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/*
2 * Detect hard and soft lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
6 * Note: Most of this code is borrowed heavily from the original softlockup
7 * detector, so thanks to Ingo for the initial implementation.
8 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
9 * to those contributors as well.
10 */
11
12#define pr_fmt(fmt) "NMI watchdog: " fmt
13
14#include <linux/mm.h>
15#include <linux/cpu.h>
16#include <linux/nmi.h>
17#include <linux/init.h>
18#include <linux/module.h>
19#include <linux/sysctl.h>
20#include <linux/smpboot.h>
21#include <linux/sched/rt.h>
22#include <linux/tick.h>
23#include <linux/workqueue.h>
24
25#include <asm/irq_regs.h>
26#include <linux/kvm_para.h>
27#include <linux/kthread.h>
28
29static DEFINE_MUTEX(watchdog_proc_mutex);
30
31#if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
32unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
33#else
34unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
35#endif
36int __read_mostly nmi_watchdog_enabled;
37int __read_mostly soft_watchdog_enabled;
38int __read_mostly watchdog_user_enabled;
39int __read_mostly watchdog_thresh = 10;
40
41#ifdef CONFIG_SMP
42int __read_mostly sysctl_softlockup_all_cpu_backtrace;
43int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
44#endif
45static struct cpumask watchdog_cpumask __read_mostly;
46unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
47
48/* Helper for online, unparked cpus. */
49#define for_each_watchdog_cpu(cpu) \
50 for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask)
51
52atomic_t watchdog_park_in_progress = ATOMIC_INIT(0);
53
54/*
55 * The 'watchdog_running' variable is set to 1 when the watchdog threads
56 * are registered/started and is set to 0 when the watchdog threads are
57 * unregistered/stopped, so it is an indicator whether the threads exist.
58 */
59static int __read_mostly watchdog_running;
60/*
61 * If a subsystem has a need to deactivate the watchdog temporarily, it
62 * can use the suspend/resume interface to achieve this. The content of
63 * the 'watchdog_suspended' variable reflects this state. Existing threads
64 * are parked/unparked by the lockup_detector_{suspend|resume} functions
65 * (see comment blocks pertaining to those functions for further details).
66 *
67 * 'watchdog_suspended' also prevents threads from being registered/started
68 * or unregistered/stopped via parameters in /proc/sys/kernel, so the state
69 * of 'watchdog_running' cannot change while the watchdog is deactivated
70 * temporarily (see related code in 'proc' handlers).
71 */
72static int __read_mostly watchdog_suspended;
73
74static u64 __read_mostly sample_period;
75
76static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
77static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
78static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
79static DEFINE_PER_CPU(bool, softlockup_touch_sync);
80static DEFINE_PER_CPU(bool, soft_watchdog_warn);
81static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
82static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
83static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
84static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
85static unsigned long soft_lockup_nmi_warn;
86
87unsigned int __read_mostly softlockup_panic =
88 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
89
90static int __init softlockup_panic_setup(char *str)
91{
92 softlockup_panic = simple_strtoul(str, NULL, 0);
93
94 return 1;
95}
96__setup("softlockup_panic=", softlockup_panic_setup);
97
98static int __init nowatchdog_setup(char *str)
99{
100 watchdog_enabled = 0;
101 return 1;
102}
103__setup("nowatchdog", nowatchdog_setup);
104
105static int __init nosoftlockup_setup(char *str)
106{
107 watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
108 return 1;
109}
110__setup("nosoftlockup", nosoftlockup_setup);
111
112#ifdef CONFIG_SMP
113static int __init softlockup_all_cpu_backtrace_setup(char *str)
114{
115 sysctl_softlockup_all_cpu_backtrace =
116 !!simple_strtol(str, NULL, 0);
117 return 1;
118}
119__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
120static int __init hardlockup_all_cpu_backtrace_setup(char *str)
121{
122 sysctl_hardlockup_all_cpu_backtrace =
123 !!simple_strtol(str, NULL, 0);
124 return 1;
125}
126__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
127#endif
128
129/*
130 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
131 * lockups can have false positives under extreme conditions. So we generally
132 * want a higher threshold for soft lockups than for hard lockups. So we couple
133 * the thresholds with a factor: we make the soft threshold twice the amount of
134 * time the hard threshold is.
