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.
13 #include <linux/cpu.h>
14 #include <linux/nmi.h>
15 #include <linux/init.h>
16 #include <linux/delay.h>
17 #include <linux/freezer.h>
18 #include <linux/kthread.h>
19 #include <linux/lockdep.h>
20 #include <linux/notifier.h>
21 #include <linux/module.h>
22 #include <linux/sysctl.h>
24 #include <asm/irq_regs.h>
25 #include <linux/perf_event.h>
27 int watchdog_enabled = 1;
28 int __read_mostly watchdog_thresh = 10;
30 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
31 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
32 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
33 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
34 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
35 #ifdef CONFIG_HARDLOCKUP_DETECTOR
36 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
37 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
38 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
39 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
40 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
45 * Should we panic when a soft-lockup or hard-lockup occurs:
47 #ifdef CONFIG_HARDLOCKUP_DETECTOR
48 static int hardlockup_panic =
49 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
51 static int __init hardlockup_panic_setup(char *str)
53 if (!strncmp(str, "panic", 5))
55 else if (!strncmp(str, "nopanic", 7))
57 else if (!strncmp(str, "0", 1))
61 __setup("nmi_watchdog=", hardlockup_panic_setup);
64 unsigned int __read_mostly softlockup_panic =
65 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
67 static int __init softlockup_panic_setup(char *str)
69 softlockup_panic = simple_strtoul(str, NULL, 0);
73 __setup("softlockup_panic=", softlockup_panic_setup);
75 static int __init nowatchdog_setup(char *str)
80 __setup("nowatchdog", nowatchdog_setup);
83 static int __init nosoftlockup_setup(char *str)
88 __setup("nosoftlockup", nosoftlockup_setup);
92 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
93 * lockups can have false positives under extreme conditions. So we generally
94 * want a higher threshold for soft lockups than for hard lockups. So we couple
95 * the thresholds with a factor: we make the soft threshold twice the amount of
96 * time the hard threshold is.
98 static int get_softlockup_thresh(void)
100 return watchdog_thresh * 2;
104 * Returns seconds, approximately. We don't need nanosecond
105 * resolution, and we don't need to waste time with a big divide when
108 static unsigned long get_timestamp(int this_cpu)
110 return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */
113 static unsigned long get_sample_period(void)
116 * convert watchdog_thresh from seconds to ns
117 * the divide by 5 is to give hrtimer several chances (two
118 * or three with the current relation between the soft
119 * and hard thresholds) to increment before the
120 * hardlockup detector generates a warning
122 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
125 /* Commands for resetting the watchdog */
126 static void __touch_watchdog(void)
128 int this_cpu = smp_processor_id();
130 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
133 void touch_softlockup_watchdog(void)
135 __this_cpu_write(watchdog_touch_ts, 0);
137 EXPORT_SYMBOL(touch_softlockup_watchdog);
139 void touch_all_softlockup_watchdogs(void)
144 * this is done lockless
145 * do we care if a 0 races with a timestamp?
146 * all it means is the softlock check starts one cycle later
148 for_each_online_cpu(cpu)
149 per_cpu(watchdog_touch_ts, cpu) = 0;
152 #ifdef CONFIG_HARDLOCKUP_DETECTOR
153 void touch_nmi_watchdog(void)
155 if (watchdog_enabled) {
158 for_each_present_cpu(cpu) {
159 if (per_cpu(watchdog_nmi_touch, cpu) != true)
160 per_cpu(watchdog_nmi_touch, cpu) = true;
163 touch_softlockup_watchdog();
165 EXPORT_SYMBOL(touch_nmi_watchdog);
169 void touch_softlockup_watchdog_sync(void)
171 __raw_get_cpu_var(softlockup_touch_sync) = true;
172 __raw_get_cpu_var(watchdog_touch_ts) = 0;
175 #ifdef CONFIG_HARDLOCKUP_DETECTOR
176 /* watchdog detector functions */
