Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
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.
11 */
12
13#define pr_fmt(fmt) "watchdog: " fmt
14
15#include <linux/mm.h>
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>
26
27#include <asm/irq_regs.h>
28#include <linux/kvm_para.h>
29
30static DEFINE_MUTEX(watchdog_mutex);
31
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
35#else
36# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37# define NMI_WATCHDOG_DEFAULT 0
38#endif
39
40unsigned long __read_mostly watchdog_enabled;
41int __read_mostly watchdog_user_enabled = 1;
42int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43int __read_mostly soft_watchdog_user_enabled = 1;
44int __read_mostly watchdog_thresh = 10;
45static int __read_mostly nmi_watchdog_available;
46
47struct cpumask watchdog_cpumask __read_mostly;
48unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
49
50#ifdef CONFIG_HARDLOCKUP_DETECTOR
51
52# ifdef CONFIG_SMP
53int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
54# endif /* CONFIG_SMP */
55
56/*
57 * Should we panic when a soft-lockup or hard-lockup occurs:
58 */
59unsigned int __read_mostly hardlockup_panic =
60 IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
61/*
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().
68 */
69void __init hardlockup_detector_disable(void)
70{
71 nmi_watchdog_user_enabled = 0;
72}
73
74static int __init hardlockup_panic_setup(char *str)
75{
76 if (!strncmp(str, "panic", 5))
77 hardlockup_panic = 1;
78 else if (!strncmp(str, "nopanic", 7))
79 hardlockup_panic = 0;
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;
84 return 1;
85}
86__setup("nmi_watchdog=", hardlockup_panic_setup);
87
88#endif /* CONFIG_HARDLOCKUP_DETECTOR */
89
90/*
91 * These functions can be overridden if an architecture implements its
92 * own hardlockup detector.
93 *
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.
97 */
98int __weak watchdog_nmi_enable(unsigned int cpu)
99{
100 hardlockup_detector_perf_enable();
101 return 0;
102}
103
104void __weak watchdog_nmi_disable(unsigned int cpu)
105{
106 hardlockup_detector_perf_disable();
107}
108
109/* Return 0, if a NMI watchdog is available. Error code otherwise */
110int __weak __init watchdog_nmi_probe(void)
111{
112 return hardlockup_detector_perf_init();
113}
114
115/**
116 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
117 *
118 * The reconfiguration steps are:
119 * watchdog_nmi_stop();
120 * update_variables();
121 * watchdog_nmi_start();
122 */
123void __weak watchdog_nmi_stop(void) { }
124
125/**
126 * watchdog_nmi_start - Start the watchdog after reconfiguration
127 *
128 * Counterpart to watchdog_nmi_stop().
129 *
130 * The following variables have been updated in update_variables() and
131 * contain the currently valid configuration:
132 * - watchdog_enabled
133 * - watchdog_thresh
134 * - watchdog_cpumask
135 */
136void __weak watchdog_nmi_start(void) { }
137
138/**
139 * lockup_detector_update_enable - Update the sysctl enable bit
140 *
141 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
142 * can't race with watchdog_nmi_disable().
143 */
144static void lockup_detector_update_enable(void)
145{
146 watchdog_enabled = 0;
147 if (!watchdog_user_enabled)
148 return;
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;
153}
154
155#ifdef CONFIG_SOFTLOCKUP_DETECTOR
156
157/*
158 * Delay the soflockup report when running a known slow code.
159 * It does _not_ affect the timestamp of the last successdul reschedule.
160 */
161#define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
162
163#ifdef CONFIG_SMP
164int __read_mostly sysctl_softlockup_all_cpu_backtrace;
165#endif
166
167static struct cpumask watchdog_allowed_mask __read_mostly;
168
169/* Global variables, exported for sysctl */
170unsigned int __read_mostly softlockup_panic =
171 IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
172
173static bool softlockup_initialized __read_mostly;
174static u64 __read_mostly sample_period;
175
176/* Timestamp taken after the last successful reschedule. */
177static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
178/* Timestamp of the last softlockup report. */
179static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
180static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
181static DEFINE_PER_CPU(bool, softlockup_touch_sync);
182static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
183static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
184static unsigned long soft_lockup_nmi_warn;
185
186static int __init nowatchdog_setup(char *str)
187{
188 watchdog_user_enabled = 0;
189 return 1;
190}
191__setup("nowatchdog", nowatchdog_setup);
192
193static int __init nosoftlockup_setup(char *str)
194{
195 soft_watchdog_user_enabled = 0;
196 return 1;
197}
198__setup("nosoftlockup", nosoftlockup_setup);
199
200static int __init watchdog_thresh_setup(char *str)
201{
202 get_option(&str, &watchdog_thresh);
203 return 1;
204}
205__setup("watchdog_thresh=", watchdog_thresh_setup);
206
207static void __lockup_detector_cleanup(void);
208
209/*
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.
215 */
216static int get_softlockup_thresh(void)
217{
218 return watchdog_thresh * 2;
219}
220
221/*
222 * Returns seconds, approximately. We don't need nanosecond
223 * resolution, and we don't need to waste time with a big divide when
224 * 2^30ns == 1.074s.
225 */
226static unsigned long get_timestamp(void)
227{
228 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
229}
230
231static void set_sample_period(void)
232{
233 /*
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
239 */
240 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
241 watchdog_update_hrtimer_threshold(sample_period);
242}
243
244static void update_report_ts(void)
245{
246 __this_cpu_write(watchdog_report_ts, get_timestamp());
247}
248
249/* Commands for resetting the watchdog */
250static void update_touch_ts(void)
251{
252 __this_cpu_write(watchdog_touch_ts, get_timestamp());
253 update_report_ts();
254}
255
256/**
257 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
258 *
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.
