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 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
158 * Delay the soflockup report when running a known slow code.
159 * It does _not_ affect the timestamp of the last successdul reschedule.
161 #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
164 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
167 static struct cpumask watchdog_allowed_mask __read_mostly;
169 /* Global variables, exported for sysctl */
170 unsigned int __read_mostly softlockup_panic =
171 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
173 static bool softlockup_initialized __read_mostly;
174 static u64 __read_mostly sample_period;
176 /* Timestamp taken after the last successful reschedule. */
177 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
178 /* Timestamp of the last softlockup report. */
179 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
180 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
181 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
182 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
183 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
184 static unsigned long soft_lockup_nmi_warn;
186 static int __init nowatchdog_setup(char *str)
188 watchdog_user_enabled = 0;
191 __setup("nowatchdog", nowatchdog_setup);
193 static int __init nosoftlockup_setup(char *str)
195 soft_watchdog_user_enabled = 0;
198 __setup("nosoftlockup", nosoftlockup_setup);
200 static int __init watchdog_thresh_setup(char *str)
202 get_option(&str, &watchdog_thresh);
205 __setup("watchdog_thresh=", watchdog_thresh_setup);
207 static void __lockup_detector_cleanup(void);
210 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
211 * lockups can have false positives under extreme conditions. So we generally
212 * want a higher threshold for soft lockups than for hard lockups. So we couple
213 * the thresholds with a factor: we make the soft threshold twice the amount of
214 * time the hard threshold is.
216 static int get_softlockup_thresh(void)
218 return watchdog_thresh * 2;
222 * Returns seconds, approximately. We don't need nanosecond
223 * resolution, and we don't need to waste time with a big divide when
226 static unsigned long get_timestamp(void)
228 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
231 static void set_sample_period(void)
234 * convert watchdog_thresh from seconds to ns
235 * the divide by 5 is to give hrtimer several chances (two
236 * or three with the current relation between the soft
237 * and hard thresholds) to increment before the
238 * hardlockup detector generates a warning
240 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
241 watchdog_update_hrtimer_threshold(sample_period);
244 static void update_report_ts(void)
246 __this_cpu_write(watchdog_report_ts, get_timestamp());
249 /* Commands for resetting the watchdog */
250 static void update_touch_ts(void)
252 __this_cpu_write(watchdog_touch_ts, get_timestamp());
257 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
259 * Call when the scheduler may have stalled for legitimate reasons
260 * preventing the watchdog task from executing - e.g. the scheduler
261 * entering idle state. This should only be used for scheduler events.
262 * Use touch_softlockup_watchdog() for everything else.
264 notrace void touch_softlockup_watchdog_sched(void)
267 * Preemption can be enabled. It doesn't matter which CPU's watchdog
268 * report period gets restarted here, so use the raw_ operation.
270 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
273 notrace void touch_softlockup_watchdog(void)
275 touch_softlockup_watchdog_sched();
276 wq_watchdog_touch(raw_smp_processor_id());
278 EXPORT_SYMBOL(touch_softlockup_watchdog);
280 void touch_all_softlockup_watchdogs(void)
285 * watchdog_mutex cannpt be taken here, as this might be called
286 * from (soft)interrupt context, so the access to
287 * watchdog_allowed_cpumask might race with a concurrent update.
289 * The watchdog time stamp can race against a concurrent real
290 * update as well, the only side effect might be a cycle delay for
291 * the softlockup check.
293 for_each_cpu(cpu, &watchdog_allowed_mask) {
294 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
295 wq_watchdog_touch(cpu);
299 void touch_softlockup_watchdog_sync(void)
301 __this_cpu_write(softlockup_touch_sync, true);
302 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
305 static int is_softlockup(unsigned long touch_ts,
306 unsigned long period_ts,
309 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
310 /* Warn about unreasonable delays. */
311 if (time_after(now, period_ts + get_softlockup_thresh()))
312 return now - touch_ts;
317 /* watchdog detector functions */
318 bool is_hardlockup(void)
320 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
322 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
325 __this_cpu_write(hrtimer_interrupts_saved, hrint);
329 static void watchdog_interrupt_count(void)
331 __this_cpu_inc(hrtimer_interrupts);
334 static DEFINE_PER_CPU(struct completion, softlockup_completion);
335 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
338 * The watchdog feed function - touches the timestamp.
340 * It only runs once every sample_period seconds (4 seconds by
341 * default) to reset the softlockup timestamp. If this gets delayed
342 * for more than 2*watchdog_thresh seconds then the debug-printout
343 * triggers in watchdog_timer_fn().
