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 struct cpumask watchdog_cpumask __read_mostly
;
48 unsigned long *watchdog_cpumask_bits
= cpumask_bits(&watchdog_cpumask
);
50 #ifdef CONFIG_HARDLOCKUP_DETECTOR
53 int __read_mostly sysctl_hardlockup_all_cpu_backtrace
;
54 # endif /* CONFIG_SMP */
57 * Should we panic when a soft-lockup or hard-lockup occurs:
59 unsigned int __read_mostly hardlockup_panic
=
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE
;
62 * We may not want to enable hard lockup detection by default in all cases,
63 * for example when running the kernel as a guest on a hypervisor. In these
64 * cases this function can be called to disable hard lockup detection. This
65 * function should only be executed once by the boot processor before the
66 * kernel command line parameters are parsed, because otherwise it is not
67 * possible to override this in hardlockup_panic_setup().
69 void __init
hardlockup_detector_disable(void)
71 nmi_watchdog_user_enabled
= 0;
74 static int __init
hardlockup_panic_setup(char *str
)
76 if (!strncmp(str
, "panic", 5))
78 else if (!strncmp(str
, "nopanic", 7))
80 else if (!strncmp(str
, "0", 1))
81 nmi_watchdog_user_enabled
= 0;
82 else if (!strncmp(str
, "1", 1))
83 nmi_watchdog_user_enabled
= 1;
86 __setup("nmi_watchdog=", hardlockup_panic_setup
);
88 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
91 * These functions can be overridden if an architecture implements its
92 * own hardlockup detector.
94 * watchdog_nmi_enable/disable can be implemented to start and stop when
95 * softlockup watchdog threads start and stop. The arch must select the
96 * SOFTLOCKUP_DETECTOR Kconfig.
98 int __weak
watchdog_nmi_enable(unsigned int cpu
)
100 hardlockup_detector_perf_enable();
104 void __weak
watchdog_nmi_disable(unsigned int cpu
)
106 hardlockup_detector_perf_disable();
109 /* Return 0, if a NMI watchdog is available. Error code otherwise */
110 int __weak __init
watchdog_nmi_probe(void)
112 return hardlockup_detector_perf_init();
116 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
118 * The reconfiguration steps are:
119 * watchdog_nmi_stop();
120 * update_variables();
121 * watchdog_nmi_start();
123 void __weak
watchdog_nmi_stop(void) { }
126 * watchdog_nmi_start - Start the watchdog after reconfiguration
128 * Counterpart to watchdog_nmi_stop().
130 * The following variables have been updated in update_variables() and
131 * contain the currently valid configuration:
136 void __weak
watchdog_nmi_start(void) { }
139 * lockup_detector_update_enable - Update the sysctl enable bit
141 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
142 * can't race with watchdog_nmi_disable().
144 static void lockup_detector_update_enable(void)
146 watchdog_enabled
= 0;
147 if (!watchdog_user_enabled
)
149 if (nmi_watchdog_available
&& nmi_watchdog_user_enabled
)
150 watchdog_enabled
|= NMI_WATCHDOG_ENABLED
;
151 if (soft_watchdog_user_enabled
)
152 watchdog_enabled
|= SOFT_WATCHDOG_ENABLED
;
155 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
157 #define SOFTLOCKUP_RESET ULONG_MAX
160 int __read_mostly sysctl_softlockup_all_cpu_backtrace
;
163 static struct cpumask watchdog_allowed_mask __read_mostly
;
165 /* Global variables, exported for sysctl */
166 unsigned int __read_mostly softlockup_panic
=
167 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE
;
169 static bool softlockup_initialized __read_mostly
;
170 static u64 __read_mostly sample_period
;
172 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts
);
173 static DEFINE_PER_CPU(struct hrtimer
, watchdog_hrtimer
);
174 static DEFINE_PER_CPU(bool, softlockup_touch_sync
);
175 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts
);
176 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved
);
177 static unsigned long soft_lockup_nmi_warn
;
179 static int __init
nowatchdog_setup(char *str
)
181 watchdog_user_enabled
= 0;
184 __setup("nowatchdog", nowatchdog_setup
);
186 static int __init
nosoftlockup_setup(char *str
)
188 soft_watchdog_user_enabled
= 0;
191 __setup("nosoftlockup", nosoftlockup_setup
);
193 static int __init
watchdog_thresh_setup(char *str
)
195 get_option(&str
, &watchdog_thresh
);
198 __setup("watchdog_thresh=", watchdog_thresh_setup
);
200 static void __lockup_detector_cleanup(void);
203 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
204 * lockups can have false positives under extreme conditions. So we generally
205 * want a higher threshold for soft lockups than for hard lockups. So we couple
206 * the thresholds with a factor: we make the soft threshold twice the amount of
207 * time the hard threshold is.