263 */
264notrace void touch_softlockup_watchdog_sched(void)
265{
266 /*
267 * Preemption can be enabled. It doesn't matter which CPU's watchdog
268 * report period gets restarted here, so use the raw_ operation.
269 */
270 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
271}
272
273notrace void touch_softlockup_watchdog(void)
274{
275 touch_softlockup_watchdog_sched();
276 wq_watchdog_touch(raw_smp_processor_id());
277}
278EXPORT_SYMBOL(touch_softlockup_watchdog);
279
280void touch_all_softlockup_watchdogs(void)
281{
282 int cpu;
283
284 /*
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.
288 *
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.
292 */
293 for_each_cpu(cpu, &watchdog_allowed_mask) {
294 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
295 wq_watchdog_touch(cpu);
296 }
297}
298
299void touch_softlockup_watchdog_sync(void)
300{
301 __this_cpu_write(softlockup_touch_sync, true);
302 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
303}
304
305static int is_softlockup(unsigned long touch_ts,
306 unsigned long period_ts,
307 unsigned long now)
308{
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;
313 }
314 return 0;
315}
316
317/* watchdog detector functions */
318bool is_hardlockup(void)
319{
320 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
321
322 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
323 return true;
324
325 __this_cpu_write(hrtimer_interrupts_saved, hrint);
326 return false;
327}
328
329static void watchdog_interrupt_count(void)
330{
331 __this_cpu_inc(hrtimer_interrupts);
332}
333
334static DEFINE_PER_CPU(struct completion, softlockup_completion);
335static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
336
337/*
338 * The watchdog feed function - touches the timestamp.
339 *
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().
344 */
345static int softlockup_fn(void *data)
346{
347 update_touch_ts();
348 complete(this_cpu_ptr(&softlockup_completion));
349
350 return 0;
351}
352
353/* watchdog kicker functions */
354static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
355{
356 unsigned long touch_ts, period_ts, now;
357 struct pt_regs *regs = get_irq_regs();
358 int duration;
359 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
360
361 if (!watchdog_enabled)
362 return HRTIMER_NORESTART;
363
364 /* kick the hardlockup detector */
365 watchdog_interrupt_count();
366
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(),
371 softlockup_fn, NULL,
372 this_cpu_ptr(&softlockup_stop_work));
373 }
374
375 /* .. and repeat */
376 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
377
378 /*
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.
382 */
383 now = get_timestamp();
384 /*
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.
387 */
388 kvm_check_and_clear_guest_paused();
389 /*
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.
393 */
394 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
395
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))) {
399 /*
400 * If the time stamp was touched atomically
401 * make sure the scheduler tick is up to date.
402 */
403 __this_cpu_write(softlockup_touch_sync, false);
404 sched_clock_tick();
405 }
406
407 update_report_ts();
408 return HRTIMER_RESTART;
409 }
410
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)) {
415 /*
416 * Prevent multiple soft-lockup reports if one cpu is already
417 * engaged in dumping all cpu back traces.
418 */
419 if (softlockup_all_cpu_backtrace) {
420 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
421 return HRTIMER_RESTART;
422 }
423
424 /* Start period for the next softlockup warning. */
425 update_report_ts();
426
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));
430 print_modules();
431 print_irqtrace_events(current);
432 if (regs)
433 show_regs(regs);
434 else
435 dump_stack();
436
437 if (softlockup_all_cpu_backtrace) {
438 trigger_allbutself_cpu_backtrace();
439 clear_bit_unlock(0, &soft_lockup_nmi_warn);
440 }
441
442 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
443 if (softlockup_panic)
444 panic("softlockup: hung tasks");
445 }
446
447 return HRTIMER_RESTART;
448}
449
450static void watchdog_enable(unsigned int cpu)
451{
452 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
453 struct completion *done = this_cpu_ptr(&softlockup_completion);
454
455 WARN_ON_ONCE(cpu != smp_processor_id());
456
457 init_completion(done);
458 complete(done);
459
460 /*
461 * Start the timer first to prevent the NMI watchdog triggering
462 * before the timer has a chance to fire.
463 */
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);
468
469 /* Initialize timestamp */
470 update_touch_ts();
471 /* Enable the perf event */
472 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
473 watchdog_nmi_enable(cpu);
474}
475
476static void watchdog_disable(unsigned int cpu)
477{
478 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
479
480 WARN_ON_ONCE(cpu != smp_processor_id());
481
482 /*
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.
486 */
487 watchdog_nmi_disable(cpu);
488 hrtimer_cancel(hrtimer);
489 wait_for_completion(this_cpu_ptr(&softlockup_completion));
490}
491
492static int softlockup_stop_fn(void *data)
493{
494 watchdog_disable(smp_processor_id());
495 return 0;
496}
497
498static void softlockup_stop_all(void)
499{
500 int cpu;
501
502 if (!softlockup_initialized)
503 return;
504
505 for_each_cpu(cpu, &watchdog_allowed_mask)
506 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
507
508 cpumask_clear(&watchdog_allowed_mask);
509}
510
511static int softlockup_start_fn(void *data)
512{
513 watchdog_enable(smp_processor_id());
514 return 0;
515}
516
517static void softlockup_start_all(void)
518{
519 int cpu;
520
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);
524}
525
526int lockup_detector_online_cpu(unsigned int cpu)
527{
528 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
529 watchdog_enable(cpu);
530 return 0;
531}
532
533int lockup_detector_offline_cpu(unsigned int cpu)
534{
535 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
536 watchdog_disable(cpu);
537 return 0;
538}
539
540static void __lockup_detector_reconfigure(void)
541{
542 cpus_read_lock();
543 watchdog_nmi_stop();
544
545 softlockup_stop_all();
546 set_sample_period();
547 lockup_detector_update_enable();
548 if (watchdog_enabled && watchdog_thresh)
549 softlockup_start_all();
550
551 watchdog_nmi_start();
552 cpus_read_unlock();
553 /*
554 * Must be called outside the cpus locked section to prevent
555 * recursive locking in the perf code.
