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 softlockup_panic_setup(char *str)
181 softlockup_panic = simple_strtoul(str, NULL, 0);
184 __setup("softlockup_panic=", softlockup_panic_setup);
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 /* Commands for resetting the watchdog */
245 static void update_touch_ts(void)
247 __this_cpu_write(watchdog_touch_ts, get_timestamp());
251 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
253 * Call when the scheduler may have stalled for legitimate reasons
254 * preventing the watchdog task from executing - e.g. the scheduler
255 * entering idle state. This should only be used for scheduler events.
256 * Use touch_softlockup_watchdog() for everything else.
258 notrace void touch_softlockup_watchdog_sched(void)
261 * Preemption can be enabled. It doesn't matter which CPU's timestamp
262 * gets zeroed here, so use the raw_ operation.
264 raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
267 notrace void touch_softlockup_watchdog(void)
269 touch_softlockup_watchdog_sched();
270 wq_watchdog_touch(raw_smp_processor_id());
272 EXPORT_SYMBOL(touch_softlockup_watchdog);
274 void touch_all_softlockup_watchdogs(void)
279 * watchdog_mutex cannpt be taken here, as this might be called
280 * from (soft)interrupt context, so the access to
281 * watchdog_allowed_cpumask might race with a concurrent update.
283 * The watchdog time stamp can race against a concurrent real
284 * update as well, the only side effect might be a cycle delay for
285 * the softlockup check.
287 for_each_cpu(cpu, &watchdog_allowed_mask)
288 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
289 wq_watchdog_touch(-1);
292 void touch_softlockup_watchdog_sync(void)
294 __this_cpu_write(softlockup_touch_sync, true);
295 __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
298 static int is_softlockup(unsigned long touch_ts)
300 unsigned long now = get_timestamp();
302 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
303 /* Warn about unreasonable delays. */
304 if (time_after(now, touch_ts + get_softlockup_thresh()))
305 return now - touch_ts;
310 /* watchdog detector functions */
311 bool is_hardlockup(void)
313 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
315 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
318 __this_cpu_write(hrtimer_interrupts_saved, hrint);
322 static void watchdog_interrupt_count(void)
324 __this_cpu_inc(hrtimer_interrupts);
327 static DEFINE_PER_CPU(struct completion, softlockup_completion);
328 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
331 * The watchdog thread function - touches the timestamp.
333 * It only runs once every sample_period seconds (4 seconds by
334 * default) to reset the softlockup timestamp. If this gets delayed
335 * for more than 2*watchdog_thresh seconds then the debug-printout
336 * triggers in watchdog_timer_fn().
338 static int softlockup_fn(void *data)
341 complete(this_cpu_ptr(&softlockup_completion));
346 /* watchdog kicker functions */
347 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
349 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
350 struct pt_regs *regs = get_irq_regs();
352 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
354 if (!watchdog_enabled)
355 return HRTIMER_NORESTART;
357 /* kick the hardlockup detector */
358 watchdog_interrupt_count();
360 /* kick the softlockup detector */
361 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
362 reinit_completion(this_cpu_ptr(&softlockup_completion));
363 stop_one_cpu_nowait(smp_processor_id(),
365 this_cpu_ptr(&softlockup_stop_work));
369 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
371 if (touch_ts == SOFTLOCKUP_RESET) {
372 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
374 * If the time stamp was touched atomically
375 * make sure the scheduler tick is up to date.
377 __this_cpu_write(softlockup_touch_sync, false);
381 /* Clear the guest paused flag on watchdog reset */
382 kvm_check_and_clear_guest_paused();
384 return HRTIMER_RESTART;
387 /* check for a softlockup
388 * This is done by making sure a high priority task is
389 * being scheduled. The task touches the watchdog to
390 * indicate it is getting cpu time. If it hasn't then
391 * this is a good indication some task is hogging the cpu
393 duration = is_softlockup(touch_ts);
394 if (unlikely(duration)) {
396 * If a virtual machine is stopped by the host it can look to
397 * the watchdog like a soft lockup, check to see if the host
398 * stopped the vm before we issue the warning
400 if (kvm_check_and_clear_guest_paused())
401 return HRTIMER_RESTART;
404 * Prevent multiple soft-lockup reports if one cpu is already
405 * engaged in dumping all cpu back traces.