345 static int softlockup_fn(void *data)
348 complete(this_cpu_ptr(&softlockup_completion));
353 /* watchdog kicker functions */
354 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
356 unsigned long touch_ts, period_ts, now;
357 struct pt_regs *regs = get_irq_regs();
359 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
361 if (!watchdog_enabled)
362 return HRTIMER_NORESTART;
364 /* kick the hardlockup detector */
365 watchdog_interrupt_count();
367 /* kick the softlockup detector */
368 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
369 reinit_completion(this_cpu_ptr(&softlockup_completion));
370 stop_one_cpu_nowait(smp_processor_id(),
372 this_cpu_ptr(&softlockup_stop_work));
376 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
379 * Read the current timestamp first. It might become invalid anytime
380 * when a virtual machine is stopped by the host or when the watchog
381 * is touched from NMI.
383 now = get_timestamp();
385 * If a virtual machine is stopped by the host it can look to
386 * the watchdog like a soft lockup. This function touches the watchdog.
388 kvm_check_and_clear_guest_paused();
390 * The stored timestamp is comparable with @now only when not touched.
391 * It might get touched anytime from NMI. Make sure that is_softlockup()
392 * uses the same (valid) value.
394 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
396 /* Reset the interval when touched by known problematic code. */
397 if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
398 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
400 * If the time stamp was touched atomically
401 * make sure the scheduler tick is up to date.
403 __this_cpu_write(softlockup_touch_sync, false);
408 return HRTIMER_RESTART;
411 /* Check for a softlockup. */
412 touch_ts = __this_cpu_read(watchdog_touch_ts);
413 duration = is_softlockup(touch_ts, period_ts, now);
414 if (unlikely(duration)) {
416 * Prevent multiple soft-lockup reports if one cpu is already
417 * engaged in dumping all cpu back traces.
419 if (softlockup_all_cpu_backtrace) {
420 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
421 return HRTIMER_RESTART;
424 /* Start period for the next softlockup warning. */
427 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
428 smp_processor_id(), duration,
429 current->comm, task_pid_nr(current));
431 print_irqtrace_events(current);
437 if (softlockup_all_cpu_backtrace) {
438 trigger_allbutself_cpu_backtrace();
439 clear_bit_unlock(0, &soft_lockup_nmi_warn);
442 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
443 if (softlockup_panic)
444 panic("softlockup: hung tasks");
447 return HRTIMER_RESTART;
450 static void watchdog_enable(unsigned int cpu)
452 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
453 struct completion *done = this_cpu_ptr(&softlockup_completion);
455 WARN_ON_ONCE(cpu != smp_processor_id());
457 init_completion(done);
461 * Start the timer first to prevent the NMI watchdog triggering
462 * before the timer has a chance to fire.
464 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
465 hrtimer->function = watchdog_timer_fn;
466 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
467 HRTIMER_MODE_REL_PINNED_HARD);
469 /* Initialize timestamp */
471 /* Enable the perf event */
472 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
473 watchdog_nmi_enable(cpu);
476 static void watchdog_disable(unsigned int cpu)
478 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
480 WARN_ON_ONCE(cpu != smp_processor_id());
483 * Disable the perf event first. That prevents that a large delay
484 * between disabling the timer and disabling the perf event causes
485 * the perf NMI to detect a false positive.
487 watchdog_nmi_disable(cpu);
488 hrtimer_cancel(hrtimer);
489 wait_for_completion(this_cpu_ptr(&softlockup_completion));
492 static int softlockup_stop_fn(void *data)
494 watchdog_disable(smp_processor_id());
498 static void softlockup_stop_all(void)
502 if (!softlockup_initialized)
505 for_each_cpu(cpu, &watchdog_allowed_mask)
506 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
508 cpumask_clear(&watchdog_allowed_mask);
511 static int softlockup_start_fn(void *data)
513 watchdog_enable(smp_processor_id());
517 static void softlockup_start_all(void)
521 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
522 for_each_cpu(cpu, &watchdog_allowed_mask)
523 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
526 int lockup_detector_online_cpu(unsigned int cpu)
528 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
529 watchdog_enable(cpu);
533 int lockup_detector_offline_cpu(unsigned int cpu)
535 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
536 watchdog_disable(cpu);
540 static void lockup_detector_reconfigure(void)
545 softlockup_stop_all();
547 lockup_detector_update_enable();
548 if (watchdog_enabled && watchdog_thresh)
549 softlockup_start_all();
551 watchdog_nmi_start();
554 * Must be called outside the cpus locked section to prevent
555 * recursive locking in the perf code.
557 __lockup_detector_cleanup();
561 * Create the watchdog infrastructure and configure the detector(s).