135 */
136static int get_softlockup_thresh(void)
137{
138 return watchdog_thresh * 2;
139}
140
141/*
142 * Returns seconds, approximately. We don't need nanosecond
143 * resolution, and we don't need to waste time with a big divide when
144 * 2^30ns == 1.074s.
145 */
146static unsigned long get_timestamp(void)
147{
148 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
149}
150
151static void set_sample_period(void)
152{
153 /*
154 * convert watchdog_thresh from seconds to ns
155 * the divide by 5 is to give hrtimer several chances (two
156 * or three with the current relation between the soft
157 * and hard thresholds) to increment before the
158 * hardlockup detector generates a warning
159 */
160 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
161}
162
163/* Commands for resetting the watchdog */
164static void __touch_watchdog(void)
165{
166 __this_cpu_write(watchdog_touch_ts, get_timestamp());
167}
168
169/**
170 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
171 *
172 * Call when the scheduler may have stalled for legitimate reasons
173 * preventing the watchdog task from executing - e.g. the scheduler
174 * entering idle state. This should only be used for scheduler events.
175 * Use touch_softlockup_watchdog() for everything else.
176 */
177void touch_softlockup_watchdog_sched(void)
178{
179 /*
180 * Preemption can be enabled. It doesn't matter which CPU's timestamp
181 * gets zeroed here, so use the raw_ operation.
182 */
183 raw_cpu_write(watchdog_touch_ts, 0);
184}
185
186void touch_softlockup_watchdog(void)
187{
188 touch_softlockup_watchdog_sched();
189 wq_watchdog_touch(raw_smp_processor_id());
190}
191EXPORT_SYMBOL(touch_softlockup_watchdog);
192
193void touch_all_softlockup_watchdogs(void)
194{
195 int cpu;
196
197 /*
198 * this is done lockless
199 * do we care if a 0 races with a timestamp?
200 * all it means is the softlock check starts one cycle later
201 */
202 for_each_watchdog_cpu(cpu)
203 per_cpu(watchdog_touch_ts, cpu) = 0;
204 wq_watchdog_touch(-1);
205}
206
207void touch_softlockup_watchdog_sync(void)
208{
209 __this_cpu_write(softlockup_touch_sync, true);
210 __this_cpu_write(watchdog_touch_ts, 0);
211}
212
213/* watchdog detector functions */
214bool is_hardlockup(void)
215{
216 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
217
218 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
219 return true;
220
221 __this_cpu_write(hrtimer_interrupts_saved, hrint);
222 return false;
223}
224
225static int is_softlockup(unsigned long touch_ts)
226{
227 unsigned long now = get_timestamp();
228
229 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
230 /* Warn about unreasonable delays. */
231 if (time_after(now, touch_ts + get_softlockup_thresh()))
232 return now - touch_ts;
233 }
234 return 0;
235}
236
237static void watchdog_interrupt_count(void)
238{
239 __this_cpu_inc(hrtimer_interrupts);
240}
241
242/*
243 * These two functions are mostly architecture specific
244 * defining them as weak here.
245 */
246int __weak watchdog_nmi_enable(unsigned int cpu)
247{
248 return 0;
249}
250void __weak watchdog_nmi_disable(unsigned int cpu)
251{
252}
253
254static int watchdog_enable_all_cpus(void);
255static void watchdog_disable_all_cpus(void);
256
257/* watchdog kicker functions */
258static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
259{
260 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
261 struct pt_regs *regs = get_irq_regs();
262 int duration;
263 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
264
265 if (atomic_read(&watchdog_park_in_progress) != 0)
266 return HRTIMER_NORESTART;
267
268 /* kick the hardlockup detector */
269 watchdog_interrupt_count();
270
271 /* kick the softlockup detector */
272 wake_up_process(__this_cpu_read(softlockup_watchdog));
273
274 /* .. and repeat */
275 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
276
277 if (touch_ts == 0) {
278 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
279 /*
280 * If the time stamp was touched atomically
281 * make sure the scheduler tick is up to date.