407 if (softlockup_all_cpu_backtrace) {
408 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
409 return HRTIMER_RESTART;
412 /* Start period for the next softlockup warning. */
415 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
416 smp_processor_id(), duration,
417 current->comm, task_pid_nr(current));
419 print_irqtrace_events(current);
425 if (softlockup_all_cpu_backtrace) {
426 trigger_allbutself_cpu_backtrace();
427 clear_bit_unlock(0, &soft_lockup_nmi_warn);
430 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
431 if (softlockup_panic)
432 panic("softlockup: hung tasks");
435 return HRTIMER_RESTART;
438 static void watchdog_enable(unsigned int cpu)
440 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
441 struct completion *done = this_cpu_ptr(&softlockup_completion);
443 WARN_ON_ONCE(cpu != smp_processor_id());
445 init_completion(done);
449 * Start the timer first to prevent the NMI watchdog triggering
450 * before the timer has a chance to fire.
452 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
453 hrtimer->function = watchdog_timer_fn;
454 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
455 HRTIMER_MODE_REL_PINNED_HARD);
457 /* Initialize timestamp */
459 /* Enable the perf event */
460 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
461 watchdog_nmi_enable(cpu);
464 static void watchdog_disable(unsigned int cpu)
466 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
468 WARN_ON_ONCE(cpu != smp_processor_id());
471 * Disable the perf event first. That prevents that a large delay
472 * between disabling the timer and disabling the perf event causes
473 * the perf NMI to detect a false positive.
475 watchdog_nmi_disable(cpu);
476 hrtimer_cancel(hrtimer);
477 wait_for_completion(this_cpu_ptr(&softlockup_completion));
480 static int softlockup_stop_fn(void *data)
482 watchdog_disable(smp_processor_id());
486 static void softlockup_stop_all(void)
490 if (!softlockup_initialized)
493 for_each_cpu(cpu, &watchdog_allowed_mask)
494 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
496 cpumask_clear(&watchdog_allowed_mask);
499 static int softlockup_start_fn(void *data)
501 watchdog_enable(smp_processor_id());
505 static void softlockup_start_all(void)
509 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
510 for_each_cpu(cpu, &watchdog_allowed_mask)
511 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
514 int lockup_detector_online_cpu(unsigned int cpu)
516 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
517 watchdog_enable(cpu);
521 int lockup_detector_offline_cpu(unsigned int cpu)
523 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
524 watchdog_disable(cpu);
528 static void __lockup_detector_reconfigure(void)
533 softlockup_stop_all();
535 lockup_detector_update_enable();
536 if (watchdog_enabled && watchdog_thresh)
537 softlockup_start_all();
539 watchdog_nmi_start();
542 * Must be called outside the cpus locked section to prevent
543 * recursive locking in the perf code.
545 __lockup_detector_cleanup();
548 void lockup_detector_reconfigure(void)
550 mutex_lock(&watchdog_mutex);
551 __lockup_detector_reconfigure();
552 mutex_unlock(&watchdog_mutex);
556 * Create the watchdog thread infrastructure and configure the detector(s).
558 * The threads are not unparked as watchdog_allowed_mask is empty. When
559 * the threads are successfully initialized, take the proper locks and
560 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
562 static __init void lockup_detector_setup(void)
565 * If sysctl is off and watchdog got disabled on the command line,
566 * nothing to do here.
568 lockup_detector_update_enable();
570 if (!IS_ENABLED(CONFIG_SYSCTL) &&
571 !(watchdog_enabled && watchdog_thresh))
574 mutex_lock(&watchdog_mutex);
575 __lockup_detector_reconfigure();
576 softlockup_initialized = true;
577 mutex_unlock(&watchdog_mutex);
580 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
581 static void __lockup_detector_reconfigure(void)
585 lockup_detector_update_enable();
586 watchdog_nmi_start();
589 void lockup_detector_reconfigure(void)
591 __lockup_detector_reconfigure();
593 static inline void lockup_detector_setup(void)
595 __lockup_detector_reconfigure();
597 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
599 static void __lockup_detector_cleanup(void)
601 lockdep_assert_held(&watchdog_mutex);
602 hardlockup_detector_perf_cleanup();
606 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
608 * Caller must not hold the cpu hotplug rwsem.
