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>
22 #include <linux/tick.h>
23 #include <linux/workqueue.h>
25 #include <asm/irq_regs.h>
26 #include <linux/kvm_para.h>
27 #include <linux/perf_event.h>
28 #include <linux/kthread.h>
31 * The run state of the lockup detectors is controlled by the content of the
32 * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
33 * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
35 * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
36 * are variables that are only used as an 'interface' between the parameters
37 * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
38 * 'watchdog_thresh' variable is handled differently because its value is not
39 * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
42 #define NMI_WATCHDOG_ENABLED_BIT 0
43 #define SOFT_WATCHDOG_ENABLED_BIT 1
44 #define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT)
45 #define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT)
47 static DEFINE_MUTEX(watchdog_proc_mutex
);
49 #ifdef CONFIG_HARDLOCKUP_DETECTOR
50 static unsigned long __read_mostly watchdog_enabled
= SOFT_WATCHDOG_ENABLED
|NMI_WATCHDOG_ENABLED
;
52 static unsigned long __read_mostly watchdog_enabled
= SOFT_WATCHDOG_ENABLED
;
54 int __read_mostly nmi_watchdog_enabled
;
55 int __read_mostly soft_watchdog_enabled
;
56 int __read_mostly watchdog_user_enabled
;
57 int __read_mostly watchdog_thresh
= 10;
60 int __read_mostly sysctl_softlockup_all_cpu_backtrace
;
61 int __read_mostly sysctl_hardlockup_all_cpu_backtrace
;
63 #define sysctl_softlockup_all_cpu_backtrace 0
64 #define sysctl_hardlockup_all_cpu_backtrace 0
66 static struct cpumask watchdog_cpumask __read_mostly
;
67 unsigned long *watchdog_cpumask_bits
= cpumask_bits(&watchdog_cpumask
);
69 /* Helper for online, unparked cpus. */
70 #define for_each_watchdog_cpu(cpu) \
71 for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
74 * The 'watchdog_running' variable is set to 1 when the watchdog threads
75 * are registered/started and is set to 0 when the watchdog threads are
76 * unregistered/stopped, so it is an indicator whether the threads exist.
78 static int __read_mostly watchdog_running
;
80 * If a subsystem has a need to deactivate the watchdog temporarily, it
81 * can use the suspend/resume interface to achieve this. The content of
82 * the 'watchdog_suspended' variable reflects this state. Existing threads
83 * are parked/unparked by the lockup_detector_{suspend|resume} functions
84 * (see comment blocks pertaining to those functions for further details).
86 * 'watchdog_suspended' also prevents threads from being registered/started
87 * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
88 * of 'watchdog_running' cannot change while the watchdog is deactivated
89 * temporarily (see related code in 'proc' handlers).
91 static int __read_mostly watchdog_suspended
;
93 static u64 __read_mostly sample_period
;
95 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
96 static DEFINE_PER_CPU(struct task_struct
*, softlockup_watchdog
);
97 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
98 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
99 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
100 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
101 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt
);
102 static DEFINE_PER_CPU(struct task_struct
*, softlockup_task_ptr_saved
);
103 #ifdef CONFIG_HARDLOCKUP_DETECTOR
104 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
105 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
106 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
107 static DEFINE_PER_CPU(struct perf_event
*, watchdog_ev
);
109 static unsigned long soft_lockup_nmi_warn
;
113 * Should we panic when a soft-lockup or hard-lockup occurs:
115 #ifdef CONFIG_HARDLOCKUP_DETECTOR
116 unsigned int __read_mostly hardlockup_panic
=
117 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
118 static unsigned long hardlockup_allcpu_dumped
;
120 * We may not want to enable hard lockup detection by default in all cases,
121 * for example when running the kernel as a guest on a hypervisor. In these
122 * cases this function can be called to disable hard lockup detection. This
123 * function should only be executed once by the boot processor before the
124 * kernel command line parameters are parsed, because otherwise it is not
125 * possible to override this in hardlockup_panic_setup().
