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/smpboot.h>
22 #include <linux/sched/rt.h>
23 #include <uapi/linux/sched/types.h>
24 #include <linux/tick.h>
25 #include <linux/workqueue.h>
26 #include <linux/sched/clock.h>
27 #include <linux/sched/debug.h>
29 #include <asm/irq_regs.h>
30 #include <linux/kvm_para.h>
31 #include <linux/kthread.h>
33 static DEFINE_MUTEX(watchdog_mutex
);
35 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) \
36 || defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU)
37 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
38 # define NMI_WATCHDOG_DEFAULT 1
40 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
41 # define NMI_WATCHDOG_DEFAULT 0
44 unsigned long __read_mostly watchdog_enabled
;
45 int __read_mostly watchdog_user_enabled
= 1;
46 int __read_mostly nmi_watchdog_user_enabled
= NMI_WATCHDOG_DEFAULT
;
47 int __read_mostly soft_watchdog_user_enabled
= 1;
48 int __read_mostly watchdog_thresh
= 10;
49 int __read_mostly nmi_watchdog_available
;
50 #if defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU)
51 int __read_mostly watchdog_other_cpu_available
= WATCHDOG_DEFAULT
;
54 struct cpumask watchdog_allowed_mask __read_mostly
;
56 struct cpumask watchdog_cpumask __read_mostly
;
57 unsigned long *watchdog_cpumask_bits
= cpumask_bits(&watchdog_cpumask
);
59 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU)
61 * Should we panic when a soft-lockup or hard-lockup occurs:
63 unsigned int __read_mostly hardlockup_panic
=
64 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
66 * We may not want to enable hard lockup detection by default in all cases,
67 * for example when running the kernel as a guest on a hypervisor. In these
68 * cases this function can be called to disable hard lockup detection. This
69 * function should only be executed once by the boot processor before the
70 * kernel command line parameters are parsed, because otherwise it is not
71 * possible to override this in hardlockup_panic_setup().
73 void __init
hardlockup_detector_disable(void)
75 nmi_watchdog_user_enabled
= 0;
78 static int __init
hardlockup_panic_setup(char *str
)
80 if (!strncmp(str
, "panic", 5))
82 else if (!strncmp(str
, "nopanic", 7))
84 else if (!strncmp(str
, "0", 1))
85 nmi_watchdog_user_enabled
= 0;
86 else if (!strncmp(str
, "1", 1))
87 nmi_watchdog_user_enabled
= 1;
90 __setup("nmi_watchdog=", hardlockup_panic_setup
);
93 int __read_mostly sysctl_hardlockup_all_cpu_backtrace
;
95 static int __init
hardlockup_all_cpu_backtrace_setup(char *str
)
97 sysctl_hardlockup_all_cpu_backtrace
= !!simple_strtol(str
, NULL
, 0);
100 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup
);
101 # endif /* CONFIG_SMP */
102 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
105 * These functions can be overridden if an architecture implements its
106 * own hardlockup detector.
108 * watchdog_nmi_enable/disable can be implemented to start and stop when
109 * softlockup watchdog threads start and stop. The arch must select the
110 * SOFTLOCKUP_DETECTOR Kconfig.
113 #ifndef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
114 int __weak
watchdog_nmi_enable(unsigned int cpu
)
116 hardlockup_detector_perf_enable();
120 void __weak
watchdog_nmi_disable(unsigned int cpu
)
122 hardlockup_detector_perf_disable();
126 /* Return 0, if a NMI watchdog is available. Error code otherwise */
127 int __weak __init
watchdog_nmi_probe(void)
129 return hardlockup_detector_perf_init();
133 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
135 * The reconfiguration steps are:
136 * watchdog_nmi_stop();
137 * update_variables();
138 * watchdog_nmi_start();
140 void __weak
watchdog_nmi_stop(void) { }
143 * watchdog_nmi_start - Start the watchdog after reconfiguration
145 * Counterpart to watchdog_nmi_stop().
147 * The following variables have been updated in update_variables() and
148 * contain the currently valid configuration:
153 void __weak
watchdog_nmi_start(void) { }
156 * lockup_detector_update_enable - Update the sysctl enable bit
158 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
159 * can't race with watchdog_nmi_disable().