610 void lockup_detector_cleanup(void)
612 mutex_lock(&watchdog_mutex);
613 __lockup_detector_cleanup();
614 mutex_unlock(&watchdog_mutex);
618 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
620 * Special interface for parisc. It prevents lockup detector warnings from
621 * the default pm_poweroff() function which busy loops forever.
623 void lockup_detector_soft_poweroff(void)
625 watchdog_enabled = 0;
630 /* Propagate any changes to the watchdog threads */
631 static void proc_watchdog_update(void)
633 /* Remove impossible cpus to keep sysctl output clean. */
634 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
635 __lockup_detector_reconfigure();
639 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
641 * caller | table->data points to | 'which'
642 * -------------------|----------------------------|--------------------------
643 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
644 * | | SOFT_WATCHDOG_ENABLED
645 * -------------------|----------------------------|--------------------------
646 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
647 * -------------------|----------------------------|--------------------------
648 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
650 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
651 void *buffer, size_t *lenp, loff_t *ppos)
653 int err, old, *param = table->data;
655 mutex_lock(&watchdog_mutex);
659 * On read synchronize the userspace interface. This is a
662 *param = (watchdog_enabled & which) != 0;
663 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
665 old = READ_ONCE(*param);
666 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
667 if (!err && old != READ_ONCE(*param))
668 proc_watchdog_update();
670 mutex_unlock(&watchdog_mutex);
675 * /proc/sys/kernel/watchdog
677 int proc_watchdog(struct ctl_table *table, int write,
678 void *buffer, size_t *lenp, loff_t *ppos)
680 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
681 table, write, buffer, lenp, ppos);
685 * /proc/sys/kernel/nmi_watchdog
687 int proc_nmi_watchdog(struct ctl_table *table, int write,
688 void *buffer, size_t *lenp, loff_t *ppos)
690 if (!nmi_watchdog_available && write)
692 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
693 table, write, buffer, lenp, ppos);
697 * /proc/sys/kernel/soft_watchdog
699 int proc_soft_watchdog(struct ctl_table *table, int write,
700 void *buffer, size_t *lenp, loff_t *ppos)
702 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
703 table, write, buffer, lenp, ppos);
707 * /proc/sys/kernel/watchdog_thresh
709 int proc_watchdog_thresh(struct ctl_table *table, int write,
710 void *buffer, size_t *lenp, loff_t *ppos)
714 mutex_lock(&watchdog_mutex);
716 old = READ_ONCE(watchdog_thresh);
717 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
719 if (!err && write && old != READ_ONCE(watchdog_thresh))
720 proc_watchdog_update();
722 mutex_unlock(&watchdog_mutex);
727 * The cpumask is the mask of possible cpus that the watchdog can run
728 * on, not the mask of cpus it is actually running on. This allows the
729 * user to specify a mask that will include cpus that have not yet
730 * been brought online, if desired.
732 int proc_watchdog_cpumask(struct ctl_table *table, int write,
733 void *buffer, size_t *lenp, loff_t *ppos)
737 mutex_lock(&watchdog_mutex);
739 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
741 proc_watchdog_update();
743 mutex_unlock(&watchdog_mutex);
746 #endif /* CONFIG_SYSCTL */
748 void __init lockup_detector_init(void)
750 if (tick_nohz_full_enabled())
751 pr_info("Disabling watchdog on nohz_full cores by default\n");
753 cpumask_copy(&watchdog_cpumask,
754 housekeeping_cpumask(HK_FLAG_TIMER));
756 if (!watchdog_nmi_probe())
757 nmi_watchdog_available = true;
758 lockup_detector_setup();