127 void hardlockup_detector_disable(void)
129 watchdog_enabled
&= ~NMI_WATCHDOG_ENABLED
;
132 static int __init
hardlockup_panic_setup(char *str
)
134 if (!strncmp(str
, "panic", 5))
135 hardlockup_panic
= 1;
136 else if (!strncmp(str
, "nopanic", 7))
137 hardlockup_panic
= 0;
138 else if (!strncmp(str
, "0", 1))
139 watchdog_enabled
&= ~NMI_WATCHDOG_ENABLED
;
140 else if (!strncmp(str
, "1", 1))
141 watchdog_enabled
|= NMI_WATCHDOG_ENABLED
;
144 __setup("nmi_watchdog=", hardlockup_panic_setup
);
147 unsigned int __read_mostly softlockup_panic
=
148 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE
;
150 static int __init
softlockup_panic_setup(char *str
)
152 softlockup_panic
= simple_strtoul(str
, NULL
, 0);
156 __setup("softlockup_panic=", softlockup_panic_setup
);
158 static int __init
nowatchdog_setup(char *str
)
160 watchdog_enabled
= 0;
163 __setup("nowatchdog", nowatchdog_setup
);
165 static int __init
nosoftlockup_setup(char *str
)
167 watchdog_enabled
&= ~SOFT_WATCHDOG_ENABLED
;
170 __setup("nosoftlockup", nosoftlockup_setup
);
173 static int __init
softlockup_all_cpu_backtrace_setup(char *str
)
175 sysctl_softlockup_all_cpu_backtrace
=
176 !!simple_strtol(str
, NULL
, 0);
179 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup
);
180 static int __init
hardlockup_all_cpu_backtrace_setup(char *str
)
182 sysctl_hardlockup_all_cpu_backtrace
=
183 !!simple_strtol(str
, NULL
, 0);
186 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup
);
190 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
191 * lockups can have false positives under extreme conditions. So we generally
192 * want a higher threshold for soft lockups than for hard lockups. So we couple
193 * the thresholds with a factor: we make the soft threshold twice the amount of
194 * time the hard threshold is.
196 static int get_softlockup_thresh(void)
198 return watchdog_thresh
* 2;
202 * Returns seconds, approximately. We don't need nanosecond
203 * resolution, and we don't need to waste time with a big divide when
206 static unsigned long get_timestamp(void)
208 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
211 static void set_sample_period(void)
214 * convert watchdog_thresh from seconds to ns
215 * the divide by 5 is to give hrtimer several chances (two
216 * or three with the current relation between the soft
217 * and hard thresholds) to increment before the
218 * hardlockup detector generates a warning
220 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
223 /* Commands for resetting the watchdog */
224 static void __touch_watchdog(void)
226 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
230 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
232 * Call when the scheduler may have stalled for legitimate reasons
233 * preventing the watchdog task from executing - e.g. the scheduler
234 * entering idle state. This should only be used for scheduler events.
235 * Use touch_softlockup_watchdog() for everything else.
237 void touch_softlockup_watchdog_sched(void)
240 * Preemption can be enabled. It doesn't matter which CPU's timestamp
241 * gets zeroed here, so use the raw_ operation.
243 raw_cpu_write(watchdog_touch_ts
, 0);
246 void touch_softlockup_watchdog(void)
248 touch_softlockup_watchdog_sched();
249 wq_watchdog_touch(raw_smp_processor_id());
251 EXPORT_SYMBOL(touch_softlockup_watchdog
);
253 void touch_all_softlockup_watchdogs(void)
258 * this is done lockless
259 * do we care if a 0 races with a timestamp?
