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>
28 * The run state of the lockup detectors is controlled by the content of the
29 * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
30 * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
32 * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
33 * are variables that are only used as an 'interface' between the parameters
34 * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
35 * 'watchdog_thresh' variable is handled differently because its value is not
36 * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
39 #define NMI_WATCHDOG_ENABLED_BIT 0
40 #define SOFT_WATCHDOG_ENABLED_BIT 1
41 #define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT)
42 #define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT)
44 #ifdef CONFIG_HARDLOCKUP_DETECTOR
45 static unsigned long __read_mostly watchdog_enabled
= SOFT_WATCHDOG_ENABLED
|NMI_WATCHDOG_ENABLED
;
47 static unsigned long __read_mostly watchdog_enabled
= SOFT_WATCHDOG_ENABLED
;
49 int __read_mostly nmi_watchdog_enabled
;
50 int __read_mostly soft_watchdog_enabled
;
51 int __read_mostly watchdog_user_enabled
;
52 int __read_mostly watchdog_thresh
= 10;
55 int __read_mostly sysctl_softlockup_all_cpu_backtrace
;
57 #define sysctl_softlockup_all_cpu_backtrace 0
60 static int __read_mostly watchdog_running
;
61 static u64 __read_mostly sample_period
;
63 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
64 static DEFINE_PER_CPU(struct task_struct
*, softlockup_watchdog
);
65 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
66 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
67 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
68 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
69 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt
);
70 static DEFINE_PER_CPU(struct task_struct
*, softlockup_task_ptr_saved
);
71 #ifdef CONFIG_HARDLOCKUP_DETECTOR
72 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
73 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
74 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
75 static DEFINE_PER_CPU(struct perf_event
*, watchdog_ev
);
77 static unsigned long soft_lockup_nmi_warn
;
81 * Should we panic when a soft-lockup or hard-lockup occurs:
83 #ifdef CONFIG_HARDLOCKUP_DETECTOR
84 static int hardlockup_panic
=
85 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
87 static bool hardlockup_detector_enabled
= true;
89 * We may not want to enable hard lockup detection by default in all cases,
90 * for example when running the kernel as a guest on a hypervisor. In these
91 * cases this function can be called to disable hard lockup detection. This
92 * function should only be executed once by the boot processor before the
93 * kernel command line parameters are parsed, because otherwise it is not
94 * possible to override this in hardlockup_panic_setup().
96 void watchdog_enable_hardlockup_detector(bool val
)
98 hardlockup_detector_enabled
= val
;
101 bool watchdog_hardlockup_detector_is_enabled(void)
103 return hardlockup_detector_enabled
;
106 static int __init
hardlockup_panic_setup(char *str
)
108 if (!strncmp(str
, "panic", 5))
109 hardlockup_panic
= 1;
110 else if (!strncmp(str
, "nopanic", 7))
111 hardlockup_panic
= 0;
112 else if (!strncmp(str
, "0", 1))
113 watchdog_user_enabled
= 0;
114 else if (!strncmp(str
, "1", 1) || !strncmp(str
, "2", 1)) {
116 * Setting 'nmi_watchdog=1' or 'nmi_watchdog=2' (legacy option)
117 * has the same effect.
119 watchdog_user_enabled
= 1;
120 watchdog_enable_hardlockup_detector(true);
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_user_enabled
= 0;
143 __setup("nowatchdog", nowatchdog_setup
);
146 static int __init
nosoftlockup_setup(char *str
)
148 watchdog_user_enabled
= 0;
151 __setup("nosoftlockup", nosoftlockup_setup
);
154 static int __init
softlockup_all_cpu_backtrace_setup(char *str
)
156 sysctl_softlockup_all_cpu_backtrace
=
157 !!simple_strtol(str
, NULL
, 0);
160 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup
);
164 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
165 * lockups can have false positives under extreme conditions. So we generally
166 * want a higher threshold for soft lockups than for hard lockups. So we couple
167 * the thresholds with a factor: we make the soft threshold twice the amount of
168 * time the hard threshold is.