161 static void lockup_detector_update_enable(void)
163 watchdog_enabled
= 0;
164 if (!watchdog_user_enabled
)
166 #if defined(CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU)
167 if (watchdog_other_cpu_available
&& nmi_watchdog_user_enabled
)
168 watchdog_enabled
|= NMI_WATCHDOG_ENABLED
;
170 if (nmi_watchdog_available
&& nmi_watchdog_user_enabled
)
171 watchdog_enabled
|= NMI_WATCHDOG_ENABLED
;
172 if (soft_watchdog_user_enabled
)
173 watchdog_enabled
|= SOFT_WATCHDOG_ENABLED
;
176 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
178 /* Global variables, exported for sysctl */
179 unsigned int __read_mostly softlockup_panic
=
180 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE
;
182 static bool softlockup_threads_initialized __read_mostly
;
183 static u64 __read_mostly sample_period
;
184 static unsigned long __read_mostly hardlockup_thresh
;
186 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
187 static DEFINE_PER_CPU(unsigned long, hardlockup_touch_ts
);
188 static DEFINE_PER_CPU(struct task_struct
*, softlockup_watchdog
);
189 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
190 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
191 static DEFINE_PER_CPU(bool, soft_watchdog_warn
);
192 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
193 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt
);
194 static DEFINE_PER_CPU(struct task_struct
*, softlockup_task_ptr_saved
);
195 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
196 static unsigned long soft_lockup_nmi_warn
;
198 static int __init
softlockup_panic_setup(char *str
)
200 softlockup_panic
= simple_strtoul(str
, NULL
, 0);
203 __setup("softlockup_panic=", softlockup_panic_setup
);
205 static int __init
nowatchdog_setup(char *str
)
207 watchdog_user_enabled
= 0;
210 __setup("nowatchdog", nowatchdog_setup
);
212 static int __init
nosoftlockup_setup(char *str
)
214 soft_watchdog_user_enabled
= 0;
217 __setup("nosoftlockup", nosoftlockup_setup
);
220 int __read_mostly sysctl_softlockup_all_cpu_backtrace
;
222 static int __init
softlockup_all_cpu_backtrace_setup(char *str
)
224 sysctl_softlockup_all_cpu_backtrace
= !!simple_strtol(str
, NULL
, 0);
227 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup
);
230 static void __lockup_detector_cleanup(void);
233 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
234 * lockups can have false positives under extreme conditions. So we generally
235 * want a higher threshold for soft lockups than for hard lockups. So we couple
236 * the thresholds with a factor: we make the soft threshold twice the amount of
237 * time the hard threshold is.
239 static int get_softlockup_thresh(void)
241 return watchdog_thresh
* 2;
245 * Returns seconds, approximately. We don't need nanosecond
246 * resolution, and we don't need to waste time with a big divide when
249 static unsigned long get_timestamp(void)
251 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
254 static void set_sample_period(void)
257 * convert watchdog_thresh from seconds to ns
258 * the divide by 5 is to give hrtimer several chances (two
259 * or three with the current relation between the soft
260 * and hard thresholds) to increment before the
261 * hardlockup detector generates a warning
263 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
264 watchdog_update_hrtimer_threshold(sample_period
);
265 hardlockup_thresh
= sample_period
* 3 / NSEC_PER_SEC
;
268 /* Commands for resetting the watchdog */
269 static void __touch_watchdog(void)
271 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
272 __this_cpu_write(hardlockup_touch_ts
, get_timestamp());
276 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
278 * Call when the scheduler may have stalled for legitimate reasons
279 * preventing the watchdog task from executing - e.g. the scheduler
280 * entering idle state. This should only be used for scheduler events.
281 * Use touch_softlockup_watchdog() for everything else.
283 notrace
void touch_softlockup_watchdog_sched(void)
286 * Preemption can be enabled. It doesn't matter which CPU's timestamp
287 * gets zeroed here, so use the raw_ operation.
289 raw_cpu_write(watchdog_touch_ts
, 0);
292 notrace
void touch_softlockup_watchdog(void)
294 touch_softlockup_watchdog_sched();
295 wq_watchdog_touch(raw_smp_processor_id());
297 EXPORT_SYMBOL(touch_softlockup_watchdog
);
299 void touch_all_softlockup_watchdogs(void)
304 * watchdog_mutex cannpt be taken here, as this might be called
305 * from (soft)interrupt context, so the access to
306 * watchdog_allowed_cpumask might race with a concurrent update.
308 * The watchdog time stamp can race against a concurrent real
309 * update as well, the only side effect might be a cycle delay for
310 * the softlockup check.
