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/smpboot.h>
22#include <linux/sched/rt.h>
23#include <uapi/linux/sched/types.h>
24#include <linux/tick.h>
25#include <linux/workqueue.h>
26#include <linux/sched/clock.h>
27#include <linux/sched/debug.h>
28#include <linux/sched/isolation.h>
29
30#include <asm/irq_regs.h>
31#include <linux/kvm_para.h>
32#include <linux/kthread.h>
33
34static DEFINE_MUTEX(watchdog_mutex);
35
36#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
37# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
38# define NMI_WATCHDOG_DEFAULT 1
39#else
40# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
41# define NMI_WATCHDOG_DEFAULT 0
42#endif
43
44unsigned long __read_mostly watchdog_enabled;
45int __read_mostly watchdog_user_enabled = 1;
46int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
47int __read_mostly soft_watchdog_user_enabled = 1;
48int __read_mostly watchdog_thresh = 10;
49int __read_mostly nmi_watchdog_available;
50
51struct cpumask watchdog_allowed_mask __read_mostly;
52
53struct cpumask watchdog_cpumask __read_mostly;
54unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
55
56#ifdef CONFIG_HARDLOCKUP_DETECTOR
57/*
58 * Should we panic when a soft-lockup or hard-lockup occurs:
59 */
60unsigned int __read_mostly hardlockup_panic =
61 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
62/*
63 * We may not want to enable hard lockup detection by default in all cases,
64 * for example when running the kernel as a guest on a hypervisor. In these
65 * cases this function can be called to disable hard lockup detection. This
66 * function should only be executed once by the boot processor before the
67 * kernel command line parameters are parsed, because otherwise it is not
68 * possible to override this in hardlockup_panic_setup().
69 */
70void __init hardlockup_detector_disable(void)
71{
72 nmi_watchdog_user_enabled = 0;
73}
74
75static int __init hardlockup_panic_setup(char *str)
76{
77 if (!strncmp(str, "panic", 5))
78 hardlockup_panic = 1;
79 else if (!strncmp(str, "nopanic", 7))
80 hardlockup_panic = 0;
81 else if (!strncmp(str, "0", 1))
82 nmi_watchdog_user_enabled = 0;
83 else if (!strncmp(str, "1", 1))
84 nmi_watchdog_user_enabled = 1;
85 return 1;
86}
87__setup("nmi_watchdog=", hardlockup_panic_setup);
88
89# ifdef CONFIG_SMP
90int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
91
92static int __init hardlockup_all_cpu_backtrace_setup(char *str)
93{
94 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
95 return 1;
96}
97__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
98# endif /* CONFIG_SMP */
99#endif /* CONFIG_HARDLOCKUP_DETECTOR */
100
101/*
102 * These functions can be overridden if an architecture implements its
103 * own hardlockup detector.
104 *
105 * watchdog_nmi_enable/disable can be implemented to start and stop when
106 * softlockup watchdog threads start and stop. The arch must select the
107 * SOFTLOCKUP_DETECTOR Kconfig.
108 */
109int __weak watchdog_nmi_enable(unsigned int cpu)
110{
111 hardlockup_detector_perf_enable();
112 return 0;
113}
114
115void __weak watchdog_nmi_disable(unsigned int cpu)
116{
117 hardlockup_detector_perf_disable();
118}
119
120/* Return 0, if a NMI watchdog is available. Error code otherwise */
121int __weak __init watchdog_nmi_probe(void)
122{
123 return hardlockup_detector_perf_init();
124}
125
126/**
127 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
128 *
129 * The reconfiguration steps are:
130 * watchdog_nmi_stop();
131 * update_variables();
132 * watchdog_nmi_start();
133 */
134void __weak watchdog_nmi_stop(void) { }
135
136/**
137 * watchdog_nmi_start - Start the watchdog after reconfiguration
138 *
139 * Counterpart to watchdog_nmi_stop().
