdog thread function - touches the timestamp.
390 * It only runs once every sample_period seconds (4 seconds by
391 * default) to reset the softlockup timestamp. If this gets delayed
392 * for more than 2*watchdog_thresh seconds then the debug-printout
393 * triggers in watchdog_timer_fn().
395 static void watchdog(unsigned int cpu)
397 __this_cpu_write(soft_lockup_hrtimer_cnt,
398 __this_cpu_read(hrtimer_interrupts));
402 #ifdef CONFIG_HARDLOCKUP_DETECTOR
404 * People like the simple clean cpu node info on boot.
405 * Reduce the watchdog noise by only printing messages
406 * that are different from what cpu0 displayed.
408 static unsigned long cpu0_err;
410 static int watchdog_nmi_enable(unsigned int cpu)
412 struct perf_event_attr *wd_attr;
413 struct perf_event *event = per_cpu(watchdog_ev, cpu);
415 /* is it already setup and enabled? */
416 if (event && event->state > PERF_EVENT_STATE_OFF)
419 /* it is setup but not enabled */
423 wd_attr = &wd_hw_attr;
424 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
426 /* Try to register using hardware perf events */
427 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
429 /* save cpu0 error for future comparision */
430 if (cpu == 0 && IS_ERR(event))
431 cpu0_err = PTR_ERR(event);
433 if (!IS_ERR(event)) {
434 /* only print for cpu0 or different than cpu0 */
435 if (cpu == 0 || cpu0_err)
436 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
440 /* skip displaying the same error again */
441 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
442 return PTR_ERR(event);
444 /* vary the KERN level based on the returned errno */
445 if (PTR_ERR(event) == -EOPNOTSUPP)
446 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
447 else if (PTR_ERR(event) == -ENOENT)
448 pr_warning("disabled (cpu%i): hardware events not enabled\n",
451 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
452 cpu, PTR_ERR(event));
453 return PTR_ERR(event);
457 per_cpu(watchdog_ev, cpu) = event;
459 perf_event_enable(per_cpu(watchdog_ev, cpu));
464 static void watchdog_nmi_disable(unsigned int cpu)
466 struct perf_event *event = per_cpu(watchdog_ev, cpu);
469 perf_event_disable(event);
470 per_cpu(watchdog_ev, cpu) = NULL;
472 /* should be in cleanup, but blocks oprofile */
473 perf_event_release_kernel(event);
478 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
479 static void watchdog_nmi_disable(unsigned int cpu) { return; }
480 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
482 /* prepare/enable/disable routines */
483 /* sysctl functions */
485 static void watchdog_enable_all_cpus(void)
489 if (watchdog_disabled) {
490 watchdog_disabled = 0;
491 for_each_online_cpu(cpu)
492 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
496 static void watchdog_disable_all_cpus(void)
500 if (!watchdog_disabled) {
501 watchdog_disabled = 1;
502 for_each_online_cpu(cpu)
503 kthread_park(per_cpu(softlockup_watchdog, cpu));
508 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
511 int proc_dowatchdog(struct ctl_table *table, int write,
512 void __user *buffer, size_t *lenp, loff_t *ppos)
516 if (watchdog_disabled < 0)
519 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
525 * Watchdog threads shouldn't be enabled if they are
526 * disabled. The 'watchdog_disabled' variable check in
527 * watchdog_*_all_cpus() function takes care of this.
529 if (watchdog_enabled && watchdog_thresh)
530 watchdog_enable_all_cpus();
532 watchdog_disable_all_cpus();
536 #endif /* CONFIG_SYSCTL */
538 static struct smp_hotplug_thread watchdog_threads = {
539 .store = &softlockup_watchdog,
540 .thread_should_run = watchdog_should_run,
541 .thread_fn = watchdog,
542 .thread_comm = "watchdog/%u",
543 .setup = watchdog_enable,
544 .park = watchdog_disable,
545 .unpark = watchdog_enable,
548 void __init lockup_detector_init(void)
551 if (smpboot_register_percpu_thread(&watchdog_threads)) {
552 pr_err("Failed to create watchdog threads, disabled\n");
553 watchdog_disabled = -ENODEV;