177 static int is_hardlockup(void)
179 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
181 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
184 __this_cpu_write(hrtimer_interrupts_saved, hrint);
189 static int is_softlockup(unsigned long touch_ts)
191 unsigned long now = get_timestamp(smp_processor_id());
193 /* Warn about unreasonable delays: */
194 if (time_after(now, touch_ts + get_softlockup_thresh()))
195 return now - touch_ts;
200 #ifdef CONFIG_HARDLOCKUP_DETECTOR
202 static struct perf_event_attr wd_hw_attr = {
203 .type = PERF_TYPE_HARDWARE,
204 .config = PERF_COUNT_HW_CPU_CYCLES,
205 .size = sizeof(struct perf_event_attr),
210 /* Callback function for perf event subsystem */
211 static void watchdog_overflow_callback(struct perf_event *event,
212 struct perf_sample_data *data,
213 struct pt_regs *regs)
215 /* Ensure the watchdog never gets throttled */
216 event->hw.interrupts = 0;
218 if (__this_cpu_read(watchdog_nmi_touch) == true) {
219 __this_cpu_write(watchdog_nmi_touch, false);
223 /* check for a hardlockup
224 * This is done by making sure our timer interrupt
225 * is incrementing. The timer interrupt should have
226 * fired multiple times before we overflow'd. If it hasn't
227 * then this is a good indication the cpu is stuck
229 if (is_hardlockup()) {
230 int this_cpu = smp_processor_id();
232 /* only print hardlockups once */
233 if (__this_cpu_read(hard_watchdog_warn) == true)
236 if (hardlockup_panic)
237 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
239 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
241 __this_cpu_write(hard_watchdog_warn, true);
245 __this_cpu_write(hard_watchdog_warn, false);
248 static void watchdog_interrupt_count(void)
250 __this_cpu_inc(hrtimer_interrupts);
253 static inline void watchdog_interrupt_count(void) { return; }
254 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
256 /* watchdog kicker functions */
257 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
259 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
260 struct pt_regs *regs = get_irq_regs();
263 /* kick the hardlockup detector */
264 watchdog_interrupt_count();
266 /* kick the softlockup detector */
267 wake_up_process(__this_cpu_read(softlockup_watchdog));
270 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
273 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
275 * If the time stamp was touched atomically
276 * make sure the scheduler tick is up to date.
278 __this_cpu_write(softlockup_touch_sync, false);
282 return HRTIMER_RESTART;
285 /* check for a softlockup
286 * This is done by making sure a high priority task is
287 * being scheduled. The task touches the watchdog to
288 * indicate it is getting cpu time. If it hasn't then
289 * this is a good indication some task is hogging the cpu
291 duration = is_softlockup(touch_ts);
292 if (unlikely(duration)) {
294 if (__this_cpu_read(soft_watchdog_warn) == true)
295 return HRTIMER_RESTART;
297 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
298 smp_processor_id(), duration,
299 current->comm, task_pid_nr(current));
301 print_irqtrace_events(current);
307 if (softlockup_panic)
308 panic("softlockup: hung tasks");
309 __this_cpu_write(soft_watchdog_warn, true);
311 __this_cpu_write(soft_watchdog_warn, false);
313 return HRTIMER_RESTART;
318 * The watchdog thread - touches the timestamp.
320 static int watchdog(void *unused)
322 struct sched_param param = { .sched_priority = 0 };
323 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
325 /* initialize timestamp */
328 /* kick off the timer for the hardlockup detector */
329 /* done here because hrtimer_start can only pin to smp_processor_id() */
330 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
331 HRTIMER_MODE_REL_PINNED);
333 set_current_state(TASK_INTERRUPTIBLE);
335 * Run briefly (kicked by the hrtimer callback function) once every
336 * get_sample_period() seconds (4 seconds by default) to reset the
337 * softlockup timestamp. If this gets delayed for more than
338 * 2*watchdog_thresh seconds then the debug-printout triggers in
339 * watchdog_timer_fn().