140 *
141 * The following variables have been updated in update_variables() and
142 * contain the currently valid configuration:
143 * - watchdog_enabled
144 * - watchdog_thresh
145 * - watchdog_cpumask
146 */
147void __weak watchdog_nmi_start(void) { }
148
149/**
150 * lockup_detector_update_enable - Update the sysctl enable bit
151 *
152 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
153 * can't race with watchdog_nmi_disable().
154 */
155static void lockup_detector_update_enable(void)
156{
157 watchdog_enabled = 0;
158 if (!watchdog_user_enabled)
159 return;
160 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
161 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
162 if (soft_watchdog_user_enabled)
163 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
164}
165
166#ifdef CONFIG_SOFTLOCKUP_DETECTOR
167
168/* Global variables, exported for sysctl */
169unsigned int __read_mostly softlockup_panic =
170 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
171
172static bool softlockup_threads_initialized __read_mostly;
173static u64 __read_mostly sample_period;
174
175static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
176static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
177static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
178static DEFINE_PER_CPU(bool, softlockup_touch_sync);
179static DEFINE_PER_CPU(bool, soft_watchdog_warn);
180static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
181static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
182static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
183static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
184static unsigned long soft_lockup_nmi_warn;
185
186static int __init softlockup_panic_setup(char *str)
187{
188 softlockup_panic = simple_strtoul(str, NULL, 0);
189 return 1;
190}
191__setup("softlockup_panic=", softlockup_panic_setup);
192
193static int __init nowatchdog_setup(char *str)
194{
195 watchdog_user_enabled = 0;
196 return 1;
197}
198__setup("nowatchdog", nowatchdog_setup);
199
200static int __init nosoftlockup_setup(char *str)
201{
202 soft_watchdog_user_enabled = 0;
203 return 1;
204}
205__setup("nosoftlockup", nosoftlockup_setup);
206
207#ifdef CONFIG_SMP
208int __read_mostly sysctl_softlockup_all_cpu_backtrace;
209
210static int __init softlockup_all_cpu_backtrace_setup(char *str)
211{
212 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
213 return 1;
214}
215__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
216#endif
217
218static void __lockup_detector_cleanup(void);
219
220/*
221 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
222 * lockups can have false positives under extreme conditions. So we generally
223 * want a higher threshold for soft lockups than for hard lockups. So we couple
224 * the thresholds with a factor: we make the soft threshold twice the amount of
225 * time the hard threshold is.
226 */
227static int get_softlockup_thresh(void)
228{
229 return watchdog_thresh * 2;
230}
231
232/*
233 * Returns seconds, approximately. We don't need nanosecond
234 * resolution, and we don't need to waste time with a big divide when
235 * 2^30ns == 1.074s.
236 */
237static unsigned long get_timestamp(void)
238{
239 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
240}
241
242static void set_sample_period(void)
243{
244 /*
245 * convert watchdog_thresh from seconds to ns
246 * the divide by 5 is to give hrtimer several chances (two
247 * or three with the current relation between the soft
248 * and hard thresholds) to increment before the
249 * hardlockup detector generates a warning
250 */
251 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
252 watchdog_update_hrtimer_threshold(sample_period);
253}
254
255/* Commands for resetting the watchdog */
256static void __touch_watchdog(void)
257{
258 __this_cpu_write(watchdog_touch_ts, get_timestamp());
259}
260
261/**
262 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
263 *
264 * Call when the scheduler may have stalled for legitimate reasons
265 * preventing the watchdog task from executing - e.g. the scheduler
266 * entering idle state. This should only be used for scheduler events.
267 * Use touch_softlockup_watchdog() for everything else.
268 */
269void touch_softlockup_watchdog_sched(void)
270{
271 /*
272 * Preemption can be enabled. It doesn't matter which CPU's timestamp
273 * gets zeroed here, so use the raw_ operation.