312 for_each_cpu(cpu
, &watchdog_allowed_mask
)
313 per_cpu(watchdog_touch_ts
, cpu
) = 0;
314 wq_watchdog_touch(-1);
317 void touch_softlockup_watchdog_sync(void)
319 __this_cpu_write(softlockup_touch_sync
, true);
320 __this_cpu_write(watchdog_touch_ts
, 0);
323 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
324 static void watchdog_check_hardlockup_other_cpu(void);
326 static inline void watchdog_check_hardlockup_other_cpu(void) { return; }
329 static int is_softlockup(unsigned long touch_ts
)
331 unsigned long now
= get_timestamp();
333 if ((watchdog_enabled
& SOFT_WATCHDOG_ENABLED
) && watchdog_thresh
){
334 /* Warn about unreasonable delays. */
335 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
336 return now
- touch_ts
;
341 /* watchdog detector functions */
342 bool is_hardlockup(void)
344 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
346 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
349 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
353 static void watchdog_interrupt_count(void)
355 __this_cpu_inc(hrtimer_interrupts
);
358 /* watchdog kicker functions */
359 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
361 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
362 struct pt_regs
*regs
= get_irq_regs();
364 int softlockup_all_cpu_backtrace
= sysctl_softlockup_all_cpu_backtrace
;
366 /* try to enable log_kevent of exynos-snapshot if log_kevent was off because of rcu stall */
367 dbg_snapshot_try_enable("log_kevent", NSEC_PER_SEC
* 15);
368 if (!watchdog_enabled
)
369 return HRTIMER_NORESTART
;
371 /* kick the hardlockup detector */
372 watchdog_interrupt_count();
374 /* test for hardlockups on the next cpu */
375 watchdog_check_hardlockup_other_cpu();
377 /* kick the softlockup detector */
378 wake_up_process(__this_cpu_read(softlockup_watchdog
));
381 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
384 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
386 * If the time stamp was touched atomically
387 * make sure the scheduler tick is up to date.
389 __this_cpu_write(softlockup_touch_sync
, false);
393 /* Clear the guest paused flag on watchdog reset */
394 kvm_check_and_clear_guest_paused();
396 return HRTIMER_RESTART
;
399 /* check for a softlockup
400 * This is done by making sure a high priority task is
401 * being scheduled. The task touches the watchdog to
402 * indicate it is getting cpu time. If it hasn't then
403 * this is a good indication some task is hogging the cpu
405 duration
= is_softlockup(touch_ts
);
406 if (unlikely(duration
)) {
408 * If a virtual machine is stopped by the host it can look to
409 * the watchdog like a soft lockup, check to see if the host
410 * stopped the vm before we issue the warning
412 if (kvm_check_and_clear_guest_paused())
413 return HRTIMER_RESTART
;
416 if (__this_cpu_read(soft_watchdog_warn
) == true) {
418 * When multiple processes are causing softlockups the
419 * softlockup detector only warns on the first one
420 * because the code relies on a full quiet cycle to
421 * re-arm. The second process prevents the quiet cycle
422 * and never gets reported. Use task pointers to detect
425 if (__this_cpu_read(softlockup_task_ptr_saved
) !=
427 __this_cpu_write(soft_watchdog_warn
, false);
430 return HRTIMER_RESTART
;
433 if (softlockup_all_cpu_backtrace
) {
434 /* Prevent multiple soft-lockup reports if one cpu is already
435 * engaged in dumping cpu back traces
437 if (test_and_set_bit(0, &soft_lockup_nmi_warn
)) {
438 /* Someone else will report us. Let's give up */
439 __this_cpu_write(soft_watchdog_warn
, true);
440 return HRTIMER_RESTART
;
444 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
445 smp_processor_id(), duration
,
446 current
->comm
, task_pid_nr(current
));
447 __this_cpu_write(softlockup_task_ptr_saved
, current
);
449 print_irqtrace_events(current
);
455 if (softlockup_all_cpu_backtrace
) {
456 /* Avoid generating two back traces for current
457 * given that one is already made above
459 trigger_allbutself_cpu_backtrace();
461 clear_bit(0, &soft_lockup_nmi_warn
);
462 /* Barrier to sync with other cpus */
463 smp_mb__after_atomic();
466 add_taint(TAINT_SOFTLOCKUP
, LOCKDEP_STILL_OK
);
467 if (softlockup_panic
)
468 panic("softlockup: hung tasks");
469 __this_cpu_write(soft_watchdog_warn
, true);
471 __this_cpu_write(soft_watchdog_warn
, false);
473 return HRTIMER_RESTART
;
476 static void watchdog_set_prio(unsigned int policy
, unsigned int prio
)
478 struct sched_param param
= { .sched_priority
= prio
};
480 sched_setscheduler(current
, policy
, ¶m
);
483 static void watchdog_enable(unsigned int cpu
)
485 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
488 * Start the timer first to prevent the NMI watchdog triggering
489 * before the timer has a chance to fire.