170 static int get_softlockup_thresh(void)
172 return watchdog_thresh
* 2;
176 * Returns seconds, approximately. We don't need nanosecond
177 * resolution, and we don't need to waste time with a big divide when
180 static unsigned long get_timestamp(void)
182 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
185 static void set_sample_period(void)
188 * convert watchdog_thresh from seconds to ns
189 * the divide by 5 is to give hrtimer several chances (two
190 * or three with the current relation between the soft
191 * and hard thresholds) to increment before the
192 * hardlockup detector generates a warning
194 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
197 /* Commands for resetting the watchdog */
198 static void __touch_watchdog(void)
200 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
203 void touch_softlockup_watchdog(void)
206 * Preemption can be enabled. It doesn't matter which CPU's timestamp
207 * gets zeroed here, so use the raw_ operation.
209 raw_cpu_write(watchdog_touch_ts
, 0);
211 EXPORT_SYMBOL(touch_softlockup_watchdog
);
213 void touch_all_softlockup_watchdogs(void)
218 * this is done lockless
219 * do we care if a 0 races with a timestamp?
220 * all it means is the softlock check starts one cycle later
222 for_each_online_cpu(cpu
)
223 per_cpu(watchdog_touch_ts
, cpu
) = 0;
226 #ifdef CONFIG_HARDLOCKUP_DETECTOR
227 void touch_nmi_watchdog(void)
230 * Using __raw here because some code paths have
231 * preemption enabled. If preemption is enabled
232 * then interrupts should be enabled too, in which
233 * case we shouldn't have to worry about the watchdog
236 raw_cpu_write(watchdog_nmi_touch
, true);
237 touch_softlockup_watchdog();
239 EXPORT_SYMBOL(touch_nmi_watchdog
);
243 void touch_softlockup_watchdog_sync(void)
245 __this_cpu_write(softlockup_touch_sync
, true);
246 __this_cpu_write(watchdog_touch_ts
, 0);
249 #ifdef CONFIG_HARDLOCKUP_DETECTOR
250 /* watchdog detector functions */
251 static int is_hardlockup(void)
253 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
255 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
258 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
263 static int is_softlockup(unsigned long touch_ts
)
265 unsigned long now
= get_timestamp();
267 /* Warn about unreasonable delays: */
268 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
269 return now
- touch_ts
;
274 #ifdef CONFIG_HARDLOCKUP_DETECTOR
276 static struct perf_event_attr wd_hw_attr
= {
277 .type
= PERF_TYPE_HARDWARE
,
278 .config
= PERF_COUNT_HW_CPU_CYCLES
,
279 .size
= sizeof(struct perf_event_attr
),
284 /* Callback function for perf event subsystem */
285 static void watchdog_overflow_callback(struct perf_event
*event
,
286 struct perf_sample_data
*data
,
287 struct pt_regs
*regs
)
289 /* Ensure the watchdog never gets throttled */
290 event
->hw
.interrupts
= 0;
292 if (__this_cpu_read(watchdog_nmi_touch
) == true) {
293 __this_cpu_write(watchdog_nmi_touch
, false);
297 /* check for a hardlockup
298 * This is done by making sure our timer interrupt
299 * is incrementing. The timer interrupt should have
300 * fired multiple times before we overflow'd. If it hasn't
301 * then this is a good indication the cpu is stuck
303 if (is_hardlockup()) {
304 int this_cpu
= smp_processor_id();
306 /* only print hardlockups once */
307 if (__this_cpu_read(hard_watchdog_warn
) == true)
310 if (hardlockup_panic
)
311 panic("Watchdog detected hard LOCKUP on cpu %d",
314 WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
317 __this_cpu_write(hard_watchdog_warn
, true);
321 __this_cpu_write(hard_watchdog_warn
, false);
324 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
326 static void watchdog_interrupt_count(void)
328 __this_cpu_inc(hrtimer_interrupts
);
331 static int watchdog_nmi_enable(unsigned int cpu
);
332 static void watchdog_nmi_disable(unsigned int cpu
);
334 /* watchdog kicker functions */
335 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
337 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
338 struct pt_regs
*regs
= get_irq_regs();
340 int softlockup_all_cpu_backtrace
= sysctl_softlockup_all_cpu_backtrace
;
342 /* kick the hardlockup detector */
343 watchdog_interrupt_count();
345 /* kick the softlockup detector */
346 wake_up_process(__this_cpu_read(softlockup_watchdog
));
349 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
352 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
354 * If the time stamp was touched atomically
355 * make sure the scheduler tick is up to date.