260 * all it means is the softlock check starts one cycle later
262 for_each_watchdog_cpu(cpu
)
263 per_cpu(watchdog_touch_ts
, cpu
) = 0;
264 wq_watchdog_touch(-1);
267 #ifdef CONFIG_HARDLOCKUP_DETECTOR
268 void touch_nmi_watchdog(void)
271 * Using __raw here because some code paths have
272 * preemption enabled. If preemption is enabled
273 * then interrupts should be enabled too, in which
274 * case we shouldn't have to worry about the watchdog
277 raw_cpu_write(watchdog_nmi_touch
, true);
278 touch_softlockup_watchdog();
280 EXPORT_SYMBOL(touch_nmi_watchdog
);
284 void touch_softlockup_watchdog_sync(void)
286 __this_cpu_write(softlockup_touch_sync
, true);
287 __this_cpu_write(watchdog_touch_ts
, 0);
290 #ifdef CONFIG_HARDLOCKUP_DETECTOR
291 /* watchdog detector functions */
292 static bool is_hardlockup(void)
294 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
296 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
299 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
304 static int is_softlockup(unsigned long touch_ts
)
306 unsigned long now
= get_timestamp();
308 if ((watchdog_enabled
& SOFT_WATCHDOG_ENABLED
) && watchdog_thresh
){
309 /* Warn about unreasonable delays. */
310 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
311 return now
- touch_ts
;
316 #ifdef CONFIG_HARDLOCKUP_DETECTOR
318 /* Can be overriden by architecture */
319 __weak
int hw_nmi_get_event(void)
321 return PERF_COUNT_HW_CPU_CYCLES
;
324 static struct perf_event_attr wd_hw_attr
= {
325 .type
= PERF_TYPE_HARDWARE
,
326 .config
= PERF_COUNT_HW_CPU_CYCLES
,
327 .size
= sizeof(struct perf_event_attr
),
332 /* Callback function for perf event subsystem */
333 static void watchdog_overflow_callback(struct perf_event
*event
,
334 struct perf_sample_data
*data
,
335 struct pt_regs
*regs
)
337 /* Ensure the watchdog never gets throttled */
338 event
->hw
.interrupts
= 0;
340 if (__this_cpu_read(watchdog_nmi_touch
) == true) {
341 __this_cpu_write(watchdog_nmi_touch
, false);
345 /* check for a hardlockup
346 * This is done by making sure our timer interrupt
347 * is incrementing. The timer interrupt should have
348 * fired multiple times before we overflow'd. If it hasn't
349 * then this is a good indication the cpu is stuck
351 if (is_hardlockup()) {
352 int this_cpu
= smp_processor_id();
353 struct pt_regs
*regs
= get_irq_regs();
355 /* only print hardlockups once */
356 if (__this_cpu_read(hard_watchdog_warn
) == true)
359 pr_emerg("Watchdog detected hard LOCKUP on cpu %d", this_cpu
);
361 print_irqtrace_events(current
);
368 * Perform all-CPU dump only once to avoid multiple hardlockups
369 * generating interleaving traces
371 if (sysctl_hardlockup_all_cpu_backtrace
&&
372 !test_and_set_bit(0, &hardlockup_allcpu_dumped
))
373 trigger_allbutself_cpu_backtrace();
375 if (hardlockup_panic
)
376 nmi_panic(regs
, "Hard LOCKUP");
378 __this_cpu_write(hard_watchdog_warn
, true);
382 __this_cpu_write(hard_watchdog_warn
, false);
385 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
387 static void watchdog_interrupt_count(void)
389 __this_cpu_inc(hrtimer_interrupts
);
392 static int watchdog_nmi_enable(unsigned int cpu
);
393 static void watchdog_nmi_disable(unsigned int cpu
);
395 static int watchdog_enable_all_cpus(void);
396 static void watchdog_disable_all_cpus(void);
398 /* watchdog kicker functions */
399 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
401 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
402 struct pt_regs
*regs
= get_irq_regs();
404 int softlockup_all_cpu_backtrace
= sysctl_softlockup_all_cpu_backtrace
;
406 /* kick the hardlockup detector */
407 watchdog_interrupt_count();
409 /* kick the softlockup detector */
410 wake_up_process(__this_cpu_read(softlockup_watchdog
));
413 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
416 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
418 * If the time stamp was touched atomically
419 * make sure the scheduler tick is up to date.