274 */
275 raw_cpu_write(watchdog_touch_ts, 0);
276}
277
278void touch_softlockup_watchdog(void)
279{
280 touch_softlockup_watchdog_sched();
281 wq_watchdog_touch(raw_smp_processor_id());
282}
283EXPORT_SYMBOL(touch_softlockup_watchdog);
284
285void touch_all_softlockup_watchdogs(void)
286{
287 int cpu;
288
289 /*
290 * watchdog_mutex cannpt be taken here, as this might be called
291 * from (soft)interrupt context, so the access to
292 * watchdog_allowed_cpumask might race with a concurrent update.
293 *
294 * The watchdog time stamp can race against a concurrent real
295 * update as well, the only side effect might be a cycle delay for
296 * the softlockup check.
297 */
298 for_each_cpu(cpu, &watchdog_allowed_mask)
299 per_cpu(watchdog_touch_ts, cpu) = 0;
300 wq_watchdog_touch(-1);
301}
302
303void touch_softlockup_watchdog_sync(void)
304{
305 __this_cpu_write(softlockup_touch_sync, true);
306 __this_cpu_write(watchdog_touch_ts, 0);
307}
308
309static int is_softlockup(unsigned long touch_ts)
310{
311 unsigned long now = get_timestamp();
312
313 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
314 /* Warn about unreasonable delays. */
315 if (time_after(now, touch_ts + get_softlockup_thresh()))
316 return now - touch_ts;
317 }
318 return 0;
319}
320
321/* watchdog detector functions */
322bool is_hardlockup(void)
323{
324 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
325
326 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
327 return true;
328
329 __this_cpu_write(hrtimer_interrupts_saved, hrint);
330 return false;
331}
332
333static void watchdog_interrupt_count(void)
334{
335 __this_cpu_inc(hrtimer_interrupts);
336}
337
338/* watchdog kicker functions */
339static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
340{
341 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
342 struct pt_regs *regs = get_irq_regs();
343 int duration;
344 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
345
346 if (!watchdog_enabled)
347 return HRTIMER_NORESTART;
348
349 /* kick the hardlockup detector */
350 watchdog_interrupt_count();
351
352 /* kick the softlockup detector */
353 wake_up_process(__this_cpu_read(softlockup_watchdog));
354
355 /* .. and repeat */
356 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
357
358 if (touch_ts == 0) {
359 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
360 /*
361 * If the time stamp was touched atomically
362 * make sure the scheduler tick is up to date.
363 */
364 __this_cpu_write(softlockup_touch_sync, false);
365 sched_clock_tick();
366 }
367
368 /* Clear the guest paused flag on watchdog reset */
369 kvm_check_and_clear_guest_paused();
370 __touch_watchdog();
371 return HRTIMER_RESTART;
372 }
373
374 /* check for a softlockup
375 * This is done by making sure a high priority task is
376 * being scheduled. The task touches the watchdog to
377 * indicate it is getting cpu time. If it hasn't then
378 * this is a good indication some task is hogging the cpu
379 */
380 duration = is_softlockup(touch_ts);
381 if (unlikely(duration)) {
382 /*
383 * If a virtual machine is stopped by the host it can look to
384 * the watchdog like a soft lockup, check to see if the host
385 * stopped the vm before we issue the warning
386 */
387 if (kvm_check_and_clear_guest_paused())
388 return HRTIMER_RESTART;
389
390 /* only warn once */
391 if (__this_cpu_read(soft_watchdog_warn) == true) {
392 /*
393 * When multiple processes are causing softlockups the
394 * softlockup detector only warns on the first one
395 * because the code relies on a full quiet cycle to
396 * re-arm. The second process prevents the quiet cycle
397 * and never gets reported. Use task pointers to detect
398 * this.
