1 // SPDX-License-Identifier: GPL-2.0
3 * Detect hard and soft lockups on a system
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
7 * Note: Most of this code is borrowed heavily from the original softlockup
8 * detector, so thanks to Ingo for the initial implementation.
9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
10 * to those contributors as well.
13 #define pr_fmt(fmt) "watchdog: " fmt
16 #include <linux/cpu.h>
17 #include <linux/nmi.h>
18 #include <linux/init.h>
19 #include <linux/module.h>
20 #include <linux/sysctl.h>
21 #include <linux/tick.h>
22 #include <linux/sched/clock.h>
23 #include <linux/sched/debug.h>
24 #include <linux/sched/isolation.h>
25 #include <linux/stop_machine.h>
27 #include <asm/irq_regs.h>
28 #include <linux/kvm_para.h>
30 static DEFINE_MUTEX(watchdog_mutex
);
32 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
33 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
34 # define NMI_WATCHDOG_DEFAULT 1
36 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37 # define NMI_WATCHDOG_DEFAULT 0
40 unsigned long __read_mostly watchdog_enabled
;
41 int __read_mostly watchdog_user_enabled
= 1;
42 int __read_mostly nmi_watchdog_user_enabled
= NMI_WATCHDOG_DEFAULT
;
43 int __read_mostly soft_watchdog_user_enabled
= 1;
44 int __read_mostly watchdog_thresh
= 10;
45 static int __read_mostly nmi_watchdog_available
;
47 static struct cpumask watchdog_allowed_mask __read_mostly
;
49 struct cpumask watchdog_cpumask __read_mostly
;
50 unsigned long *watchdog_cpumask_bits
= cpumask_bits(&watchdog_cpumask
);
52 #ifdef CONFIG_HARDLOCKUP_DETECTOR
54 * Should we panic when a soft-lockup or hard-lockup occurs:
56 unsigned int __read_mostly hardlockup_panic
=
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
59 * We may not want to enable hard lockup detection by default in all cases,
60 * for example when running the kernel as a guest on a hypervisor. In these
61 * cases this function can be called to disable hard lockup detection. This
62 * function should only be executed once by the boot processor before the
63 * kernel command line parameters are parsed, because otherwise it is not
64 * possible to override this in hardlockup_panic_setup().
66 void __init
hardlockup_detector_disable(void)
68 nmi_watchdog_user_enabled
= 0;
71 static int __init
hardlockup_panic_setup(char *str
)
73 if (!strncmp(str
, "panic", 5))
75 else if (!strncmp(str
, "nopanic", 7))
77 else if (!strncmp(str
, "0", 1))
78 nmi_watchdog_user_enabled
= 0;
79 else if (!strncmp(str
, "1", 1))
80 nmi_watchdog_user_enabled
= 1;
83 __setup("nmi_watchdog=", hardlockup_panic_setup
);
86 int __read_mostly sysctl_hardlockup_all_cpu_backtrace
;
88 static int __init
hardlockup_all_cpu_backtrace_setup(char *str
)
90 sysctl_hardlockup_all_cpu_backtrace
= !!simple_strtol(str
, NULL
, 0);
93 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup
);
94 # endif /* CONFIG_SMP */
95 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
98 * These functions can be overridden if an architecture implements its
99 * own hardlockup detector.
101 * watchdog_nmi_enable/disable can be implemented to start and stop when
102 * softlockup watchdog threads start and stop. The arch must select the
103 * SOFTLOCKUP_DETECTOR Kconfig.
105 int __weak
watchdog_nmi_enable(unsigned int cpu
)
107 hardlockup_detector_perf_enable();
111 void __weak
watchdog_nmi_disable(unsigned int cpu
)
113 hardlockup_detector_perf_disable();
116 /* Return 0, if a NMI watchdog is available. Error code otherwise */
117 int __weak __init
watchdog_nmi_probe(void)
119 return hardlockup_detector_perf_init();
123 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
125 * The reconfiguration steps are:
126 * watchdog_nmi_stop();
127 * update_variables();
128 * watchdog_nmi_start();
130 void __weak
watchdog_nmi_stop(void) { }
133 * watchdog_nmi_start - Start the watchdog after reconfiguration
135 * Counterpart to watchdog_nmi_stop().