357 __this_cpu_write(softlockup_touch_sync
, false);
361 /* Clear the guest paused flag on watchdog reset */
362 kvm_check_and_clear_guest_paused();
364 return HRTIMER_RESTART
;
367 /* check for a softlockup
368 * This is done by making sure a high priority task is
369 * being scheduled. The task touches the watchdog to
370 * indicate it is getting cpu time. If it hasn't then
371 * this is a good indication some task is hogging the cpu
373 duration
= is_softlockup(touch_ts
);
374 if (unlikely(duration
)) {
376 * If a virtual machine is stopped by the host it can look to
377 * the watchdog like a soft lockup, check to see if the host
378 * stopped the vm before we issue the warning
380 if (kvm_check_and_clear_guest_paused())
381 return HRTIMER_RESTART
;
384 if (__this_cpu_read(soft_watchdog_warn
) == true) {
386 * When multiple processes are causing softlockups the
387 * softlockup detector only warns on the first one
388 * because the code relies on a full quiet cycle to
389 * re-arm. The second process prevents the quiet cycle
390 * and never gets reported. Use task pointers to detect
393 if (__this_cpu_read(softlockup_task_ptr_saved
) !=
395 __this_cpu_write(soft_watchdog_warn
, false);
398 return HRTIMER_RESTART
;
401 if (softlockup_all_cpu_backtrace
) {
402 /* Prevent multiple soft-lockup reports if one cpu is already
403 * engaged in dumping cpu back traces
405 if (test_and_set_bit(0, &soft_lockup_nmi_warn
)) {
406 /* Someone else will report us. Let's give up */
407 __this_cpu_write(soft_watchdog_warn
, true);
408 return HRTIMER_RESTART
;
412 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
413 smp_processor_id(), duration
,
414 current
->comm
, task_pid_nr(current
));
415 __this_cpu_write(softlockup_task_ptr_saved
, current
);
417 print_irqtrace_events(current
);
423 if (softlockup_all_cpu_backtrace
) {
424 /* Avoid generating two back traces for current
425 * given that one is already made above
427 trigger_allbutself_cpu_backtrace();
429 clear_bit(0, &soft_lockup_nmi_warn
);
430 /* Barrier to sync with other cpus */
431 smp_mb__after_atomic();
434 add_taint(TAINT_SOFTLOCKUP
, LOCKDEP_STILL_OK
);
435 if (softlockup_panic
)
436 panic("softlockup: hung tasks");
437 __this_cpu_write(soft_watchdog_warn
, true);
439 __this_cpu_write(soft_watchdog_warn
, false);
441 return HRTIMER_RESTART
;
444 static void watchdog_set_prio(unsigned int policy
, unsigned int prio
)
446 struct sched_param param
= { .sched_priority
= prio
};
448 sched_setscheduler(current
, policy
, ¶m
);
451 static void watchdog_enable(unsigned int cpu
)
453 struct hrtimer
*hrtimer
= raw_cpu_ptr(&watchdog_hrtimer
);
455 /* kick off the timer for the hardlockup detector */
456 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
457 hrtimer
->function
= watchdog_timer_fn
;
459 /* Enable the perf event */
460 watchdog_nmi_enable(cpu
);
462 /* done here because hrtimer_start can only pin to smp_processor_id() */
463 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
464 HRTIMER_MODE_REL_PINNED
);
466 /* initialize timestamp */
467 watchdog_set_prio(SCHED_FIFO
, MAX_RT_PRIO
- 1);
471 static void watchdog_disable(unsigned int cpu
)
473 struct hrtimer
*hrtimer
= raw_cpu_ptr(&watchdog_hrtimer
);
475 watchdog_set_prio(SCHED_NORMAL
, 0);
476 hrtimer_cancel(hrtimer
);
477 /* disable the perf event */
478 watchdog_nmi_disable(cpu
);
481 static void watchdog_cleanup(unsigned int cpu
, bool online
)
483 watchdog_disable(cpu
);
486 static int watchdog_should_run(unsigned int cpu
)
488 return __this_cpu_read(hrtimer_interrupts
) !=
489 __this_cpu_read(soft_lockup_hrtimer_cnt
);
493 * The watchdog thread function - touches the timestamp.