421 __this_cpu_write(softlockup_touch_sync
, false);
425 /* Clear the guest paused flag on watchdog reset */
426 kvm_check_and_clear_guest_paused();
428 return HRTIMER_RESTART
;
431 /* check for a softlockup
432 * This is done by making sure a high priority task is
433 * being scheduled. The task touches the watchdog to
434 * indicate it is getting cpu time. If it hasn't then
435 * this is a good indication some task is hogging the cpu
437 duration
= is_softlockup(touch_ts
);
438 if (unlikely(duration
)) {
440 * If a virtual machine is stopped by the host it can look to
441 * the watchdog like a soft lockup, check to see if the host
442 * stopped the vm before we issue the warning
444 if (kvm_check_and_clear_guest_paused())
445 return HRTIMER_RESTART
;
448 if (__this_cpu_read(soft_watchdog_warn
) == true) {
450 * When multiple processes are causing softlockups the
451 * softlockup detector only warns on the first one
452 * because the code relies on a full quiet cycle to
453 * re-arm. The second process prevents the quiet cycle
454 * and never gets reported. Use task pointers to detect
457 if (__this_cpu_read(softlockup_task_ptr_saved
) !=
459 __this_cpu_write(soft_watchdog_warn
, false);
462 return HRTIMER_RESTART
;
465 if (softlockup_all_cpu_backtrace
) {
466 /* Prevent multiple soft-lockup reports if one cpu is already
467 * engaged in dumping cpu back traces
469 if (test_and_set_bit(0, &soft_lockup_nmi_warn
)) {
470 /* Someone else will report us. Let's give up */
471 __this_cpu_write(soft_watchdog_warn
, true);
472 return HRTIMER_RESTART
;
476 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
477 smp_processor_id(), duration
,
478 current
->comm
, task_pid_nr(current
));
479 __this_cpu_write(softlockup_task_ptr_saved
, current
);
481 print_irqtrace_events(current
);
487 if (softlockup_all_cpu_backtrace
) {
488 /* Avoid generating two back traces for current
489 * given that one is already made above
491 trigger_allbutself_cpu_backtrace();
493 clear_bit(0, &soft_lockup_nmi_warn
);
494 /* Barrier to sync with other cpus */
495 smp_mb__after_atomic();
498 add_taint(TAINT_SOFTLOCKUP
, LOCKDEP_STILL_OK
);
499 if (softlockup_panic
)
500 panic("softlockup: hung tasks");
501 __this_cpu_write(soft_watchdog_warn
, true);
503 __this_cpu_write(soft_watchdog_warn
, false);
505 return HRTIMER_RESTART
;
508 static void watchdog_set_prio(unsigned int policy
, unsigned int prio
)
510 struct sched_param param
= { .sched_priority
= prio
};
512 sched_setscheduler(current
, policy
, ¶m
);
515 static void watchdog_enable(unsigned int cpu
)
517 struct hrtimer
*hrtimer
= raw_cpu_ptr(&watchdog_hrtimer
);
519 /* kick off the timer for the hardlockup detector */
520 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
521 hrtimer
->function
= watchdog_timer_fn
;
523 /* Enable the perf event */
524 watchdog_nmi_enable(cpu
);
526 /* done here because hrtimer_start can only pin to smp_processor_id() */
527 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
528 HRTIMER_MODE_REL_PINNED
);
530 /* initialize timestamp */
531 watchdog_set_prio(SCHED_FIFO
, MAX_RT_PRIO
- 1);
535 static void watchdog_disable(unsigned int cpu
)
537 struct hrtimer
*hrtimer
= raw_cpu_ptr(&watchdog_hrtimer
);
539 watchdog_set_prio(SCHED_NORMAL
, 0);
540 hrtimer_cancel(hrtimer
);
541 /* disable the perf event */
542 watchdog_nmi_disable(cpu
);
545 static void watchdog_cleanup(unsigned int cpu
, bool online
)
547 watchdog_disable(cpu
);
550 static int watchdog_should_run(unsigned int cpu
)
552 return __this_cpu_read(hrtimer_interrupts
) !=
553 __this_cpu_read(soft_lockup_hrtimer_cnt
);
557 * The watchdog thread function - touches the timestamp.
559 * It only runs once every sample_period seconds (4 seconds by
560 * default) to reset the softlockup timestamp. If this gets delayed
561 * for more than 2*watchdog_thresh seconds then the debug-printout
562 * triggers in watchdog_timer_fn().
564 static void watchdog(unsigned int cpu
)
566 __this_cpu_write(soft_lockup_hrtimer_cnt
,
567 __this_cpu_read(hrtimer_interrupts
));
571 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
572 * failure path. Check for failures that can occur asynchronously -
573 * for example, when CPUs are on-lined - and shut down the hardware
574 * perf event on each CPU accordingly.