137 * The following variables have been updated in update_variables() and
138 * contain the currently valid configuration:
143 void __weak
watchdog_nmi_start(void) { }
146 * lockup_detector_update_enable - Update the sysctl enable bit
148 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
149 * can't race with watchdog_nmi_disable().
151 static void lockup_detector_update_enable(void)
153 watchdog_enabled
= 0;
154 if (!watchdog_user_enabled
)
156 if (nmi_watchdog_available
&& nmi_watchdog_user_enabled
)
157 watchdog_enabled
|= NMI_WATCHDOG_ENABLED
;
158 if (soft_watchdog_user_enabled
)
159 watchdog_enabled
|= SOFT_WATCHDOG_ENABLED
;
162 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
164 #define SOFTLOCKUP_RESET ULONG_MAX
166 /* Global variables, exported for sysctl */
167 unsigned int __read_mostly softlockup_panic
=
168 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE
;
170 static bool softlockup_initialized __read_mostly
;
171 static u64 __read_mostly sample_period
;
173 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
174 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
175 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
176 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
177 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
178 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
179 static unsigned long soft_lockup_nmi_warn
;
181 static int __init
softlockup_panic_setup(char *str
)
183 softlockup_panic
= simple_strtoul(str
, NULL
, 0);
186 __setup("softlockup_panic=", softlockup_panic_setup
);
188 static int __init
nowatchdog_setup(char *str
)
190 watchdog_user_enabled
= 0;
193 __setup("nowatchdog", nowatchdog_setup
);
195 static int __init
nosoftlockup_setup(char *str
)
197 soft_watchdog_user_enabled
= 0;
200 __setup("nosoftlockup", nosoftlockup_setup
);
202 static int __init
watchdog_thresh_setup(char *str
)
204 get_option(&str
, &watchdog_thresh
);
207 __setup("watchdog_thresh=", watchdog_thresh_setup
);
210 int __read_mostly sysctl_softlockup_all_cpu_backtrace
;
212 static int __init
softlockup_all_cpu_backtrace_setup(char *str
)
214 sysctl_softlockup_all_cpu_backtrace
= !!simple_strtol(str
, NULL
, 0);
217 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup
);
220 static void __lockup_detector_cleanup(void);
223 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
224 * lockups can have false positives under extreme conditions. So we generally
225 * want a higher threshold for soft lockups than for hard lockups. So we couple
226 * the thresholds with a factor: we make the soft threshold twice the amount of
227 * time the hard threshold is.
229 static int get_softlockup_thresh(void)
231 return watchdog_thresh
* 2;
235 * Returns seconds, approximately. We don't need nanosecond
236 * resolution, and we don't need to waste time with a big divide when
239 static unsigned long get_timestamp(void)
241 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
244 static void set_sample_period(void)
247 * convert watchdog_thresh from seconds to ns
248 * the divide by 5 is to give hrtimer several chances (two
249 * or three with the current relation between the soft
250 * and hard thresholds) to increment before the
251 * hardlockup detector generates a warning
253 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
254 watchdog_update_hrtimer_threshold(sample_period
);
257 /* Commands for resetting the watchdog */
258 static void __touch_watchdog(void)
260 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
264 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
266 * Call when the scheduler may have stalled for legitimate reasons
267 * preventing the watchdog task from executing - e.g. the scheduler
268 * entering idle state. This should only be used for scheduler events.
269 * Use touch_softlockup_watchdog() for everything else.
271 notrace
void touch_softlockup_watchdog_sched(void)
274 * Preemption can be enabled. It doesn't matter which CPU's timestamp
275 * gets zeroed here, so use the raw_ operation.
277 raw_cpu_write(watchdog_touch_ts
, SOFTLOCKUP_RESET
);
280 notrace
void touch_softlockup_watchdog(void)
282 touch_softlockup_watchdog_sched();
283 wq_watchdog_touch(raw_smp_processor_id());
285 EXPORT_SYMBOL(touch_softlockup_watchdog
);
287 void touch_all_softlockup_watchdogs(void)
292 * watchdog_mutex cannpt be taken here, as this might be called
293 * from (soft)interrupt context, so the access to
294 * watchdog_allowed_cpumask might race with a concurrent update.