556 */
557 __lockup_detector_cleanup();
558}
559
560void lockup_detector_reconfigure(void)
561{
562 mutex_lock(&watchdog_mutex);
563 __lockup_detector_reconfigure();
564 mutex_unlock(&watchdog_mutex);
565}
566
567/*
568 * Create the watchdog infrastructure and configure the detector(s).
569 */
570static __init void lockup_detector_setup(void)
571{
572 /*
573 * If sysctl is off and watchdog got disabled on the command line,
574 * nothing to do here.
575 */
576 lockup_detector_update_enable();
577
578 if (!IS_ENABLED(CONFIG_SYSCTL) &&
579 !(watchdog_enabled && watchdog_thresh))
580 return;
581
582 mutex_lock(&watchdog_mutex);
583 __lockup_detector_reconfigure();
584 softlockup_initialized = true;
585 mutex_unlock(&watchdog_mutex);
586}
587
588#else /* CONFIG_SOFTLOCKUP_DETECTOR */
589static void __lockup_detector_reconfigure(void)
590{
591 cpus_read_lock();
592 watchdog_nmi_stop();
593 lockup_detector_update_enable();
594 watchdog_nmi_start();
595 cpus_read_unlock();
596}
597void lockup_detector_reconfigure(void)
598{
599 __lockup_detector_reconfigure();
600}
601static inline void lockup_detector_setup(void)
602{
603 __lockup_detector_reconfigure();
604}
605#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
606
607static void __lockup_detector_cleanup(void)
608{
609 lockdep_assert_held(&watchdog_mutex);
610 hardlockup_detector_perf_cleanup();
611}
612
613/**
614 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
615 *
616 * Caller must not hold the cpu hotplug rwsem.
617 */
618void lockup_detector_cleanup(void)
619{
620 mutex_lock(&watchdog_mutex);
621 __lockup_detector_cleanup();
622 mutex_unlock(&watchdog_mutex);
623}
624
625/**
626 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
627 *
628 * Special interface for parisc. It prevents lockup detector warnings from
629 * the default pm_poweroff() function which busy loops forever.
630 */
631void lockup_detector_soft_poweroff(void)
632{
633 watchdog_enabled = 0;
634}
635
636#ifdef CONFIG_SYSCTL
637
638/* Propagate any changes to the watchdog infrastructure */
639static void proc_watchdog_update(void)
640{
641 /* Remove impossible cpus to keep sysctl output clean. */
642 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
643 __lockup_detector_reconfigure();
644}
645
646/*
647 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
648 *
649 * caller | table->data points to | 'which'
650 * -------------------|----------------------------|--------------------------
651 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
652 * | | SOFT_WATCHDOG_ENABLED
653 * -------------------|----------------------------|--------------------------
654 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
655 * -------------------|----------------------------|--------------------------
656 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
657 */
658static int proc_watchdog_common(int which, struct ctl_table *table, int write,
659 void *buffer, size_t *lenp, loff_t *ppos)
660{
661 int err, old, *param = table->data;
662
663 mutex_lock(&watchdog_mutex);
664
665 if (!write) {
666 /*
667 * On read synchronize the userspace interface. This is a
668 * racy snapshot.
669 */
670 *param = (watchdog_enabled & which) != 0;
671 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
672 } else {
673 old = READ_ONCE(*param);
674 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
675 if (!err && old != READ_ONCE(*param))
676 proc_watchdog_update();
677 }
678 mutex_unlock(&watchdog_mutex);
679 return err;
680}
681
682/*
683 * /proc/sys/kernel/watchdog
684 */
685int proc_watchdog(struct ctl_table *table, int write,
686 void *buffer, size_t *lenp, loff_t *ppos)
687{
688 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
689 table, write, buffer, lenp, ppos);
690}
691
692/*
693 * /proc/sys/kernel/nmi_watchdog
694 */
695int proc_nmi_watchdog(struct ctl_table *table, int write,
696 void *buffer, size_t *lenp, loff_t *ppos)
697{
698 if (!nmi_watchdog_available && write)
699 return -ENOTSUPP;
700 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
701 table, write, buffer, lenp, ppos);
702}
703
704/*
705 * /proc/sys/kernel/soft_watchdog
706 */
707int proc_soft_watchdog(struct ctl_table *table, int write,
708 void *buffer, size_t *lenp, loff_t *ppos)
709{
710 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
711 table, write, buffer, lenp, ppos);
712}
713
714/*
715 * /proc/sys/kernel/watchdog_thresh
716 */
717int proc_watchdog_thresh(struct ctl_table *table, int write,
718 void *buffer, size_t *lenp, loff_t *ppos)
719{
720 int err, old;
721
722 mutex_lock(&watchdog_mutex);
723
724 old = READ_ONCE(watchdog_thresh);
725 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
726
727 if (!err && write && old != READ_ONCE(watchdog_thresh))
728 proc_watchdog_update();
729
730 mutex_unlock(&watchdog_mutex);
731 return err;
732}
733
734/*
735 * The cpumask is the mask of possible cpus that the watchdog can run
736 * on, not the mask of cpus it is actually running on. This allows the
737 * user to specify a mask that will include cpus that have not yet
738 * been brought online, if desired.