495 * It only runs once every sample_period seconds (4 seconds by
496 * default) to reset the softlockup timestamp. If this gets delayed
497 * for more than 2*watchdog_thresh seconds then the debug-printout
498 * triggers in watchdog_timer_fn().
500 static void watchdog(unsigned int cpu
)
502 __this_cpu_write(soft_lockup_hrtimer_cnt
,
503 __this_cpu_read(hrtimer_interrupts
));
507 #ifdef CONFIG_HARDLOCKUP_DETECTOR
509 * People like the simple clean cpu node info on boot.
510 * Reduce the watchdog noise by only printing messages
511 * that are different from what cpu0 displayed.
513 static unsigned long cpu0_err
;
515 static int watchdog_nmi_enable(unsigned int cpu
)
517 struct perf_event_attr
*wd_attr
;
518 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
521 * Some kernels need to default hard lockup detection to
522 * 'disabled', for example a guest on a hypervisor.
524 if (!watchdog_hardlockup_detector_is_enabled()) {
525 event
= ERR_PTR(-ENOENT
);
529 /* is it already setup and enabled? */
530 if (event
&& event
->state
> PERF_EVENT_STATE_OFF
)
533 /* it is setup but not enabled */
537 wd_attr
= &wd_hw_attr
;
538 wd_attr
->sample_period
= hw_nmi_get_sample_period(watchdog_thresh
);
540 /* Try to register using hardware perf events */
541 event
= perf_event_create_kernel_counter(wd_attr
, cpu
, NULL
, watchdog_overflow_callback
, NULL
);
544 /* save cpu0 error for future comparision */
545 if (cpu
== 0 && IS_ERR(event
))
546 cpu0_err
= PTR_ERR(event
);
548 if (!IS_ERR(event
)) {
549 /* only print for cpu0 or different than cpu0 */
550 if (cpu
== 0 || cpu0_err
)
551 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
555 /* skip displaying the same error again */
556 if (cpu
> 0 && (PTR_ERR(event
) == cpu0_err
))
557 return PTR_ERR(event
);
559 /* vary the KERN level based on the returned errno */
560 if (PTR_ERR(event
) == -EOPNOTSUPP
)
561 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu
);
562 else if (PTR_ERR(event
) == -ENOENT
)
563 pr_warn("disabled (cpu%i): hardware events not enabled\n",
566 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
567 cpu
, PTR_ERR(event
));
568 return PTR_ERR(event
);
572 per_cpu(watchdog_ev
, cpu
) = event
;
574 perf_event_enable(per_cpu(watchdog_ev
, cpu
));
579 static void watchdog_nmi_disable(unsigned int cpu
)
581 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
584 perf_event_disable(event
);
585 per_cpu(watchdog_ev
, cpu
) = NULL
;
587 /* should be in cleanup, but blocks oprofile */
588 perf_event_release_kernel(event
);
591 /* watchdog_nmi_enable() expects this to be zero initially. */
596 static int watchdog_nmi_enable(unsigned int cpu
) { return 0; }
597 static void watchdog_nmi_disable(unsigned int cpu
) { return; }
598 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
600 static struct smp_hotplug_thread watchdog_threads
= {
601 .store
= &softlockup_watchdog
,
602 .thread_should_run
= watchdog_should_run
,
603 .thread_fn
= watchdog
,
604 .thread_comm
= "watchdog/%u",
605 .setup
= watchdog_enable
,
606 .cleanup
= watchdog_cleanup
,
607 .park
= watchdog_disable
,
608 .unpark
= watchdog_enable
,
611 static void restart_watchdog_hrtimer(void *info
)
613 struct hrtimer
*hrtimer
= raw_cpu_ptr(&watchdog_hrtimer
);
617 * No need to cancel and restart hrtimer if it is currently executing
618 * because it will reprogram itself with the new period now.