209 static int get_softlockup_thresh(void)
211 return watchdog_thresh
* 2;
215 * Returns seconds, approximately. We don't need nanosecond
216 * resolution, and we don't need to waste time with a big divide when
219 static unsigned long get_timestamp(void)
221 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
224 static void set_sample_period(void)
227 * convert watchdog_thresh from seconds to ns
228 * the divide by 5 is to give hrtimer several chances (two
229 * or three with the current relation between the soft
230 * and hard thresholds) to increment before the
231 * hardlockup detector generates a warning
233 sample_period
= get_softlockup_thresh() * ((u64
)NSEC_PER_SEC
/ 5);
234 watchdog_update_hrtimer_threshold(sample_period
);
237 /* Commands for resetting the watchdog */
238 static void update_touch_ts(void)
240 __this_cpu_write(watchdog_touch_ts
, get_timestamp());
244 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
246 * Call when the scheduler may have stalled for legitimate reasons
247 * preventing the watchdog task from executing - e.g. the scheduler
248 * entering idle state. This should only be used for scheduler events.
249 * Use touch_softlockup_watchdog() for everything else.
251 notrace
void touch_softlockup_watchdog_sched(void)
254 * Preemption can be enabled. It doesn't matter which CPU's timestamp
255 * gets zeroed here, so use the raw_ operation.
257 raw_cpu_write(watchdog_touch_ts
, SOFTLOCKUP_RESET
);
260 notrace
void touch_softlockup_watchdog(void)
262 touch_softlockup_watchdog_sched();
263 wq_watchdog_touch(raw_smp_processor_id());
265 EXPORT_SYMBOL(touch_softlockup_watchdog
);
267 void touch_all_softlockup_watchdogs(void)
272 * watchdog_mutex cannpt be taken here, as this might be called
273 * from (soft)interrupt context, so the access to
274 * watchdog_allowed_cpumask might race with a concurrent update.
276 * The watchdog time stamp can race against a concurrent real
277 * update as well, the only side effect might be a cycle delay for
278 * the softlockup check.
280 for_each_cpu(cpu
, &watchdog_allowed_mask
)
281 per_cpu(watchdog_touch_ts
, cpu
) = SOFTLOCKUP_RESET
;
282 wq_watchdog_touch(-1);
285 void touch_softlockup_watchdog_sync(void)
287 __this_cpu_write(softlockup_touch_sync
, true);
288 __this_cpu_write(watchdog_touch_ts
, SOFTLOCKUP_RESET
);
291 static int is_softlockup(unsigned long touch_ts
)
293 unsigned long now
= get_timestamp();
295 if ((watchdog_enabled
& SOFT_WATCHDOG_ENABLED
) && watchdog_thresh
){
296 /* Warn about unreasonable delays. */
297 if (time_after(now
, touch_ts
+ get_softlockup_thresh()))
298 return now
- touch_ts
;
303 /* watchdog detector functions */
304 bool is_hardlockup(void)
306 unsigned long hrint
= __this_cpu_read(hrtimer_interrupts
);
308 if (__this_cpu_read(hrtimer_interrupts_saved
) == hrint
)
311 __this_cpu_write(hrtimer_interrupts_saved
, hrint
);
315 static void watchdog_interrupt_count(void)
317 __this_cpu_inc(hrtimer_interrupts
);
320 static DEFINE_PER_CPU(struct completion
, softlockup_completion
);
321 static DEFINE_PER_CPU(struct cpu_stop_work
, softlockup_stop_work
);
324 * The watchdog thread function - touches the timestamp.