739 */
740int proc_watchdog_cpumask(struct ctl_table *table, int write,
741 void *buffer, size_t *lenp, loff_t *ppos)
742{
743 int err;
744
745 mutex_lock(&watchdog_mutex);
746
747 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
748 if (!err && write)
749 proc_watchdog_update();
750
751 mutex_unlock(&watchdog_mutex);
752 return err;
753}
754
755static const int sixty = 60;
756
757static struct ctl_table watchdog_sysctls[] = {
758 {
759 .procname = "watchdog",
760 .data = &watchdog_user_enabled,
761 .maxlen = sizeof(int),
762 .mode = 0644,
763 .proc_handler = proc_watchdog,
764 .extra1 = SYSCTL_ZERO,
765 .extra2 = SYSCTL_ONE,
766 },
767 {
768 .procname = "watchdog_thresh",
769 .data = &watchdog_thresh,
770 .maxlen = sizeof(int),
771 .mode = 0644,
772 .proc_handler = proc_watchdog_thresh,
773 .extra1 = SYSCTL_ZERO,
774 .extra2 = (void *)&sixty,
775 },
776 {
777 .procname = "nmi_watchdog",
778 .data = &nmi_watchdog_user_enabled,
779 .maxlen = sizeof(int),
780 .mode = NMI_WATCHDOG_SYSCTL_PERM,
781 .proc_handler = proc_nmi_watchdog,
782 .extra1 = SYSCTL_ZERO,
783 .extra2 = SYSCTL_ONE,
784 },
785 {
786 .procname = "watchdog_cpumask",
787 .data = &watchdog_cpumask_bits,
788 .maxlen = NR_CPUS,
789 .mode = 0644,
790 .proc_handler = proc_watchdog_cpumask,
791 },
792#ifdef CONFIG_SOFTLOCKUP_DETECTOR
793 {
794 .procname = "soft_watchdog",
795 .data = &soft_watchdog_user_enabled,
796 .maxlen = sizeof(int),
797 .mode = 0644,
798 .proc_handler = proc_soft_watchdog,
799 .extra1 = SYSCTL_ZERO,
800 .extra2 = SYSCTL_ONE,
801 },
802 {
803 .procname = "softlockup_panic",
804 .data = &softlockup_panic,
805 .maxlen = sizeof(int),
806 .mode = 0644,
807 .proc_handler = proc_dointvec_minmax,
808 .extra1 = SYSCTL_ZERO,
809 .extra2 = SYSCTL_ONE,
810 },
811#ifdef CONFIG_SMP
812 {
813 .procname = "softlockup_all_cpu_backtrace",
814 .data = &sysctl_softlockup_all_cpu_backtrace,
815 .maxlen = sizeof(int),
816 .mode = 0644,
817 .proc_handler = proc_dointvec_minmax,
818 .extra1 = SYSCTL_ZERO,
819 .extra2 = SYSCTL_ONE,
820 },
821#endif /* CONFIG_SMP */
822#endif
823#ifdef CONFIG_HARDLOCKUP_DETECTOR
824 {
825 .procname = "hardlockup_panic",
826 .data = &hardlockup_panic,
827 .maxlen = sizeof(int),
828 .mode = 0644,
829 .proc_handler = proc_dointvec_minmax,
830 .extra1 = SYSCTL_ZERO,
831 .extra2 = SYSCTL_ONE,
832 },
833#ifdef CONFIG_SMP
834 {
835 .procname = "hardlockup_all_cpu_backtrace",
836 .data = &sysctl_hardlockup_all_cpu_backtrace,
837 .maxlen = sizeof(int),
838 .mode = 0644,
839 .proc_handler = proc_dointvec_minmax,
840 .extra1 = SYSCTL_ZERO,
841 .extra2 = SYSCTL_ONE,
842 },
843#endif /* CONFIG_SMP */
844#endif
845 {}
846};
847
848static void __init watchdog_sysctl_init(void)
849{
850 register_sysctl_init("kernel", watchdog_sysctls);
851}
852#else
853#define watchdog_sysctl_init() do { } while (0)
854#endif /* CONFIG_SYSCTL */
855
856void __init lockup_detector_init(void)
857{
858 if (tick_nohz_full_enabled())
859 pr_info("Disabling watchdog on nohz_full cores by default\n");
860
861 cpumask_copy(&watchdog_cpumask,
862 housekeeping_cpumask(HK_TYPE_TIMER));
863
864 if (!watchdog_nmi_probe())
865 nmi_watchdog_available = true;
866 lockup_detector_setup();
867 watchdog_sysctl_init();
868}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
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.