296 * The watchdog time stamp can race against a concurrent real
297 * update as well, the only side effect might be a cycle delay for
298 * the softlockup check.
300 for_each_cpu(cpu
, &watchdog_allowed_mask
)
301 per_cpu(watchdog_touch_ts
, cpu
) = SOFTLOCKUP_RESET
;
302 wq_watchdog_touch(-1);
305 void touch_softlockup_watchdog_sync(void)
307 __this_cpu_write(softlockup_touch_sync
, true);
308 __this_cpu_write(watchdog_touch_ts
, SOFTLOCKUP_RESET
);
311 static int is_softlockup(unsigned long touch_ts
)
313 unsigned long now
= get_timestamp();
315 if ((watchdog_enabled
& SOFT_WATCHDOG_ENABLED
) && watchdog_thresh
){
316 /* Warn about unreasonable delays. */
317 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
318 return now
- touch_ts
;
323 /* watchdog detector functions */
324 bool is_hardlockup(void)
326 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
328 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
331 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
335 static void watchdog_interrupt_count(void)
337 __this_cpu_inc(hrtimer_interrupts
);
340 static DEFINE_PER_CPU(struct completion
, softlockup_completion
);
341 static DEFINE_PER_CPU(struct cpu_stop_work
, softlockup_stop_work
);
344 * The watchdog thread function - touches the timestamp.
346 * It only runs once every sample_period seconds (4 seconds by
347 * default) to reset the softlockup timestamp. If this gets delayed
348 * for more than 2*watchdog_thresh seconds then the debug-printout
349 * triggers in watchdog_timer_fn().
351 static int softlockup_fn(void *data
)
354 complete(this_cpu_ptr(&softlockup_completion
));
359 /* watchdog kicker functions */
360 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
362 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
363 struct pt_regs
*regs
= get_irq_regs();
365 int softlockup_all_cpu_backtrace
= sysctl_softlockup_all_cpu_backtrace
;
367 if (!watchdog_enabled
)
368 return HRTIMER_NORESTART
;
370 /* kick the hardlockup detector */
371 watchdog_interrupt_count();
373 /* kick the softlockup detector */
374 if (completion_done(this_cpu_ptr(&softlockup_completion
))) {
375 reinit_completion(this_cpu_ptr(&softlockup_completion
));
376 stop_one_cpu_nowait(smp_processor_id(),
378 this_cpu_ptr(&softlockup_stop_work
));
382 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
384 if (touch_ts
== SOFTLOCKUP_RESET
) {
385 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
387 * If the time stamp was touched atomically
388 * make sure the scheduler tick is up to date.
390 __this_cpu_write(softlockup_touch_sync
, false);
394 /* Clear the guest paused flag on watchdog reset */
395 kvm_check_and_clear_guest_paused();
397 return HRTIMER_RESTART
;
400 /* check for a softlockup
401 * This is done by making sure a high priority task is
402 * being scheduled. The task touches the watchdog to
403 * indicate it is getting cpu time. If it hasn't then
404 * this is a good indication some task is hogging the cpu
406 duration
= is_softlockup(touch_ts
);
407 if (unlikely(duration
)) {
409 * If a virtual machine is stopped by the host it can look to
410 * the watchdog like a soft lockup, check to see if the host
411 * stopped the vm before we issue the warning
413 if (kvm_check_and_clear_guest_paused())
414 return HRTIMER_RESTART
;
417 if (__this_cpu_read(soft_watchdog_warn
) == true)
418 return HRTIMER_RESTART
;
420 if (softlockup_all_cpu_backtrace
) {
421 /* Prevent multiple soft-lockup reports if one cpu is already
422 * engaged in dumping cpu back traces
424 if (test_and_set_bit(0, &soft_lockup_nmi_warn
)) {
425 /* Someone else will report us. Let's give up */
426 __this_cpu_write(soft_watchdog_warn
, true);
427 return HRTIMER_RESTART
;
431 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
432 smp_processor_id(), duration
,
433 current
->comm
, task_pid_nr(current
));
435 print_irqtrace_events(current
);
441 if (softlockup_all_cpu_backtrace
) {
442 /* Avoid generating two back traces for current
443 * given that one is already made above
445 trigger_allbutself_cpu_backtrace();
447 clear_bit(0, &soft_lockup_nmi_warn
);
448 /* Barrier to sync with other cpus */
449 smp_mb__after_atomic();
452 add_taint(TAINT_SOFTLOCKUP
, LOCKDEP_STILL_OK
);
453 if (softlockup_panic
)
454 panic("softlockup: hung tasks");
455 __this_cpu_write(soft_watchdog_warn
, true);
457 __this_cpu_write(soft_watchdog_warn
, false);
459 return HRTIMER_RESTART
;
462 static void watchdog_enable(unsigned int cpu
)
464 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
465 struct completion
*done
= this_cpu_ptr(&softlockup_completion
);
467 WARN_ON_ONCE(cpu
!= smp_processor_id());
469 init_completion(done
);
473 * Start the timer first to prevent the NMI watchdog triggering
474 * before the timer has a chance to fire.