326 * It only runs once every sample_period seconds (4 seconds by
327 * default) to reset the softlockup timestamp. If this gets delayed
328 * for more than 2*watchdog_thresh seconds then the debug-printout
329 * triggers in watchdog_timer_fn().
331 static int softlockup_fn(void *data
)
334 complete(this_cpu_ptr(&softlockup_completion
));
339 /* watchdog kicker functions */
340 static enum hrtimer_restart
watchdog_timer_fn(struct hrtimer
*hrtimer
)
342 unsigned long touch_ts
= __this_cpu_read(watchdog_touch_ts
);
343 struct pt_regs
*regs
= get_irq_regs();
345 int softlockup_all_cpu_backtrace
= sysctl_softlockup_all_cpu_backtrace
;
347 if (!watchdog_enabled
)
348 return HRTIMER_NORESTART
;
350 /* kick the hardlockup detector */
351 watchdog_interrupt_count();
353 /* kick the softlockup detector */
354 if (completion_done(this_cpu_ptr(&softlockup_completion
))) {
355 reinit_completion(this_cpu_ptr(&softlockup_completion
));
356 stop_one_cpu_nowait(smp_processor_id(),
358 this_cpu_ptr(&softlockup_stop_work
));
362 hrtimer_forward_now(hrtimer
, ns_to_ktime(sample_period
));
364 if (touch_ts
== SOFTLOCKUP_RESET
) {
365 if (unlikely(__this_cpu_read(softlockup_touch_sync
))) {
367 * If the time stamp was touched atomically
368 * make sure the scheduler tick is up to date.
370 __this_cpu_write(softlockup_touch_sync
, false);
374 /* Clear the guest paused flag on watchdog reset */
375 kvm_check_and_clear_guest_paused();
377 return HRTIMER_RESTART
;
380 /* check for a softlockup
381 * This is done by making sure a high priority task is
382 * being scheduled. The task touches the watchdog to
383 * indicate it is getting cpu time. If it hasn't then
384 * this is a good indication some task is hogging the cpu
386 duration
= is_softlockup(touch_ts
);
387 if (unlikely(duration
)) {
389 * If a virtual machine is stopped by the host it can look to
390 * the watchdog like a soft lockup, check to see if the host
391 * stopped the vm before we issue the warning
393 if (kvm_check_and_clear_guest_paused())
394 return HRTIMER_RESTART
;
397 * Prevent multiple soft-lockup reports if one cpu is already
398 * engaged in dumping all cpu back traces.
400 if (softlockup_all_cpu_backtrace
) {
401 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn
))
402 return HRTIMER_RESTART
;
405 /* Start period for the next softlockup warning. */
408 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
409 smp_processor_id(), duration
,
410 current
->comm
, task_pid_nr(current
));
412 print_irqtrace_events(current
);
418 if (softlockup_all_cpu_backtrace
) {
419 trigger_allbutself_cpu_backtrace();
420 clear_bit_unlock(0, &soft_lockup_nmi_warn
);
423 add_taint(TAINT_SOFTLOCKUP
, LOCKDEP_STILL_OK
);
424 if (softlockup_panic
)
425 panic("softlockup: hung tasks");
428 return HRTIMER_RESTART
;
431 static void watchdog_enable(unsigned int cpu
)
433 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
434 struct completion
*done
= this_cpu_ptr(&softlockup_completion
);
436 WARN_ON_ONCE(cpu
!= smp_processor_id());
438 init_completion(done
);
442 * Start the timer first to prevent the NMI watchdog triggering
443 * before the timer has a chance to fire.