491 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
492 hrtimer
->function
= watchdog_timer_fn
;
493 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
494 HRTIMER_MODE_REL_PINNED
);
496 /* Initialize timestamp */
498 /* Enable the perf event */
499 if (watchdog_enabled
& NMI_WATCHDOG_ENABLED
)
500 watchdog_nmi_enable(cpu
);
502 watchdog_set_prio(SCHED_FIFO
, MAX_RT_PRIO
- 1);
505 static void watchdog_disable(unsigned int cpu
)
507 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
509 watchdog_set_prio(SCHED_NORMAL
, 0);
511 * Disable the perf event first. That prevents that a large delay
512 * between disabling the timer and disabling the perf event causes
513 * the perf NMI to detect a false positive.
515 watchdog_nmi_disable(cpu
);
516 hrtimer_cancel(hrtimer
);
519 static void watchdog_cleanup(unsigned int cpu
, bool online
)
521 watchdog_disable(cpu
);
524 static int watchdog_should_run(unsigned int cpu
)
526 return __this_cpu_read(hrtimer_interrupts
) !=
527 __this_cpu_read(soft_lockup_hrtimer_cnt
);
531 * The watchdog thread function - touches the timestamp.
533 * It only runs once every sample_period seconds (4 seconds by
534 * default) to reset the softlockup timestamp. If this gets delayed
535 * for more than 2*watchdog_thresh seconds then the debug-printout
536 * triggers in watchdog_timer_fn().
538 static void watchdog(unsigned int cpu
)
540 __this_cpu_write(soft_lockup_hrtimer_cnt
,
541 __this_cpu_read(hrtimer_interrupts
));
545 static struct smp_hotplug_thread watchdog_threads
= {
546 .store
= &softlockup_watchdog
,
547 .thread_should_run
= watchdog_should_run
,
548 .thread_fn
= watchdog
,
549 .thread_comm
= "watchdog/%u",
550 .setup
= watchdog_enable
,
551 .cleanup
= watchdog_cleanup
,
552 .park
= watchdog_disable
,
553 .unpark
= watchdog_enable
,
556 static void softlockup_update_smpboot_threads(void)
558 lockdep_assert_held(&watchdog_mutex
);
560 if (!softlockup_threads_initialized
)
563 smpboot_update_cpumask_percpu_thread(&watchdog_threads
,
564 &watchdog_allowed_mask
);
567 /* Temporarily park all watchdog threads */
568 static void softlockup_park_all_threads(void)
570 cpumask_clear(&watchdog_allowed_mask
);
571 softlockup_update_smpboot_threads();
574 /* Unpark enabled threads */
575 static void softlockup_unpark_threads(void)
577 cpumask_copy(&watchdog_allowed_mask
, &watchdog_cpumask
);
578 softlockup_update_smpboot_threads();
581 static void lockup_detector_reconfigure(void)
585 softlockup_park_all_threads();
587 lockup_detector_update_enable();
588 if (watchdog_enabled
&& watchdog_thresh
)
589 softlockup_unpark_threads();
590 watchdog_nmi_start();
593 * Must be called outside the cpus locked section to prevent
594 * recursive locking in the perf code.
596 __lockup_detector_cleanup();
600 * Create the watchdog thread infrastructure and configure the detector(s).
602 * The threads are not unparked as watchdog_allowed_mask is empty. When
603 * the threads are sucessfully initialized, take the proper locks and
604 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
606 static __init
void lockup_detector_setup(void)
611 * If sysctl is off and watchdog got disabled on the command line,
612 * nothing to do here.
614 lockup_detector_update_enable();
616 if (!IS_ENABLED(CONFIG_SYSCTL
) &&
617 !(watchdog_enabled
&& watchdog_thresh
))
620 ret
= smpboot_register_percpu_thread_cpumask(&watchdog_threads
,
621 &watchdog_allowed_mask
);
623 pr_err("Failed to initialize soft lockup detector threads\n");
627 mutex_lock(&watchdog_mutex
);
628 softlockup_threads_initialized
= true;
629 lockup_detector_reconfigure();
630 mutex_unlock(&watchdog_mutex
);
633 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
634 static inline int watchdog_park_threads(void) { return 0; }
635 static inline void watchdog_unpark_threads(void) { }
636 static inline int watchdog_enable_all_cpus(void) { return 0; }
637 static inline void watchdog_disable_all_cpus(void) { }
638 static void lockup_detector_reconfigure(void)
642 lockup_detector_update_enable();
643 watchdog_nmi_start();
646 static inline void lockup_detector_setup(void)
648 lockup_detector_reconfigure();
650 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
652 static void __lockup_detector_cleanup(void)
654 lockdep_assert_held(&watchdog_mutex
);
655 hardlockup_detector_perf_cleanup();
659 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
661 * Caller must not hold the cpu hotplug rwsem.