476 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
477 hrtimer
->function
= watchdog_timer_fn
;
478 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
479 HRTIMER_MODE_REL_PINNED_HARD
);
481 /* Initialize timestamp */
483 /* Enable the perf event */
484 if (watchdog_enabled
& NMI_WATCHDOG_ENABLED
)
485 watchdog_nmi_enable(cpu
);
488 static void watchdog_disable(unsigned int cpu
)
490 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
492 WARN_ON_ONCE(cpu
!= smp_processor_id());
495 * Disable the perf event first. That prevents that a large delay
496 * between disabling the timer and disabling the perf event causes
497 * the perf NMI to detect a false positive.
499 watchdog_nmi_disable(cpu
);
500 hrtimer_cancel(hrtimer
);
501 wait_for_completion(this_cpu_ptr(&softlockup_completion
));
504 static int softlockup_stop_fn(void *data
)
506 watchdog_disable(smp_processor_id());
510 static void softlockup_stop_all(void)
514 if (!softlockup_initialized
)
517 for_each_cpu(cpu
, &watchdog_allowed_mask
)
518 smp_call_on_cpu(cpu
, softlockup_stop_fn
, NULL
, false);
520 cpumask_clear(&watchdog_allowed_mask
);
523 static int softlockup_start_fn(void *data
)
525 watchdog_enable(smp_processor_id());
529 static void softlockup_start_all(void)
533 cpumask_copy(&watchdog_allowed_mask
, &watchdog_cpumask
);
534 for_each_cpu(cpu
, &watchdog_allowed_mask
)
535 smp_call_on_cpu(cpu
, softlockup_start_fn
, NULL
, false);
538 int lockup_detector_online_cpu(unsigned int cpu
)
540 if (cpumask_test_cpu(cpu
, &watchdog_allowed_mask
))
541 watchdog_enable(cpu
);
545 int lockup_detector_offline_cpu(unsigned int cpu
)
547 if (cpumask_test_cpu(cpu
, &watchdog_allowed_mask
))
548 watchdog_disable(cpu
);
552 static void lockup_detector_reconfigure(void)
557 softlockup_stop_all();
559 lockup_detector_update_enable();
560 if (watchdog_enabled
&& watchdog_thresh
)
561 softlockup_start_all();
563 watchdog_nmi_start();
566 * Must be called outside the cpus locked section to prevent
567 * recursive locking in the perf code.
569 __lockup_detector_cleanup();
573 * Create the watchdog thread infrastructure and configure the detector(s).
575 * The threads are not unparked as watchdog_allowed_mask is empty. When
576 * the threads are successfully initialized, take the proper locks and
577 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
579 static __init
void lockup_detector_setup(void)
582 * If sysctl is off and watchdog got disabled on the command line,
583 * nothing to do here.
585 lockup_detector_update_enable();
587 if (!IS_ENABLED(CONFIG_SYSCTL
) &&
588 !(watchdog_enabled
&& watchdog_thresh
))
591 mutex_lock(&watchdog_mutex
);
592 lockup_detector_reconfigure();
593 softlockup_initialized
= true;
594 mutex_unlock(&watchdog_mutex
);
597 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
598 static void lockup_detector_reconfigure(void)
602 lockup_detector_update_enable();
603 watchdog_nmi_start();
606 static inline void lockup_detector_setup(void)
608 lockup_detector_reconfigure();
610 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
612 static void __lockup_detector_cleanup(void)
614 lockdep_assert_held(&watchdog_mutex
);
615 hardlockup_detector_perf_cleanup();
619 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
621 * Caller must not hold the cpu hotplug rwsem.