563 static __init void lockup_detector_setup(void)
566 * If sysctl is off and watchdog got disabled on the command line,
567 * nothing to do here.
569 lockup_detector_update_enable();
571 if (!IS_ENABLED(CONFIG_SYSCTL) &&
572 !(watchdog_enabled && watchdog_thresh))
575 mutex_lock(&watchdog_mutex);
576 lockup_detector_reconfigure();
577 softlockup_initialized = true;
578 mutex_unlock(&watchdog_mutex);
581 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
582 static void lockup_detector_reconfigure(void)
586 lockup_detector_update_enable();
587 watchdog_nmi_start();
590 static inline void lockup_detector_setup(void)
592 lockup_detector_reconfigure();
594 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
596 static void __lockup_detector_cleanup(void)
598 lockdep_assert_held(&watchdog_mutex);
599 hardlockup_detector_perf_cleanup();
603 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
605 * Caller must not hold the cpu hotplug rwsem.
607 void lockup_detector_cleanup(void)
609 mutex_lock(&watchdog_mutex);
610 __lockup_detector_cleanup();
611 mutex_unlock(&watchdog_mutex);
615 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
617 * Special interface for parisc. It prevents lockup detector warnings from
618 * the default pm_poweroff() function which busy loops forever.
620 void lockup_detector_soft_poweroff(void)
622 watchdog_enabled = 0;
627 /* Propagate any changes to the watchdog infrastructure */
628 static void proc_watchdog_update(void)
630 /* Remove impossible cpus to keep sysctl output clean. */
631 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
632 lockup_detector_reconfigure();
636 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
638 * caller | table->data points to | 'which'
639 * -------------------|----------------------------|--------------------------
640 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
641 * | | SOFT_WATCHDOG_ENABLED
642 * -------------------|----------------------------|--------------------------
643 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
644 * -------------------|----------------------------|--------------------------
645 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
647 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
648 void *buffer, size_t *lenp, loff_t *ppos)
650 int err, old, *param = table->data;
652 mutex_lock(&watchdog_mutex);
656 * On read synchronize the userspace interface. This is a
659 *param = (watchdog_enabled & which) != 0;
660 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
662 old = READ_ONCE(*param);
663 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
664 if (!err && old != READ_ONCE(*param))
665 proc_watchdog_update();
667 mutex_unlock(&watchdog_mutex);
672 * /proc/sys/kernel/watchdog
674 int proc_watchdog(struct ctl_table *table, int write,
675 void *buffer, size_t *lenp, loff_t *ppos)
677 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
678 table, write, buffer, lenp, ppos);
682 * /proc/sys/kernel/nmi_watchdog
684 int proc_nmi_watchdog(struct ctl_table *table, int write,
685 void *buffer, size_t *lenp, loff_t *ppos)
687 if (!nmi_watchdog_available && write)
689 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
690 table, write, buffer, lenp, ppos);
694 * /proc/sys/kernel/soft_watchdog
696 int proc_soft_watchdog(struct ctl_table *table, int write,
697 void *buffer, size_t *lenp, loff_t *ppos)
699 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
700 table, write, buffer, lenp, ppos);
704 * /proc/sys/kernel/watchdog_thresh
706 int proc_watchdog_thresh(struct ctl_table *table, int write,
707 void *buffer, size_t *lenp, loff_t *ppos)
711 mutex_lock(&watchdog_mutex);
713 old = READ_ONCE(watchdog_thresh);
714 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
716 if (!err && write && old != READ_ONCE(watchdog_thresh))
717 proc_watchdog_update();
719 mutex_unlock(&watchdog_mutex);
724 * The cpumask is the mask of possible cpus that the watchdog can run
725 * on, not the mask of cpus it is actually running on. This allows the
726 * user to specify a mask that will include cpus that have not yet
727 * been brought online, if desired.
729 int proc_watchdog_cpumask(struct ctl_table *table, int write,
730 void *buffer, size_t *lenp, loff_t *ppos)
734 mutex_lock(&watchdog_mutex);
736 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
738 proc_watchdog_update();
740 mutex_unlock(&watchdog_mutex);
743 #endif /* CONFIG_SYSCTL */
745 void __init lockup_detector_init(void)
747 if (tick_nohz_full_enabled())
748 pr_info("Disabling watchdog on nohz_full cores by default\n");
750 cpumask_copy(&watchdog_cpumask,
751 housekeeping_cpumask(HK_FLAG_TIMER));
753 if (!watchdog_nmi_probe())
754 nmi_watchdog_available = true;
755 lockup_detector_setup();