576 * The only non-obvious place this bit can be cleared is through
577 * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a
578 * pr_info here would be too noisy as it would result in a message
579 * every few seconds if the hardlockup was disabled but the softlockup
582 if (!(watchdog_enabled
& NMI_WATCHDOG_ENABLED
))
583 watchdog_nmi_disable(cpu
);
586 #ifdef CONFIG_HARDLOCKUP_DETECTOR
588 * People like the simple clean cpu node info on boot.
589 * Reduce the watchdog noise by only printing messages
590 * that are different from what cpu0 displayed.
592 static unsigned long cpu0_err
;
594 static int watchdog_nmi_enable(unsigned int cpu
)
596 struct perf_event_attr
*wd_attr
;
597 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
599 /* nothing to do if the hard lockup detector is disabled */
600 if (!(watchdog_enabled
& NMI_WATCHDOG_ENABLED
))
603 /* is it already setup and enabled? */
604 if (event
&& event
->state
> PERF_EVENT_STATE_OFF
)
607 /* it is setup but not enabled */
611 wd_attr
= &wd_hw_attr
;
612 wd_attr
->sample_period
= hw_nmi_get_sample_period(watchdog_thresh
);
613 wd_attr
->config
= hw_nmi_get_event();
615 /* Try to register using hardware perf events */
616 event
= perf_event_create_kernel_counter(wd_attr
, cpu
, NULL
, watchdog_overflow_callback
, NULL
);
618 /* save cpu0 error for future comparision */
619 if (cpu
== 0 && IS_ERR(event
))
620 cpu0_err
= PTR_ERR(event
);
622 if (!IS_ERR(event
)) {
623 /* only print for cpu0 or different than cpu0 */
624 if (cpu
== 0 || cpu0_err
)
625 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
630 * Disable the hard lockup detector if _any_ CPU fails to set up
631 * set up the hardware perf event. The watchdog() function checks
632 * the NMI_WATCHDOG_ENABLED bit periodically.
634 * The barriers are for syncing up watchdog_enabled across all the
635 * cpus, as clear_bit() does not use barriers.
637 smp_mb__before_atomic();
638 clear_bit(NMI_WATCHDOG_ENABLED_BIT
, &watchdog_enabled
);
639 smp_mb__after_atomic();
641 /* skip displaying the same error again */
642 if (cpu
> 0 && (PTR_ERR(event
) == cpu0_err
))
643 return PTR_ERR(event
);
645 /* vary the KERN level based on the returned errno */
646 if (PTR_ERR(event
) == -EOPNOTSUPP
)
647 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu
);
648 else if (PTR_ERR(event
) == -ENOENT
)
649 pr_warn("disabled (cpu%i): hardware events not enabled\n",
652 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
653 cpu
, PTR_ERR(event
));
655 pr_info("Shutting down hard lockup detector on all cpus\n");
657 return PTR_ERR(event
);
661 per_cpu(watchdog_ev
, cpu
) = event
;
663 perf_event_enable(per_cpu(watchdog_ev
, cpu
));
668 static void watchdog_nmi_disable(unsigned int cpu
)
670 struct perf_event
*event
= per_cpu(watchdog_ev
, cpu
);
673 perf_event_disable(event
);
674 per_cpu(watchdog_ev
, cpu
) = NULL
;
676 /* should be in cleanup, but blocks oprofile */
677 perf_event_release_kernel(event
);
680 /* watchdog_nmi_enable() expects this to be zero initially. */
686 static int watchdog_nmi_enable(unsigned int cpu
) { return 0; }
687 static void watchdog_nmi_disable(unsigned int cpu
) { return; }
688 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
690 static struct smp_hotplug_thread watchdog_threads
= {
691 .store
= &softlockup_watchdog
,
692 .thread_should_run
= watchdog_should_run
,
693 .thread_fn
= watchdog
,
694 .thread_comm
= "watchdog/%u",
695 .setup
= watchdog_enable
,
696 .cleanup
= watchdog_cleanup
,
697 .park
= watchdog_disable
,
698 .unpark
= watchdog_enable
,
702 * park all watchdog threads that are specified in 'watchdog_cpumask'
704 * This function returns an error if kthread_park() of a watchdog thread
705 * fails. In this situation, the watchdog threads of some CPUs can already
706 * be parked and the watchdog threads of other CPUs can still be runnable.