11 */
12
13#define pr_fmt(fmt) "watchdog: " fmt
14
15#include <linux/mm.h>
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>
26
27#include <asm/irq_regs.h>
28#include <linux/kvm_para.h>
29
30static DEFINE_MUTEX(watchdog_mutex);
31
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
35#else
36# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37# define NMI_WATCHDOG_DEFAULT 0
38#endif
39
40unsigned long __read_mostly watchdog_enabled;
41int __read_mostly watchdog_user_enabled = 1;
42int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43int __read_mostly soft_watchdog_user_enabled = 1;
44int __read_mostly watchdog_thresh = 10;
45static int __read_mostly nmi_watchdog_available;
46
47static struct cpumask watchdog_allowed_mask __read_mostly;
48
49struct cpumask watchdog_cpumask __read_mostly;
50unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
51
52#ifdef CONFIG_HARDLOCKUP_DETECTOR
53/*
54 * Should we panic when a soft-lockup or hard-lockup occurs:
55 */
56unsigned int __read_mostly hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
58/*
59 * We may not want to enable hard lockup detection by default in all cases,
60 * for example when running the kernel as a guest on a hypervisor. In these
61 * cases this function can be called to disable hard lockup detection. This
62 * function should only be executed once by the boot processor before the
63 * kernel command line parameters are parsed, because otherwise it is not
64 * possible to override this in hardlockup_panic_setup().
65 */
66void __init hardlockup_detector_disable(void)
67{
68 nmi_watchdog_user_enabled = 0;
69}
70
71static int __init hardlockup_panic_setup(char *str)
72{
73 if (!strncmp(str, "panic", 5))
74 hardlockup_panic = 1;
75 else if (!strncmp(str, "nopanic", 7))
76 hardlockup_panic = 0;
77 else if (!strncmp(str, "0", 1))
78 nmi_watchdog_user_enabled = 0;
79 else if (!strncmp(str, "1", 1))
80 nmi_watchdog_user_enabled = 1;
81 return 1;
82}
83__setup("nmi_watchdog=", hardlockup_panic_setup);
84
85# ifdef CONFIG_SMP
86int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
87
88static int __init hardlockup_all_cpu_backtrace_setup(char *str)
89{
90 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
91 return 1;
92}
93__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
94# endif /* CONFIG_SMP */
95#endif /* CONFIG_HARDLOCKUP_DETECTOR */
96
97/*
98 * These functions can be overridden if an architecture implements its
99 * own hardlockup detector.
100 *
101 * watchdog_nmi_enable/disable can be implemented to start and stop when
102 * softlockup watchdog threads start and stop. The arch must select the
103 * SOFTLOCKUP_DETECTOR Kconfig.
104 */
105int __weak watchdog_nmi_enable(unsigned int cpu)
106{
107 hardlockup_detector_perf_enable();
108 return 0;
109}
110
111void __weak watchdog_nmi_disable(unsigned int cpu)
112{
113 hardlockup_detector_perf_disable();
114}
115
116/* Return 0, if a NMI watchdog is available. Error code otherwise */
117int __weak __init watchdog_nmi_probe(void)
118{
119 return hardlockup_detector_perf_init();
120}
121
122/**
123 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
124 *
125 * The reconfiguration steps are:
126 * watchdog_nmi_stop();
127 * update_variables();
128 * watchdog_nmi_start();
129 */
130void __weak watchdog_nmi_stop(void) { }
131
132/**
133 * watchdog_nmi_start - Start the watchdog after reconfiguration
134 *
135 * Counterpart to watchdog_nmi_stop().
136 *
137 * The following variables have been updated in update_variables() and
138 * contain the currently valid configuration:
139 * - watchdog_enabled
140 * - watchdog_thresh
141 * - watchdog_cpumask
142 */
143void __weak watchdog_nmi_start(void) { }
144
145/**
146 * lockup_detector_update_enable - Update the sysctl enable bit
147 *
148 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
149 * can't race with watchdog_nmi_disable().
150 */
151static void lockup_detector_update_enable(void)
152{
153 watchdog_enabled = 0;
154 if (!watchdog_user_enabled)
155 return;
156 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
157 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
158 if (soft_watchdog_user_enabled)
159 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
160}
161
162#ifdef CONFIG_SOFTLOCKUP_DETECTOR
163
164/* Global variables, exported for sysctl */
165unsigned int __read_mostly softlockup_panic =
166 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
167
168static bool softlockup_initialized __read_mostly;
169static u64 __read_mostly sample_period;
170
171static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
172static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
173static DEFINE_PER_CPU(bool, softlockup_touch_sync);
174static DEFINE_PER_CPU(bool, soft_watchdog_warn);
175static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
176static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
177static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
178static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
179static unsigned long soft_lockup_nmi_warn;
180
181static int __init softlockup_panic_setup(char *str)
182{
183 softlockup_panic = simple_strtoul(str, NULL, 0);
184 return 1;
185}
186__setup("softlockup_panic=", softlockup_panic_setup);
187
188static int __init nowatchdog_setup(char *str)
189{
190 watchdog_user_enabled = 0;
191 return 1;
192}
193__setup("nowatchdog", nowatchdog_setup);
194
195static int __init nosoftlockup_setup(char *str)
196{
197 soft_watchdog_user_enabled = 0;
198 return 1;
199}
200__setup("nosoftlockup", nosoftlockup_setup);
201
202static int __init watchdog_thresh_setup(char *str)
203{
204 get_option(&str, &watchdog_thresh);
205 return 1;
206}
207__setup("watchdog_thresh=", watchdog_thresh_setup);
208
209#ifdef CONFIG_SMP
210int __read_mostly sysctl_softlockup_all_cpu_backtrace;
211
212static int __init softlockup_all_cpu_backtrace_setup(char *str)
213{
214 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
215 return 1;
216}
217__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
218#endif
219
220static void __lockup_detector_cleanup(void);
221
222/*
223 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
224 * lockups can have false positives under extreme conditions. So we generally
225 * want a higher threshold for soft lockups than for hard lockups. So we couple
226 * the thresholds with a factor: we make the soft threshold twice the amount of
227 * time the hard threshold is.