445 hrtimer_init(hrtimer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL_HARD
);
446 hrtimer
->function
= watchdog_timer_fn
;
447 hrtimer_start(hrtimer
, ns_to_ktime(sample_period
),
448 HRTIMER_MODE_REL_PINNED_HARD
);
450 /* Initialize timestamp */
452 /* Enable the perf event */
453 if (watchdog_enabled
& NMI_WATCHDOG_ENABLED
)
454 watchdog_nmi_enable(cpu
);
457 static void watchdog_disable(unsigned int cpu
)
459 struct hrtimer
*hrtimer
= this_cpu_ptr(&watchdog_hrtimer
);
461 WARN_ON_ONCE(cpu
!= smp_processor_id());
464 * Disable the perf event first. That prevents that a large delay
465 * between disabling the timer and disabling the perf event causes
466 * the perf NMI to detect a false positive.
468 watchdog_nmi_disable(cpu
);
469 hrtimer_cancel(hrtimer
);
470 wait_for_completion(this_cpu_ptr(&softlockup_completion
));
473 static int softlockup_stop_fn(void *data
)
475 watchdog_disable(smp_processor_id());
479 static void softlockup_stop_all(void)
483 if (!softlockup_initialized
)
486 for_each_cpu(cpu
, &watchdog_allowed_mask
)
487 smp_call_on_cpu(cpu
, softlockup_stop_fn
, NULL
, false);
489 cpumask_clear(&watchdog_allowed_mask
);
492 static int softlockup_start_fn(void *data
)
494 watchdog_enable(smp_processor_id());
498 static void softlockup_start_all(void)
502 cpumask_copy(&watchdog_allowed_mask
, &watchdog_cpumask
);
503 for_each_cpu(cpu
, &watchdog_allowed_mask
)
504 smp_call_on_cpu(cpu
, softlockup_start_fn
, NULL
, false);
507 int lockup_detector_online_cpu(unsigned int cpu
)
509 if (cpumask_test_cpu(cpu
, &watchdog_allowed_mask
))
510 watchdog_enable(cpu
);
514 int lockup_detector_offline_cpu(unsigned int cpu
)
516 if (cpumask_test_cpu(cpu
, &watchdog_allowed_mask
))
517 watchdog_disable(cpu
);
521 static void lockup_detector_reconfigure(void)
526 softlockup_stop_all();
528 lockup_detector_update_enable();
529 if (watchdog_enabled
&& watchdog_thresh
)
530 softlockup_start_all();
532 watchdog_nmi_start();
535 * Must be called outside the cpus locked section to prevent
536 * recursive locking in the perf code.
538 __lockup_detector_cleanup();
542 * Create the watchdog thread infrastructure and configure the detector(s).
544 * The threads are not unparked as watchdog_allowed_mask is empty. When
545 * the threads are successfully initialized, take the proper locks and
546 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
548 static __init
void lockup_detector_setup(void)
551 * If sysctl is off and watchdog got disabled on the command line,
552 * nothing to do here.
554 lockup_detector_update_enable();
556 if (!IS_ENABLED(CONFIG_SYSCTL
) &&
557 !(watchdog_enabled
&& watchdog_thresh
))
560 mutex_lock(&watchdog_mutex
);
561 lockup_detector_reconfigure();
562 softlockup_initialized
= true;
563 mutex_unlock(&watchdog_mutex
);
566 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
567 static void lockup_detector_reconfigure(void)
571 lockup_detector_update_enable();
572 watchdog_nmi_start();
575 static inline void lockup_detector_setup(void)
577 lockup_detector_reconfigure();
579 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
581 static void __lockup_detector_cleanup(void)
583 lockdep_assert_held(&watchdog_mutex
);
584 hardlockup_detector_perf_cleanup();
588 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
590 * Caller must not hold the cpu hotplug rwsem.
592 void lockup_detector_cleanup(void)
594 mutex_lock(&watchdog_mutex
);
595 __lockup_detector_cleanup();
596 mutex_unlock(&watchdog_mutex
);
600 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
602 * Special interface for parisc. It prevents lockup detector warnings from
603 * the default pm_poweroff() function which busy loops forever.