663 void lockup_detector_cleanup(void)
665 mutex_lock(&watchdog_mutex
);
666 __lockup_detector_cleanup();
667 mutex_unlock(&watchdog_mutex
);
671 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
673 * Special interface for parisc. It prevents lockup detector warnings from
674 * the default pm_poweroff() function which busy loops forever.
676 void lockup_detector_soft_poweroff(void)
678 watchdog_enabled
= 0;
683 /* Propagate any changes to the watchdog threads */
684 static void proc_watchdog_update(void)
686 /* Remove impossible cpus to keep sysctl output clean. */
687 cpumask_and(&watchdog_cpumask
, &watchdog_cpumask
, cpu_possible_mask
);
688 lockup_detector_reconfigure();
692 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
694 * caller | table->data points to | 'which'
695 * -------------------|----------------------------|--------------------------
696 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
697 * | | SOFT_WATCHDOG_ENABLED
698 * -------------------|----------------------------|--------------------------
699 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
700 * -------------------|----------------------------|--------------------------
701 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
703 static int proc_watchdog_common(int which
, struct ctl_table
*table
, int write
,
704 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
706 int err
, old
, *param
= table
->data
;
708 mutex_lock(&watchdog_mutex
);
712 * On read synchronize the userspace interface. This is a
715 *param
= (watchdog_enabled
& which
) != 0;
716 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
718 old
= READ_ONCE(*param
);
719 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
720 if (!err
&& old
!= READ_ONCE(*param
))
721 proc_watchdog_update();
723 mutex_unlock(&watchdog_mutex
);
728 * /proc/sys/kernel/watchdog
730 int proc_watchdog(struct ctl_table
*table
, int write
,
731 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
733 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
|SOFT_WATCHDOG_ENABLED
,
734 table
, write
, buffer
, lenp
, ppos
);
738 * /proc/sys/kernel/nmi_watchdog
740 int proc_nmi_watchdog(struct ctl_table
*table
, int write
,
741 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
743 if (!nmi_watchdog_available
&& write
)
745 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
,
746 table
, write
, buffer
, lenp
, ppos
);
750 * /proc/sys/kernel/soft_watchdog
752 int proc_soft_watchdog(struct ctl_table
*table
, int write
,
753 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
755 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED
,
756 table
, write
, buffer
, lenp
, ppos
);
760 * /proc/sys/kernel/watchdog_thresh
762 int proc_watchdog_thresh(struct ctl_table
*table
, int write
,
763 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
767 mutex_lock(&watchdog_mutex
);
769 old
= READ_ONCE(watchdog_thresh
);
770 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
772 if (!err
&& write
&& old
!= READ_ONCE(watchdog_thresh
))
773 proc_watchdog_update();
775 mutex_unlock(&watchdog_mutex
);
780 * The cpumask is the mask of possible cpus that the watchdog can run
781 * on, not the mask of cpus it is actually running on. This allows the
782 * user to specify a mask that will include cpus that have not yet
783 * been brought online, if desired.