707 * Callers are expected to handle this special condition as appropriate in
710 * This function may only be called in a context that is protected against
711 * races with CPU hotplug - for example, via get_online_cpus().
713 static int watchdog_park_threads(void)
717 for_each_watchdog_cpu(cpu
) {
718 ret
= kthread_park(per_cpu(softlockup_watchdog
, cpu
));
727 * unpark all watchdog threads that are specified in 'watchdog_cpumask'
729 * This function may only be called in a context that is protected against
730 * races with CPU hotplug - for example, via get_online_cpus().
732 static void watchdog_unpark_threads(void)
736 for_each_watchdog_cpu(cpu
)
737 kthread_unpark(per_cpu(softlockup_watchdog
, cpu
));
741 * Suspend the hard and soft lockup detector by parking the watchdog threads.
743 int lockup_detector_suspend(void)
748 mutex_lock(&watchdog_proc_mutex
);
750 * Multiple suspend requests can be active in parallel (counted by
751 * the 'watchdog_suspended' variable). If the watchdog threads are
752 * running, the first caller takes care that they will be parked.
753 * The state of 'watchdog_running' cannot change while a suspend
754 * request is active (see related code in 'proc' handlers).
756 if (watchdog_running
&& !watchdog_suspended
)
757 ret
= watchdog_park_threads();
760 watchdog_suspended
++;
762 watchdog_disable_all_cpus();
763 pr_err("Failed to suspend lockup detectors, disabled\n");
764 watchdog_enabled
= 0;
767 mutex_unlock(&watchdog_proc_mutex
);
773 * Resume the hard and soft lockup detector by unparking the watchdog threads.
775 void lockup_detector_resume(void)
777 mutex_lock(&watchdog_proc_mutex
);
779 watchdog_suspended
--;
781 * The watchdog threads are unparked if they were previously running
782 * and if there is no more active suspend request.
784 if (watchdog_running
&& !watchdog_suspended
)
785 watchdog_unpark_threads();
787 mutex_unlock(&watchdog_proc_mutex
);
791 static int update_watchdog_all_cpus(void)
795 ret
= watchdog_park_threads();
799 watchdog_unpark_threads();
804 static int watchdog_enable_all_cpus(void)
808 if (!watchdog_running
) {
809 err
= smpboot_register_percpu_thread_cpumask(&watchdog_threads
,
812 pr_err("Failed to create watchdog threads, disabled\n");
814 watchdog_running
= 1;
817 * Enable/disable the lockup detectors or
818 * change the sample period 'on the fly'.
820 err
= update_watchdog_all_cpus();
823 watchdog_disable_all_cpus();
824 pr_err("Failed to update lockup detectors, disabled\n");
829 watchdog_enabled
= 0;
834 static void watchdog_disable_all_cpus(void)
836 if (watchdog_running
) {
837 watchdog_running
= 0;
838 smpboot_unregister_percpu_thread(&watchdog_threads
);
845 * Update the run state of the lockup detectors.
847 static int proc_watchdog_update(void)
852 * Watchdog threads won't be started if they are already active.
853 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
854 * care of this. If those threads are already active, the sample
855 * period will be updated and the lockup detectors will be enabled
856 * or disabled 'on the fly'.
858 if (watchdog_enabled
&& watchdog_thresh
)
859 err
= watchdog_enable_all_cpus();
861 watchdog_disable_all_cpus();
868 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
870 * caller | table->data points to | 'which' contains the flag(s)
871 * -------------------|-----------------------|-----------------------------
872 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
873 * | | with SOFT_WATCHDOG_ENABLED
874 * -------------------|-----------------------|-----------------------------
875 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
876 * -------------------|-----------------------|-----------------------------
877 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
879 static int proc_watchdog_common(int which
, struct ctl_table
*table
, int write
,
880 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
883 int *watchdog_param
= (int *)table
->data
;
886 mutex_lock(&watchdog_proc_mutex
);
888 if (watchdog_suspended
) {
889 /* no parameter changes allowed while watchdog is suspended */
895 * If the parameter is being read return the state of the corresponding
896 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
897 * run state of the lockup detectors.