341 while (!kthread_should_stop()) {
345 if (kthread_should_stop())
348 set_current_state(TASK_INTERRUPTIBLE);
350 __set_current_state(TASK_RUNNING);
351 sched_setscheduler(current, SCHED_NORMAL, ¶m);
356 #ifdef CONFIG_HARDLOCKUP_DETECTOR
357 static int watchdog_nmi_enable(int cpu)
359 struct perf_event_attr *wd_attr;
360 struct perf_event *event = per_cpu(watchdog_ev, cpu);
362 /* is it already setup and enabled? */
363 if (event && event->state > PERF_EVENT_STATE_OFF)
366 /* it is setup but not enabled */
370 wd_attr = &wd_hw_attr;
371 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
373 /* Try to register using hardware perf events */
374 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
375 if (!IS_ERR(event)) {
376 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
381 /* vary the KERN level based on the returned errno */
382 if (PTR_ERR(event) == -EOPNOTSUPP)
383 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
384 else if (PTR_ERR(event) == -ENOENT)
385 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
387 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
388 return PTR_ERR(event);
392 per_cpu(watchdog_ev, cpu) = event;
394 perf_event_enable(per_cpu(watchdog_ev, cpu));
399 static void watchdog_nmi_disable(int cpu)
401 struct perf_event *event = per_cpu(watchdog_ev, cpu);
404 perf_event_disable(event);
405 per_cpu(watchdog_ev, cpu) = NULL;
407 /* should be in cleanup, but blocks oprofile */
408 perf_event_release_kernel(event);
413 static int watchdog_nmi_enable(int cpu) { return 0; }
414 static void watchdog_nmi_disable(int cpu) { return; }
415 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
417 /* prepare/enable/disable routines */
418 static void watchdog_prepare_cpu(int cpu)
420 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
422 WARN_ON(per_cpu(softlockup_watchdog, cpu));
423 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
424 hrtimer->function = watchdog_timer_fn;
427 static int watchdog_enable(int cpu)
429 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
432 /* enable the perf event */
433 err = watchdog_nmi_enable(cpu);
435 /* Regardless of err above, fall through and start softlockup */
437 /* create the watchdog thread */
439 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
440 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
442 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
444 /* if hardlockup hasn't already set this */
446 /* and disable the perf event */
447 watchdog_nmi_disable(cpu);
451 sched_setscheduler(p, SCHED_FIFO, ¶m);
452 kthread_bind(p, cpu);
453 per_cpu(watchdog_touch_ts, cpu) = 0;
454 per_cpu(softlockup_watchdog, cpu) = p;
462 static void watchdog_disable(int cpu)
464 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
465 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
468 * cancel the timer first to stop incrementing the stats
469 * and waking up the kthread
471 hrtimer_cancel(hrtimer);
473 /* disable the perf event */
474 watchdog_nmi_disable(cpu);
476 /* stop the watchdog thread */
478 per_cpu(softlockup_watchdog, cpu) = NULL;
483 /* sysctl functions */
485 static void watchdog_enable_all_cpus(void)
489 watchdog_enabled = 0;
491 for_each_online_cpu(cpu)
492 if (!watchdog_enable(cpu))
493 /* if any cpu succeeds, watchdog is considered
494 enabled for the system */
495 watchdog_enabled = 1;
497 if (!watchdog_enabled)
498 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
502 static void watchdog_disable_all_cpus(void)
506 for_each_online_cpu(cpu)
507 watchdog_disable(cpu);
509 /* if all watchdogs are disabled, then they are disabled for the system */
510 watchdog_enabled = 0;
515 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
518 int proc_dowatchdog(struct ctl_table *table, int write,
519 void __user *buffer, size_t *lenp, loff_t *ppos)
523 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
527 if (watchdog_enabled && watchdog_thresh)
528 watchdog_enable_all_cpus();
530 watchdog_disable_all_cpus();
535 #endif /* CONFIG_SYSCTL */
539 * Create/destroy watchdog threads as CPUs come and go:
542 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
544 int hotcpu = (unsigned long)hcpu;
548 case CPU_UP_PREPARE_FROZEN:
549 watchdog_prepare_cpu(hotcpu);
552 case CPU_ONLINE_FROZEN:
553 if (watchdog_enabled)
554 watchdog_enable(hotcpu);
556 #ifdef CONFIG_HOTPLUG_CPU
557 case CPU_UP_CANCELED:
558 case CPU_UP_CANCELED_FROZEN:
559 watchdog_disable(hotcpu);
562 case CPU_DEAD_FROZEN:
563 watchdog_disable(hotcpu);
565 #endif /* CONFIG_HOTPLUG_CPU */
569 * hardlockup and softlockup are not important enough
570 * to block cpu bring up. Just always succeed and
571 * rely on printk output to flag problems.
576 static struct notifier_block __cpuinitdata cpu_nfb = {
577 .notifier_call = cpu_callback
580 void __init lockup_detector_init(void)
582 void *cpu = (void *)(long)smp_processor_id();
585 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
586 WARN_ON(notifier_to_errno(err));
588 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
589 register_cpu_notifier(&cpu_nfb);