785 int proc_watchdog_cpumask(struct ctl_table
*table
, int write
,
786 void __user
*buffer
, size_t *lenp
, loff_t
*ppos
)
790 mutex_lock(&watchdog_mutex
);
792 err
= proc_do_large_bitmap(table
, write
, buffer
, lenp
, ppos
);
794 proc_watchdog_update();
796 mutex_unlock(&watchdog_mutex
);
799 #endif /* CONFIG_SYSCTL */
801 void __init
lockup_detector_init(void)
803 #ifdef CONFIG_NO_HZ_FULL
804 if (tick_nohz_full_enabled()) {
805 pr_info("Disabling watchdog on nohz_full cores by default\n");
806 cpumask_copy(&watchdog_cpumask
, housekeeping_mask
);
808 cpumask_copy(&watchdog_cpumask
, cpu_possible_mask
);
810 cpumask_copy(&watchdog_cpumask
, cpu_possible_mask
);
813 if (!watchdog_nmi_probe())
814 nmi_watchdog_available
= true;
815 lockup_detector_setup();
818 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
819 static DEFINE_PER_CPU(bool, hard_watchdog_warn
);
820 static DEFINE_PER_CPU(bool, watchdog_nmi_touch
);
821 static cpumask_t __read_mostly watchdog_cpus
;
822 ATOMIC_NOTIFIER_HEAD(hardlockup_notifier_list
);
823 EXPORT_SYMBOL(hardlockup_notifier_list
);
825 static unsigned int watchdog_next_cpu(unsigned int cpu
)
827 cpumask_t cpus
= watchdog_cpus
;
828 unsigned int next_cpu
;
830 next_cpu
= cpumask_next(cpu
, &cpus
);
831 if (next_cpu
>= nr_cpu_ids
)
832 next_cpu
= cpumask_first(&cpus
);
840 static int is_hardlockup_other_cpu(unsigned int cpu
)
842 unsigned long hrint
= per_cpu(hrtimer_interrupts
, cpu
);
844 if (per_cpu(hrtimer_interrupts_saved
, cpu
) == hrint
) {
845 unsigned long now
= get_timestamp();
846 unsigned long touch_ts
= per_cpu(hardlockup_touch_ts
, cpu
);
848 if (time_after(now
, touch_ts
) &&
849 (now
- touch_ts
>= hardlockup_thresh
))
853 per_cpu(hrtimer_interrupts_saved
, cpu
) = hrint
;
857 static void watchdog_check_hardlockup_other_cpu(void)
859 unsigned int next_cpu
;
862 * Test for hardlockups every 3 samples. The sample period is
863 * watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over
864 * watchdog_thresh (over by 20%).
866 if (__this_cpu_read(hrtimer_interrupts
) % 3 != 0)
869 /* check for a hardlockup on the next cpu */
870 next_cpu
= watchdog_next_cpu(smp_processor_id());
871 if (next_cpu
>= nr_cpu_ids
)
876 if (per_cpu(watchdog_nmi_touch
, next_cpu
) == true) {
877 per_cpu(watchdog_nmi_touch
, next_cpu
) = false;
881 if (is_hardlockup_other_cpu(next_cpu
)) {
883 if (per_cpu(hard_watchdog_warn
, next_cpu
) == true)
886 if (hardlockup_panic
) {
887 dbg_snapshot_set_hardlockup(hardlockup_panic
);
888 atomic_notifier_call_chain(&hardlockup_notifier_list
, 0, (void *)&next_cpu
);
889 panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu
);
891 WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu
);
894 per_cpu(hard_watchdog_warn
, next_cpu
) = true;
896 per_cpu(hard_watchdog_warn
, next_cpu
) = false;
900 void touch_nmi_watchdog(void)
903 * Using __raw here because some code paths have
904 * preemption enabled. If preemption is enabled
905 * then interrupts should be enabled too, in which
906 * case we shouldn't have to worry about the watchdog
909 raw_cpu_write(watchdog_nmi_touch
, true);
910 arch_touch_nmi_watchdog();
911 touch_softlockup_watchdog();
913 EXPORT_SYMBOL(touch_nmi_watchdog
);
915 int watchdog_nmi_enable(unsigned int cpu
)
918 * The new cpu will be marked online before the first hrtimer interrupt
919 * runs on it. If another cpu tests for a hardlockup on the new cpu
920 * before it has run its first hrtimer, it will get a false positive.
921 * Touch the watchdog on the new cpu to delay the first check for at
922 * least 3 sampling periods to guarantee one hrtimer has run on the new
925 per_cpu(watchdog_nmi_touch
, cpu
) = true;
927 cpumask_set_cpu(cpu
, &watchdog_cpus
);
931 void watchdog_nmi_disable(unsigned int cpu
)
933 unsigned int next_cpu
= watchdog_next_cpu(cpu
);
936 * Offlining this cpu will cause the cpu before this one to start
937 * checking the one after this one. If this cpu just finished checking
938 * the next cpu and updating hrtimer_interrupts_saved, and then the
939 * previous cpu checks it within one sample period, it will trigger a
940 * false positive. Touch the watchdog on the next cpu to prevent it.
942 if (next_cpu
< nr_cpu_ids
)
943 per_cpu(watchdog_nmi_touch
, next_cpu
) = true;
945 cpumask_clear_cpu(cpu
, &watchdog_cpus
);