228 */
229static int get_softlockup_thresh(void)
230{
231 return watchdog_thresh * 2;
232}
233
234/*
235 * Returns seconds, approximately. We don't need nanosecond
236 * resolution, and we don't need to waste time with a big divide when
237 * 2^30ns == 1.074s.
238 */
239static unsigned long get_timestamp(void)
240{
241 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
242}
243
244static void set_sample_period(void)
245{
246 /*
247 * convert watchdog_thresh from seconds to ns
248 * the divide by 5 is to give hrtimer several chances (two
249 * or three with the current relation between the soft
250 * and hard thresholds) to increment before the
251 * hardlockup detector generates a warning
252 */
253 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
254 watchdog_update_hrtimer_threshold(sample_period);
255}
256
257/* Commands for resetting the watchdog */
258static void __touch_watchdog(void)
259{
260 __this_cpu_write(watchdog_touch_ts, get_timestamp());
261}
262
263/**
264 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
265 *
266 * Call when the scheduler may have stalled for legitimate reasons
267 * preventing the watchdog task from executing - e.g. the scheduler
268 * entering idle state. This should only be used for scheduler events.
269 * Use touch_softlockup_watchdog() for everything else.
270 */
271notrace void touch_softlockup_watchdog_sched(void)
272{
273 /*
274 * Preemption can be enabled. It doesn't matter which CPU's timestamp
275 * gets zeroed here, so use the raw_ operation.
276 */
277 raw_cpu_write(watchdog_touch_ts, 0);
278}
279
280notrace void touch_softlockup_watchdog(void)
281{
282 touch_softlockup_watchdog_sched();
283 wq_watchdog_touch(raw_smp_processor_id());
284}
285EXPORT_SYMBOL(touch_softlockup_watchdog);
286
287void touch_all_softlockup_watchdogs(void)
288{
289 int cpu;
290
291 /*
292 * watchdog_mutex cannpt be taken here, as this might be called
293 * from (soft)interrupt context, so the access to
294 * watchdog_allowed_cpumask might race with a concurrent update.
295 *
296 * The watchdog time stamp can race against a concurrent real
297 * update as well, the only side effect might be a cycle delay for
298 * the softlockup check.
299 */
300 for_each_cpu(cpu, &watchdog_allowed_mask)
301 per_cpu(watchdog_touch_ts, cpu) = 0;
302 wq_watchdog_touch(-1);
303}
304
305void touch_softlockup_watchdog_sync(void)
306{
307 __this_cpu_write(softlockup_touch_sync, true);
308 __this_cpu_write(watchdog_touch_ts, 0);
309}
310
311static int is_softlockup(unsigned long touch_ts)
312{
313 unsigned long now = get_timestamp();
314
315 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
316 /* Warn about unreasonable delays. */
317 if (time_after(now, touch_ts + get_softlockup_thresh()))
318 return now - touch_ts;
319 }
320 return 0;
321}
322
323/* watchdog detector functions */
324bool is_hardlockup(void)
325{
326 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
327
328 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
329 return true;
330
331 __this_cpu_write(hrtimer_interrupts_saved, hrint);
332 return false;
333}
334
335static void watchdog_interrupt_count(void)
336{
337 __this_cpu_inc(hrtimer_interrupts);
338}
339
340static DEFINE_PER_CPU(struct completion, softlockup_completion);
341static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
342
343/*
344 * The watchdog thread function - touches the timestamp.
345 *
346 * It only runs once every sample_period seconds (4 seconds by
347 * default) to reset the softlockup timestamp. If this gets delayed
348 * for more than 2*watchdog_thresh seconds then the debug-printout
349 * triggers in watchdog_timer_fn().
350 */
351static int softlockup_fn(void *data)
352{
353 __this_cpu_write(soft_lockup_hrtimer_cnt,
354 __this_cpu_read(hrtimer_interrupts));
355 __touch_watchdog();
356 complete(this_cpu_ptr(&softlockup_completion));
357
358 return 0;
359}
360
361/* watchdog kicker functions */
362static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
363{
364 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
365 struct pt_regs *regs = get_irq_regs();
366 int duration;
367 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
368
369 if (!watchdog_enabled)
370 return HRTIMER_NORESTART;
371
372 /* kick the hardlockup detector */
373 watchdog_interrupt_count();
374
375 /* kick the softlockup detector */
376 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
377 reinit_completion(this_cpu_ptr(&softlockup_completion));
378 stop_one_cpu_nowait(smp_processor_id(),
379 softlockup_fn, NULL,
380 this_cpu_ptr(&softlockup_stop_work));
381 }
382
383 /* .. and repeat */
384 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
385
386 if (touch_ts == 0) {
387 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
388 /*
389 * If the time stamp was touched atomically
390 * make sure the scheduler tick is up to date.