623 void lockup_detector_cleanup(void)
625 mutex_lock(&watchdog_mutex
);
626 __lockup_detector_cleanup();
627 mutex_unlock(&watchdog_mutex
);
631 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
633 * Special interface for parisc. It prevents lockup detector warnings from
634 * the default pm_poweroff() function which busy loops forever.
636 void lockup_detector_soft_poweroff(void)
638 watchdog_enabled
= 0;
643 /* Propagate any changes to the watchdog threads */
644 static void proc_watchdog_update(void)
646 /* Remove impossible cpus to keep sysctl output clean. */
647 cpumask_and(&watchdog_cpumask
, &watchdog_cpumask
, cpu_possible_mask
);
648 lockup_detector_reconfigure();
652 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
654 * caller | table->data points to | 'which'
655 * -------------------|----------------------------|--------------------------
656 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
657 * | | SOFT_WATCHDOG_ENABLED
658 * -------------------|----------------------------|--------------------------
659 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
660 * -------------------|----------------------------|--------------------------
661 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
663 static int proc_watchdog_common(int which
, struct ctl_table
*table
, int write
,
664 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
666 int err
, old
, *param
= table
->data
;
668 mutex_lock(&watchdog_mutex
);
672 * On read synchronize the userspace interface. This is a
675 *param
= (watchdog_enabled
& which
) != 0;
676 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
678 old
= READ_ONCE(*param
);
679 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
680 if (!err
&& old
!= READ_ONCE(*param
))
681 proc_watchdog_update();
683 mutex_unlock(&watchdog_mutex
);
688 * /proc/sys/kernel/watchdog
690 int proc_watchdog(struct ctl_table
*table
, int write
,
691 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
693 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
|SOFT_WATCHDOG_ENABLED
,
694 table
, write
, buffer
, lenp
, ppos
);
698 * /proc/sys/kernel/nmi_watchdog
700 int proc_nmi_watchdog(struct ctl_table
*table
, int write
,
701 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
703 if (!nmi_watchdog_available
&& write
)
705 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
,
706 table
, write
, buffer
, lenp
, ppos
);
710 * /proc/sys/kernel/soft_watchdog
712 int proc_soft_watchdog(struct ctl_table
*table
, int write
,
713 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
715 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED
,
716 table
, write
, buffer
, lenp
, ppos
);
720 * /proc/sys/kernel/watchdog_thresh
722 int proc_watchdog_thresh(struct ctl_table
*table
, int write
,
723 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
727 mutex_lock(&watchdog_mutex
);
729 old
= READ_ONCE(watchdog_thresh
);
730 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
732 if (!err
&& write
&& old
!= READ_ONCE(watchdog_thresh
))
733 proc_watchdog_update();
735 mutex_unlock(&watchdog_mutex
);
740 * The cpumask is the mask of possible cpus that the watchdog can run
741 * on, not the mask of cpus it is actually running on. This allows the
742 * user to specify a mask that will include cpus that have not yet
743 * been brought online, if desired.
745 int proc_watchdog_cpumask(struct ctl_table
*table
, int write
,
746 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
750 mutex_lock(&watchdog_mutex
);
752 err
= proc_do_large_bitmap(table
, write
, buffer
, lenp
, ppos
);
754 proc_watchdog_update();
756 mutex_unlock(&watchdog_mutex
);
759 #endif /* CONFIG_SYSCTL */
761 void __init
lockup_detector_init(void)
763 if (tick_nohz_full_enabled())
764 pr_info("Disabling watchdog on nohz_full cores by default\n");
766 cpumask_copy(&watchdog_cpumask
,
767 housekeeping_cpumask(HK_FLAG_TIMER
));
769 if (!watchdog_nmi_probe())
770 nmi_watchdog_available
= true;
771 lockup_detector_setup();