605 void lockup_detector_soft_poweroff(void)
607 watchdog_enabled
= 0;
612 /* Propagate any changes to the watchdog threads */
613 static void proc_watchdog_update(void)
615 /* Remove impossible cpus to keep sysctl output clean. */
616 cpumask_and(&watchdog_cpumask
, &watchdog_cpumask
, cpu_possible_mask
);
617 lockup_detector_reconfigure();
621 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
623 * caller | table->data points to | 'which'
624 * -------------------|----------------------------|--------------------------
625 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
626 * | | SOFT_WATCHDOG_ENABLED
627 * -------------------|----------------------------|--------------------------
628 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
629 * -------------------|----------------------------|--------------------------
630 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
632 static int proc_watchdog_common(int which
, struct ctl_table
*table
, int write
,
633 void *buffer
, size_t *lenp
, loff_t
*ppos
)
635 int err
, old
, *param
= table
->data
;
637 mutex_lock(&watchdog_mutex
);
641 * On read synchronize the userspace interface. This is a
644 *param
= (watchdog_enabled
& which
) != 0;
645 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
647 old
= READ_ONCE(*param
);
648 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
649 if (!err
&& old
!= READ_ONCE(*param
))
650 proc_watchdog_update();
652 mutex_unlock(&watchdog_mutex
);
657 * /proc/sys/kernel/watchdog
659 int proc_watchdog(struct ctl_table
*table
, int write
,
660 void *buffer
, size_t *lenp
, loff_t
*ppos
)
662 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
|SOFT_WATCHDOG_ENABLED
,
663 table
, write
, buffer
, lenp
, ppos
);
667 * /proc/sys/kernel/nmi_watchdog
669 int proc_nmi_watchdog(struct ctl_table
*table
, int write
,
670 void *buffer
, size_t *lenp
, loff_t
*ppos
)
672 if (!nmi_watchdog_available
&& write
)
674 return proc_watchdog_common(NMI_WATCHDOG_ENABLED
,
675 table
, write
, buffer
, lenp
, ppos
);
679 * /proc/sys/kernel/soft_watchdog
681 int proc_soft_watchdog(struct ctl_table
*table
, int write
,
682 void *buffer
, size_t *lenp
, loff_t
*ppos
)
684 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED
,
685 table
, write
, buffer
, lenp
, ppos
);
689 * /proc/sys/kernel/watchdog_thresh
691 int proc_watchdog_thresh(struct ctl_table
*table
, int write
,
692 void *buffer
, size_t *lenp
, loff_t
*ppos
)
696 mutex_lock(&watchdog_mutex
);
698 old
= READ_ONCE(watchdog_thresh
);
699 err
= proc_dointvec_minmax(table
, write
, buffer
, lenp
, ppos
);
701 if (!err
&& write
&& old
!= READ_ONCE(watchdog_thresh
))
702 proc_watchdog_update();
704 mutex_unlock(&watchdog_mutex
);
709 * The cpumask is the mask of possible cpus that the watchdog can run
710 * on, not the mask of cpus it is actually running on. This allows the
711 * user to specify a mask that will include cpus that have not yet
712 * been brought online, if desired.
714 int proc_watchdog_cpumask(struct ctl_table
*table
, int write
,
715 void *buffer
, size_t *lenp
, loff_t
*ppos
)
719 mutex_lock(&watchdog_mutex
);
721 err
= proc_do_large_bitmap(table
, write
, buffer
, lenp
, ppos
);
723 proc_watchdog_update();
725 mutex_unlock(&watchdog_mutex
);
728 #endif /* CONFIG_SYSCTL */
730 void __init
lockup_detector_init(void)
732 if (tick_nohz_full_enabled())
733 pr_info("Disabling watchdog on nohz_full cores by default\n");
735 cpumask_copy(&watchdog_cpumask
,
736 housekeeping_cpumask(HK_FLAG_TIMER
));
738 if (!watchdog_nmi_probe())
739 nmi_watchdog_available
= true;
740 lockup_detector_setup();