282 */
283 __this_cpu_write(softlockup_touch_sync, false);
284 sched_clock_tick();
285 }
286
287 /* Clear the guest paused flag on watchdog reset */
288 kvm_check_and_clear_guest_paused();
289 __touch_watchdog();
290 return HRTIMER_RESTART;
291 }
292
293 /* check for a softlockup
294 * This is done by making sure a high priority task is
295 * being scheduled. The task touches the watchdog to
296 * indicate it is getting cpu time. If it hasn't then
297 * this is a good indication some task is hogging the cpu
298 */
299 duration = is_softlockup(touch_ts);
300 if (unlikely(duration)) {
301 /*
302 * If a virtual machine is stopped by the host it can look to
303 * the watchdog like a soft lockup, check to see if the host
304 * stopped the vm before we issue the warning
305 */
306 if (kvm_check_and_clear_guest_paused())
307 return HRTIMER_RESTART;
308
309 /* only warn once */
310 if (__this_cpu_read(soft_watchdog_warn) == true) {
311 /*
312 * When multiple processes are causing softlockups the
313 * softlockup detector only warns on the first one
314 * because the code relies on a full quiet cycle to
315 * re-arm. The second process prevents the quiet cycle
316 * and never gets reported. Use task pointers to detect
317 * this.
318 */
319 if (__this_cpu_read(softlockup_task_ptr_saved) !=
320 current) {
321 __this_cpu_write(soft_watchdog_warn, false);
322 __touch_watchdog();
323 }
324 return HRTIMER_RESTART;
325 }
326
327 if (softlockup_all_cpu_backtrace) {
328 /* Prevent multiple soft-lockup reports if one cpu is already
329 * engaged in dumping cpu back traces
330 */
331 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
332 /* Someone else will report us. Let's give up */
333 __this_cpu_write(soft_watchdog_warn, true);
334 return HRTIMER_RESTART;
335 }
336 }
337
338 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
339 smp_processor_id(), duration,
340 current->comm, task_pid_nr(current));
341 __this_cpu_write(softlockup_task_ptr_saved, current);
342 print_modules();
343 print_irqtrace_events(current);
344 if (regs)
345 show_regs(regs);
346 else
347 dump_stack();
348
349 if (softlockup_all_cpu_backtrace) {
350 /* Avoid generating two back traces for current
351 * given that one is already made above
352 */
353 trigger_allbutself_cpu_backtrace();
354
355 clear_bit(0, &soft_lockup_nmi_warn);
356 /* Barrier to sync with other cpus */
357 smp_mb__after_atomic();
358 }
359
360 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
361 if (softlockup_panic)
362 panic("softlockup: hung tasks");
363 __this_cpu_write(soft_watchdog_warn, true);
364 } else
365 __this_cpu_write(soft_watchdog_warn, false);
366
367 return HRTIMER_RESTART;
368}
369
370static void watchdog_set_prio(unsigned int policy, unsigned int prio)
371{
372 struct sched_param param = { .sched_priority = prio };
373
374 sched_setscheduler(current, policy, ¶m);
375}
376
377static void watchdog_enable(unsigned int cpu)
378{
379 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
380
381 /* kick off the timer for the hardlockup detector */
382 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
383 hrtimer->function = watchdog_timer_fn;
384
385 /* Enable the perf event */
386 watchdog_nmi_enable(cpu);
387
388 /* done here because hrtimer_start can only pin to smp_processor_id() */
389 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
390 HRTIMER_MODE_REL_PINNED);
391
392 /* initialize timestamp */
393 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
394 __touch_watchdog();
395}
396
397static void watchdog_disable(unsigned int cpu)
398{
399 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
400
401 watchdog_set_prio(SCHED_NORMAL, 0);
402 hrtimer_cancel(hrtimer);
403 /* disable the perf event */
404 watchdog_nmi_disable(cpu);
405}
406
407static void watchdog_cleanup(unsigned int cpu, bool online)
408{
409 watchdog_disable(cpu);
410}
411
412static int watchdog_should_run(unsigned int cpu)
413{
414 return __this_cpu_read(hrtimer_interrupts) !=
415 __this_cpu_read(soft_lockup_hrtimer_cnt);
416}
417
418/*
419 * The watchdog thread function - touches the timestamp.