619 * We should never see it unqueued here because we are running per-cpu
620 * with interrupts disabled.
622 ret
= hrtimer_try_to_cancel(hrtimer
);
624 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
625 HRTIMER_MODE_REL_PINNED
);
628 static void update_timers(int cpu
)
631 * Make sure that perf event counter will adopt to a new
632 * sampling period. Updating the sampling period directly would
633 * be much nicer but we do not have an API for that now so
634 * let's use a big hammer.
635 * Hrtimer will adopt the new period on the next tick but this
636 * might be late already so we have to restart the timer as well.
638 watchdog_nmi_disable(cpu
);
639 smp_call_function_single(cpu
, restart_watchdog_hrtimer
, NULL
, 1);
640 watchdog_nmi_enable(cpu
);
643 static void update_timers_all_cpus(void)
648 for_each_online_cpu(cpu
)
653 static int watchdog_enable_all_cpus(bool sample_period_changed
)
657 if (!watchdog_running
) {
658 err
= smpboot_register_percpu_thread(&watchdog_threads
);
660 pr_err("Failed to create watchdog threads, disabled\n");
662 watchdog_running
= 1;
663 } else if (sample_period_changed
) {
664 update_timers_all_cpus();
670 /* prepare/enable/disable routines */
671 /* sysctl functions */
673 static void watchdog_disable_all_cpus(void)
675 if (watchdog_running
) {
676 watchdog_running
= 0;
677 smpboot_unregister_percpu_thread(&watchdog_threads
);
682 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
685 int proc_dowatchdog(struct ctl_table
*table
, int write
,
686 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
688 int err
, old_thresh
, old_enabled
;
690 static DEFINE_MUTEX(watchdog_proc_mutex
);
692 mutex_lock(&watchdog_proc_mutex
);
693 old_thresh
= ACCESS_ONCE(watchdog_thresh
);
694 old_enabled
= ACCESS_ONCE(watchdog_user_enabled
);
695 old_hardlockup
= watchdog_hardlockup_detector_is_enabled();
697 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
703 * Watchdog threads shouldn't be enabled if they are
704 * disabled. The 'watchdog_running' variable check in
705 * watchdog_*_all_cpus() function takes care of this.
707 if (watchdog_user_enabled
&& watchdog_thresh
) {
709 * Prevent a change in watchdog_thresh accidentally overriding
710 * the enablement of the hardlockup detector.
712 if (watchdog_user_enabled
!= old_enabled
)
713 watchdog_enable_hardlockup_detector(true);
714 err
= watchdog_enable_all_cpus(old_thresh
!= watchdog_thresh
);
716 watchdog_disable_all_cpus();
718 /* Restore old values on failure */
720 watchdog_thresh
= old_thresh
;
721 watchdog_user_enabled
= old_enabled
;
722 watchdog_enable_hardlockup_detector(old_hardlockup
);
725 mutex_unlock(&watchdog_proc_mutex
);
728 #endif /* CONFIG_SYSCTL */
730 void __init
lockup_detector_init(void)
734 if (watchdog_user_enabled
)
735 watchdog_enable_all_cpus(false);