391 */
392 __this_cpu_write(softlockup_touch_sync, false);
393 sched_clock_tick();
394 }
395
396 /* Clear the guest paused flag on watchdog reset */
397 kvm_check_and_clear_guest_paused();
398 __touch_watchdog();
399 return HRTIMER_RESTART;
400 }
401
402 /* check for a softlockup
403 * This is done by making sure a high priority task is
404 * being scheduled. The task touches the watchdog to
405 * indicate it is getting cpu time. If it hasn't then
406 * this is a good indication some task is hogging the cpu
407 */
408 duration = is_softlockup(touch_ts);
409 if (unlikely(duration)) {
410 /*
411 * If a virtual machine is stopped by the host it can look to
412 * the watchdog like a soft lockup, check to see if the host
413 * stopped the vm before we issue the warning
414 */
415 if (kvm_check_and_clear_guest_paused())
416 return HRTIMER_RESTART;
417
418 /* only warn once */
419 if (__this_cpu_read(soft_watchdog_warn) == true) {
420 /*
421 * When multiple processes are causing softlockups the
422 * softlockup detector only warns on the first one
423 * because the code relies on a full quiet cycle to
424 * re-arm. The second process prevents the quiet cycle
425 * and never gets reported. Use task pointers to detect
426 * this.
427 */
428 if (__this_cpu_read(softlockup_task_ptr_saved) !=
429 current) {
430 __this_cpu_write(soft_watchdog_warn, false);
431 __touch_watchdog();
432 }
433 return HRTIMER_RESTART;
434 }
435
436 if (softlockup_all_cpu_backtrace) {
437 /* Prevent multiple soft-lockup reports if one cpu is already
438 * engaged in dumping cpu back traces
439 */
440 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
441 /* Someone else will report us. Let's give up */
442 __this_cpu_write(soft_watchdog_warn, true);
443 return HRTIMER_RESTART;
444 }
445 }
446
447 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
448 smp_processor_id(), duration,
449 current->comm, task_pid_nr(current));
450 __this_cpu_write(softlockup_task_ptr_saved, current);
451 print_modules();
452 print_irqtrace_events(current);
453 if (regs)
454 show_regs(regs);
455 else
456 dump_stack();
457
458 if (softlockup_all_cpu_backtrace) {
459 /* Avoid generating two back traces for current
460 * given that one is already made above
461 */
462 trigger_allbutself_cpu_backtrace();
463
464 clear_bit(0, &soft_lockup_nmi_warn);
465 /* Barrier to sync with other cpus */
466 smp_mb__after_atomic();
467 }
468
469 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
470 if (softlockup_panic)
471 panic("softlockup: hung tasks");
472 __this_cpu_write(soft_watchdog_warn, true);
473 } else
474 __this_cpu_write(soft_watchdog_warn, false);
475
476 return HRTIMER_RESTART;
477}
478
479static void watchdog_enable(unsigned int cpu)
480{
481 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
482 struct completion *done = this_cpu_ptr(&softlockup_completion);
483
484 WARN_ON_ONCE(cpu != smp_processor_id());
485
486 init_completion(done);
487 complete(done);
488
489 /*
490 * Start the timer first to prevent the NMI watchdog triggering
491 * before the timer has a chance to fire.
492 */
493 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
494 hrtimer->function = watchdog_timer_fn;
495 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
496 HRTIMER_MODE_REL_PINNED_HARD);
497
498 /* Initialize timestamp */
499 __touch_watchdog();
500 /* Enable the perf event */
501 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
502 watchdog_nmi_enable(cpu);
503}
504
505static void watchdog_disable(unsigned int cpu)
506{
507 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
508
509 WARN_ON_ONCE(cpu != smp_processor_id());
510
511 /*
512 * Disable the perf event first. That prevents that a large delay
513 * between disabling the timer and disabling the perf event causes
514 * the perf NMI to detect a false positive.
515 */
516 watchdog_nmi_disable(cpu);
517 hrtimer_cancel(hrtimer);
518 wait_for_completion(this_cpu_ptr(&softlockup_completion));
519}
520
521static int softlockup_stop_fn(void *data)
522{
523 watchdog_disable(smp_processor_id());
524 return 0;
525}
526
527static void softlockup_stop_all(void)
528{
529 int cpu;
530
531 if (!softlockup_initialized)
532 return;
533
534 for_each_cpu(cpu, &watchdog_allowed_mask)
535 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
536
537 cpumask_clear(&watchdog_allowed_mask);
538}
539
540static int softlockup_start_fn(void *data)
541{
542 watchdog_enable(smp_processor_id());
543 return 0;
544}
545
546static void softlockup_start_all(void)
547{
548 int cpu;
549
550 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
551 for_each_cpu(cpu, &watchdog_allowed_mask)
552 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
553}
554
555int lockup_detector_online_cpu(unsigned int cpu)
556{
557 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
558 watchdog_enable(cpu);
559 return 0;
560}
561
562int lockup_detector_offline_cpu(unsigned int cpu)
563{
564 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
565 watchdog_disable(cpu);
566 return 0;
567}
568
569static void lockup_detector_reconfigure(void)
570{
571 cpus_read_lock();
572 watchdog_nmi_stop();
573
574 softlockup_stop_all();
575 set_sample_period();
576 lockup_detector_update_enable();
577 if (watchdog_enabled && watchdog_thresh)
578 softlockup_start_all();
579
580 watchdog_nmi_start();
581 cpus_read_unlock();
582 /*
583 * Must be called outside the cpus locked section to prevent
584 * recursive locking in the perf code.
585 */
586 __lockup_detector_cleanup();
587}
588
589/*
590 * Create the watchdog thread infrastructure and configure the detector(s).
591 *
592 * The threads are not unparked as watchdog_allowed_mask is empty. When
593 * the threads are successfully initialized, take the proper locks and
594 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
595 */
596static __init void lockup_detector_setup(void)
597{
598 /*
599 * If sysctl is off and watchdog got disabled on the command line,
600 * nothing to do here.