420 *
421 * It only runs once every sample_period seconds (4 seconds by
422 * default) to reset the softlockup timestamp. If this gets delayed
423 * for more than 2*watchdog_thresh seconds then the debug-printout
424 * triggers in watchdog_timer_fn().
425 */
426static void watchdog(unsigned int cpu)
427{
428 __this_cpu_write(soft_lockup_hrtimer_cnt,
429 __this_cpu_read(hrtimer_interrupts));
430 __touch_watchdog();
431
432 /*
433 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
434 * failure path. Check for failures that can occur asynchronously -
435 * for example, when CPUs are on-lined - and shut down the hardware
436 * perf event on each CPU accordingly.
437 *
438 * The only non-obvious place this bit can be cleared is through
439 * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a
440 * pr_info here would be too noisy as it would result in a message
441 * every few seconds if the hardlockup was disabled but the softlockup
442 * enabled.
443 */
444 if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
445 watchdog_nmi_disable(cpu);
446}
447
448static struct smp_hotplug_thread watchdog_threads = {
449 .store = &softlockup_watchdog,
450 .thread_should_run = watchdog_should_run,
451 .thread_fn = watchdog,
452 .thread_comm = "watchdog/%u",
453 .setup = watchdog_enable,
454 .cleanup = watchdog_cleanup,
455 .park = watchdog_disable,
456 .unpark = watchdog_enable,
457};
458
459/*
460 * park all watchdog threads that are specified in 'watchdog_cpumask'
461 *
462 * This function returns an error if kthread_park() of a watchdog thread
463 * fails. In this situation, the watchdog threads of some CPUs can already
464 * be parked and the watchdog threads of other CPUs can still be runnable.
465 * Callers are expected to handle this special condition as appropriate in
466 * their context.
467 *
468 * This function may only be called in a context that is protected against
469 * races with CPU hotplug - for example, via get_online_cpus().
470 */
471static int watchdog_park_threads(void)
472{
473 int cpu, ret = 0;
474
475 atomic_set(&watchdog_park_in_progress, 1);
476
477 for_each_watchdog_cpu(cpu) {
478 ret = kthread_park(per_cpu(softlockup_watchdog, cpu));
479 if (ret)
480 break;
481 }
482
483 atomic_set(&watchdog_park_in_progress, 0);
484
485 return ret;
486}
487
488/*
489 * unpark all watchdog threads that are specified in 'watchdog_cpumask'
490 *
491 * This function may only be called in a context that is protected against
492 * races with CPU hotplug - for example, via get_online_cpus().
493 */
494static void watchdog_unpark_threads(void)
495{
496 int cpu;
497
498 for_each_watchdog_cpu(cpu)
499 kthread_unpark(per_cpu(softlockup_watchdog, cpu));
500}
501
502/*
503 * Suspend the hard and soft lockup detector by parking the watchdog threads.
504 */
505int lockup_detector_suspend(void)
506{
507 int ret = 0;
508
509 get_online_cpus();
510 mutex_lock(&watchdog_proc_mutex);
511 /*
512 * Multiple suspend requests can be active in parallel (counted by
513 * the 'watchdog_suspended' variable). If the watchdog threads are
514 * running, the first caller takes care that they will be parked.
515 * The state of 'watchdog_running' cannot change while a suspend
516 * request is active (see related code in 'proc' handlers).