900 *watchdog_param
= (watchdog_enabled
& which
) != 0;
901 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
903 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
908 * There is a race window between fetching the current value
909 * from 'watchdog_enabled' and storing the new value. During
910 * this race window, watchdog_nmi_enable() can sneak in and
911 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
912 * The 'cmpxchg' detects this race and the loop retries.
915 old
= watchdog_enabled
;
917 * If the parameter value is not zero set the
918 * corresponding bit(s), else clear it(them).
924 } while (cmpxchg(&watchdog_enabled
, old
, new) != old
);
927 * Update the run state of the lockup detectors. There is _no_
928 * need to check the value returned by proc_watchdog_update()
929 * and to restore the previous value of 'watchdog_enabled' as
930 * both lockup detectors are disabled if proc_watchdog_update()
936 err
= proc_watchdog_update();
939 mutex_unlock(&watchdog_proc_mutex
);
945 * /proc/sys/kernel/watchdog
947 int proc_watchdog(struct ctl_table
*table
, int write
,
948 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
950 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
|SOFT_WATCHDOG_ENABLED
,
951 table
, write
, buffer
, lenp
, ppos
);
955 * /proc/sys/kernel/nmi_watchdog
957 int proc_nmi_watchdog(struct ctl_table
*table
, int write
,
958 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
960 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
,
961 table
, write
, buffer
, lenp
, ppos
);
965 * /proc/sys/kernel/soft_watchdog
967 int proc_soft_watchdog(struct ctl_table
*table
, int write
,
968 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
970 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED
,
971 table
, write
, buffer
, lenp
, ppos
);
975 * /proc/sys/kernel/watchdog_thresh
977 int proc_watchdog_thresh(struct ctl_table
*table
, int write
,
978 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
983 mutex_lock(&watchdog_proc_mutex
);
985 if (watchdog_suspended
) {
986 /* no parameter changes allowed while watchdog is suspended */
991 old
= ACCESS_ONCE(watchdog_thresh
);
992 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
998 * Update the sample period. Restore on failure.
1000 new = ACCESS_ONCE(watchdog_thresh
);
1004 set_sample_period();
1005 err
= proc_watchdog_update();
1007 watchdog_thresh
= old
;
1008 set_sample_period();
1011 mutex_unlock(&watchdog_proc_mutex
);
1017 * The cpumask is the mask of possible cpus that the watchdog can run
1018 * on, not the mask of cpus it is actually running on. This allows the
1019 * user to specify a mask that will include cpus that have not yet
1020 * been brought online, if desired.
1022 int proc_watchdog_cpumask(struct ctl_table
*table
, int write
,
1023 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
1028 mutex_lock(&watchdog_proc_mutex
);
1030 if (watchdog_suspended
) {
1031 /* no parameter changes allowed while watchdog is suspended */
1036 err
= proc_do_large_bitmap(table
, write
, buffer
, lenp
, ppos
);
1037 if (!err
&& write
) {
1038 /* Remove impossible cpus to keep sysctl output cleaner. */
1039 cpumask_and(&watchdog_cpumask
, &watchdog_cpumask
,
1042 if (watchdog_running
) {
1044 * Failure would be due to being unable to allocate
1045 * a temporary cpumask, so we are likely not in a
1046 * position to do much else to make things better.
1048 if (smpboot_update_cpumask_percpu_thread(
1049 &watchdog_threads
, &watchdog_cpumask
) != 0)
1050 pr_err("cpumask update failed\n");
1054 mutex_unlock(&watchdog_proc_mutex
);
1059 #endif /* CONFIG_SYSCTL */
1061 void __init
lockup_detector_init(void)
1063 set_sample_period();
1065 #ifdef CONFIG_NO_HZ_FULL
1066 if (tick_nohz_full_enabled()) {
1067 pr_info("Disabling watchdog on nohz_full cores by default\n");
1068 cpumask_copy(&watchdog_cpumask
, housekeeping_mask
);
1070 cpumask_copy(&watchdog_cpumask
, cpu_possible_mask
);
1072 cpumask_copy(&watchdog_cpumask
, cpu_possible_mask
);
1075 if (watchdog_enabled
)
1076 watchdog_enable_all_cpus();