601 */
602 lockup_detector_update_enable();
603
604 if (!IS_ENABLED(CONFIG_SYSCTL) &&
605 !(watchdog_enabled && watchdog_thresh))
606 return;
607
608 mutex_lock(&watchdog_mutex);
609 lockup_detector_reconfigure();
610 softlockup_initialized = true;
611 mutex_unlock(&watchdog_mutex);
612}
613
614#else /* CONFIG_SOFTLOCKUP_DETECTOR */
615static void lockup_detector_reconfigure(void)
616{
617 cpus_read_lock();
618 watchdog_nmi_stop();
619 lockup_detector_update_enable();
620 watchdog_nmi_start();
621 cpus_read_unlock();
622}
623static inline void lockup_detector_setup(void)
624{
625 lockup_detector_reconfigure();
626}
627#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
628
629static void __lockup_detector_cleanup(void)
630{
631 lockdep_assert_held(&watchdog_mutex);
632 hardlockup_detector_perf_cleanup();
633}
634
635/**
636 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
637 *
638 * Caller must not hold the cpu hotplug rwsem.
639 */
640void lockup_detector_cleanup(void)
641{
642 mutex_lock(&watchdog_mutex);
643 __lockup_detector_cleanup();
644 mutex_unlock(&watchdog_mutex);
645}
646
647/**
648 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
649 *
650 * Special interface for parisc. It prevents lockup detector warnings from
651 * the default pm_poweroff() function which busy loops forever.
652 */
653void lockup_detector_soft_poweroff(void)
654{
655 watchdog_enabled = 0;
656}
657
658#ifdef CONFIG_SYSCTL
659
660/* Propagate any changes to the watchdog threads */
661static void proc_watchdog_update(void)
662{
663 /* Remove impossible cpus to keep sysctl output clean. */
664 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
665 lockup_detector_reconfigure();
666}
667
668/*
669 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
670 *
671 * caller | table->data points to | 'which'
672 * -------------------|----------------------------|--------------------------
673 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
674 * | | SOFT_WATCHDOG_ENABLED
675 * -------------------|----------------------------|--------------------------
676 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
677 * -------------------|----------------------------|--------------------------
678 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
679 */
680static int proc_watchdog_common(int which, struct ctl_table *table, int write,
681 void __user *buffer, size_t *lenp, loff_t *ppos)
682{
683 int err, old, *param = table->data;
684
685 mutex_lock(&watchdog_mutex);
686
687 if (!write) {
688 /*
689 * On read synchronize the userspace interface. This is a
690 * racy snapshot.
691 */
692 *param = (watchdog_enabled & which) != 0;
693 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
694 } else {
695 old = READ_ONCE(*param);
696 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
697 if (!err && old != READ_ONCE(*param))
698 proc_watchdog_update();
699 }
700 mutex_unlock(&watchdog_mutex);
701 return err;
702}
703
704/*
705 * /proc/sys/kernel/watchdog
706 */
707int proc_watchdog(struct ctl_table *table, int write,
708 void __user *buffer, size_t *lenp, loff_t *ppos)
709{
710 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
711 table, write, buffer, lenp, ppos);
712}
713
714/*
715 * /proc/sys/kernel/nmi_watchdog
716 */
717int proc_nmi_watchdog(struct ctl_table *table, int write,
718 void __user *buffer, size_t *lenp, loff_t *ppos)
719{
720 if (!nmi_watchdog_available && write)
721 return -ENOTSUPP;
722 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
723 table, write, buffer, lenp, ppos);
724}
725
726/*
727 * /proc/sys/kernel/soft_watchdog
728 */
729int proc_soft_watchdog(struct ctl_table *table, int write,
730 void __user *buffer, size_t *lenp, loff_t *ppos)
731{
732 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
733 table, write, buffer, lenp, ppos);
734}
735
736/*
737 * /proc/sys/kernel/watchdog_thresh
738 */
739int proc_watchdog_thresh(struct ctl_table *table, int write,
740 void __user *buffer, size_t *lenp, loff_t *ppos)
741{
742 int err, old;
743
744 mutex_lock(&watchdog_mutex);
745
746 old = READ_ONCE(watchdog_thresh);
747 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
748
749 if (!err && write && old != READ_ONCE(watchdog_thresh))
750 proc_watchdog_update();
751
752 mutex_unlock(&watchdog_mutex);
753 return err;
754}
755
756/*
757 * The cpumask is the mask of possible cpus that the watchdog can run
758 * on, not the mask of cpus it is actually running on. This allows the
759 * user to specify a mask that will include cpus that have not yet
760 * been brought online, if desired.
761 */
762int proc_watchdog_cpumask(struct ctl_table *table, int write,
763 void __user *buffer, size_t *lenp, loff_t *ppos)
764{
765 int err;
766
767 mutex_lock(&watchdog_mutex);
768
769 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
770 if (!err && write)
771 proc_watchdog_update();
772
773 mutex_unlock(&watchdog_mutex);
774 return err;
775}
776#endif /* CONFIG_SYSCTL */
777
778void __init lockup_detector_init(void)
779{
780 if (tick_nohz_full_enabled())
781 pr_info("Disabling watchdog on nohz_full cores by default\n");
782
783 cpumask_copy(&watchdog_cpumask,
784 housekeeping_cpumask(HK_FLAG_TIMER));
785
786 if (!watchdog_nmi_probe())
787 nmi_watchdog_available = true;
788 lockup_detector_setup();
789}