517 */
518 if (watchdog_running && !watchdog_suspended)
519 ret = watchdog_park_threads();
520
521 if (ret == 0)
522 watchdog_suspended++;
523 else {
524 watchdog_disable_all_cpus();
525 pr_err("Failed to suspend lockup detectors, disabled\n");
526 watchdog_enabled = 0;
527 }
528
529 mutex_unlock(&watchdog_proc_mutex);
530
531 return ret;
532}
533
534/*
535 * Resume the hard and soft lockup detector by unparking the watchdog threads.
536 */
537void lockup_detector_resume(void)
538{
539 mutex_lock(&watchdog_proc_mutex);
540
541 watchdog_suspended--;
542 /*
543 * The watchdog threads are unparked if they were previously running
544 * and if there is no more active suspend request.
545 */
546 if (watchdog_running && !watchdog_suspended)
547 watchdog_unpark_threads();
548
549 mutex_unlock(&watchdog_proc_mutex);
550 put_online_cpus();
551}
552
553static int update_watchdog_all_cpus(void)
554{
555 int ret;
556
557 ret = watchdog_park_threads();
558 if (ret)
559 return ret;
560
561 watchdog_unpark_threads();
562
563 return 0;
564}
565
566static int watchdog_enable_all_cpus(void)
567{
568 int err = 0;
569
570 if (!watchdog_running) {
571 err = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
572 &watchdog_cpumask);
573 if (err)
574 pr_err("Failed to create watchdog threads, disabled\n");
575 else
576 watchdog_running = 1;
577 } else {
578 /*
579 * Enable/disable the lockup detectors or
580 * change the sample period 'on the fly'.
581 */
582 err = update_watchdog_all_cpus();
583
584 if (err) {
585 watchdog_disable_all_cpus();
586 pr_err("Failed to update lockup detectors, disabled\n");
587 }
588 }
589
590 if (err)
591 watchdog_enabled = 0;
592
593 return err;
594}
595
596static void watchdog_disable_all_cpus(void)
597{
598 if (watchdog_running) {
599 watchdog_running = 0;
600 smpboot_unregister_percpu_thread(&watchdog_threads);
601 }
602}
603
604#ifdef CONFIG_SYSCTL
605
606/*
607 * Update the run state of the lockup detectors.
608 */
609static int proc_watchdog_update(void)
610{
611 int err = 0;
612
613 /*
614 * Watchdog threads won't be started if they are already active.
615 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
616 * care of this. If those threads are already active, the sample
617 * period will be updated and the lockup detectors will be enabled
618 * or disabled 'on the fly'.
619 */
620 if (watchdog_enabled && watchdog_thresh)
621 err = watchdog_enable_all_cpus();
622 else
623 watchdog_disable_all_cpus();
624
625 return err;
626
627}
628
629/*
630 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
631 *
632 * caller | table->data points to | 'which' contains the flag(s)
633 * -------------------|-----------------------|-----------------------------
634 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
635 * | | with SOFT_WATCHDOG_ENABLED
636 * -------------------|-----------------------|-----------------------------
637 * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED
638 * -------------------|-----------------------|-----------------------------
639 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
640 */
641static int proc_watchdog_common(int which, struct ctl_table *table, int write,
642 void __user *buffer, size_t *lenp, loff_t *ppos)
643{
644 int err, old, new;
645 int *watchdog_param = (int *)table->data;
646
647 get_online_cpus();
648 mutex_lock(&watchdog_proc_mutex);
649
650 if (watchdog_suspended) {
651 /* no parameter changes allowed while watchdog is suspended */
652 err = -EAGAIN;
653 goto out;
654 }
655
656 /*
657 * If the parameter is being read return the state of the corresponding
658 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
659 * run state of the lockup detectors.
660 */
661 if (!write) {
662 *watchdog_param = (watchdog_enabled & which) != 0;
663 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
664 } else {
665 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
666 if (err)
667 goto out;
668
669 /*
670 * There is a race window between fetching the current value
671 * from 'watchdog_enabled' and storing the new value. During
672 * this race window, watchdog_nmi_enable() can sneak in and
673 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
674 * The 'cmpxchg' detects this race and the loop retries.
675 */
676 do {
677 old = watchdog_enabled;
678 /*
679 * If the parameter value is not zero set the
680 * corresponding bit(s), else clear it(them).
681 */
682 if (*watchdog_param)
683 new = old | which;
684 else
685 new = old & ~which;
686 } while (cmpxchg(&watchdog_enabled, old, new) != old);
687
688 /*
689 * Update the run state of the lockup detectors. There is _no_
690 * need to check the value returned by proc_watchdog_update()
691 * and to restore the previous value of 'watchdog_enabled' as
692 * both lockup detectors are disabled if proc_watchdog_update()
693 * returns an error.
694 */
695 if (old == new)
696 goto out;
697
698 err = proc_watchdog_update();
699 }
700out:
701 mutex_unlock(&watchdog_proc_mutex);
702 put_online_cpus();
703 return err;
704}
705
706/*
707 * /proc/sys/kernel/watchdog
708 */
709int proc_watchdog(struct ctl_table *table, int write,
710 void __user *buffer, size_t *lenp, loff_t *ppos)
711{
712 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
713 table, write, buffer, lenp, ppos);
714}
715
716/*
717 * /proc/sys/kernel/nmi_watchdog
718 */
719int proc_nmi_watchdog(struct ctl_table *table, int write,
720 void __user *buffer, size_t *lenp, loff_t *ppos)
721{
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, new;
743
744 get_online_cpus();
745 mutex_lock(&watchdog_proc_mutex);
746
747 if (watchdog_suspended) {
748 /* no parameter changes allowed while watchdog is suspended */
749 err = -EAGAIN;
750 goto out;
751 }
752
753 old = ACCESS_ONCE(watchdog_thresh);
754 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
755
756 if (err || !write)
757 goto out;
758
759 /*
760 * Update the sample period. Restore on failure.
761 */
762 new = ACCESS_ONCE(watchdog_thresh);
763 if (old == new)
764 goto out;
765
766 set_sample_period();
767 err = proc_watchdog_update();
768 if (err) {
769 watchdog_thresh = old;
770 set_sample_period();
771 }
772out:
773 mutex_unlock(&watchdog_proc_mutex);
774 put_online_cpus();
775 return err;
776}
777
778/*
779 * The cpumask is the mask of possible cpus that the watchdog can run
780 * on, not the mask of cpus it is actually running on. This allows the
781 * user to specify a mask that will include cpus that have not yet
782 * been brought online, if desired.
783 */
784int proc_watchdog_cpumask(struct ctl_table *table, int write,
785 void __user *buffer, size_t *lenp, loff_t *ppos)
786{
787 int err;
788
789 get_online_cpus();
790 mutex_lock(&watchdog_proc_mutex);
791
792 if (watchdog_suspended) {
793 /* no parameter changes allowed while watchdog is suspended */
794 err = -EAGAIN;
795 goto out;
796 }
797
798 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
799 if (!err && write) {
800 /* Remove impossible cpus to keep sysctl output cleaner. */
801 cpumask_and(&watchdog_cpumask, &watchdog_cpumask,
802 cpu_possible_mask);
803
804 if (watchdog_running) {
805 /*
806 * Failure would be due to being unable to allocate
807 * a temporary cpumask, so we are likely not in a
808 * position to do much else to make things better.
809 */
810 if (smpboot_update_cpumask_percpu_thread(
811 &watchdog_threads, &watchdog_cpumask) != 0)
812 pr_err("cpumask update failed\n");
813 }
814 }
815out:
816 mutex_unlock(&watchdog_proc_mutex);
817 put_online_cpus();
818 return err;
819}
820
821#endif /* CONFIG_SYSCTL */
822
823void __init lockup_detector_init(void)
824{
825 set_sample_period();
826
827#ifdef CONFIG_NO_HZ_FULL
828 if (tick_nohz_full_enabled()) {
829 pr_info("Disabling watchdog on nohz_full cores by default\n");
830 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
831 } else
832 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
833#else
834 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
835#endif
836
837 if (watchdog_enabled)
838 watchdog_enable_all_cpus();
839}