399 */
400 if (__this_cpu_read(softlockup_task_ptr_saved) !=
401 current) {
402 __this_cpu_write(soft_watchdog_warn, false);
403 __touch_watchdog();
404 }
405 return HRTIMER_RESTART;
406 }
407
408 if (softlockup_all_cpu_backtrace) {
409 /* Prevent multiple soft-lockup reports if one cpu is already
410 * engaged in dumping cpu back traces
411 */
412 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
413 /* Someone else will report us. Let's give up */
414 __this_cpu_write(soft_watchdog_warn, true);
415 return HRTIMER_RESTART;
416 }
417 }
418
419 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
420 smp_processor_id(), duration,
421 current->comm, task_pid_nr(current));
422 __this_cpu_write(softlockup_task_ptr_saved, current);
423 print_modules();
424 print_irqtrace_events(current);
425 if (regs)
426 show_regs(regs);
427 else
428 dump_stack();
429
430 if (softlockup_all_cpu_backtrace) {
431 /* Avoid generating two back traces for current
432 * given that one is already made above
433 */
434 trigger_allbutself_cpu_backtrace();
435
436 clear_bit(0, &soft_lockup_nmi_warn);
437 /* Barrier to sync with other cpus */
438 smp_mb__after_atomic();
439 }
440
441 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
442 if (softlockup_panic)
443 panic("softlockup: hung tasks");
444 __this_cpu_write(soft_watchdog_warn, true);
445 } else
446 __this_cpu_write(soft_watchdog_warn, false);
447
448 return HRTIMER_RESTART;
449}
450
451static void watchdog_set_prio(unsigned int policy, unsigned int prio)
452{
453 struct sched_param param = { .sched_priority = prio };
454
455 sched_setscheduler(current, policy, ¶m);
456}
457
458static void watchdog_enable(unsigned int cpu)
459{
460 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
461
462 /*
463 * Start the timer first to prevent the NMI watchdog triggering
464 * before the timer has a chance to fire.
465 */
466 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
467 hrtimer->function = watchdog_timer_fn;
468 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
469 HRTIMER_MODE_REL_PINNED);
470
471 /* Initialize timestamp */
472 __touch_watchdog();
473 /* Enable the perf event */
474 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
475 watchdog_nmi_enable(cpu);
476
477 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
478}
479
480static void watchdog_disable(unsigned int cpu)
481{
482 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
483
484 watchdog_set_prio(SCHED_NORMAL, 0);
485 /*
486 * Disable the perf event first. That prevents that a large delay
487 * between disabling the timer and disabling the perf event causes
488 * the perf NMI to detect a false positive.
489 */
490 watchdog_nmi_disable(cpu);
491 hrtimer_cancel(hrtimer);
492}
493
494static void watchdog_cleanup(unsigned int cpu, bool online)
495{
496 watchdog_disable(cpu);
497}
498
499static int watchdog_should_run(unsigned int cpu)
500{
501 return __this_cpu_read(hrtimer_interrupts) !=
502 __this_cpu_read(soft_lockup_hrtimer_cnt);
503}
504
505/*
506 * The watchdog thread function - touches the timestamp.
507 *
508 * It only runs once every sample_period seconds (4 seconds by
509 * default) to reset the softlockup timestamp. If this gets delayed
510 * for more than 2*watchdog_thresh seconds then the debug-printout
511 * triggers in watchdog_timer_fn().
512 */
513static void watchdog(unsigned int cpu)
514{
515 __this_cpu_write(soft_lockup_hrtimer_cnt,
516 __this_cpu_read(hrtimer_interrupts));
517 __touch_watchdog();
518}
519
520static struct smp_hotplug_thread watchdog_threads = {
521 .store = &softlockup_watchdog,
522 .thread_should_run = watchdog_should_run,
523 .thread_fn = watchdog,
524 .thread_comm = "watchdog/%u",
525 .setup = watchdog_enable,
526 .cleanup = watchdog_cleanup,
527 .park = watchdog_disable,
528 .unpark = watchdog_enable,
529};
530
531static void softlockup_update_smpboot_threads(void)
532{
533 lockdep_assert_held(&watchdog_mutex);
534
535 if (!softlockup_threads_initialized)
536 return;
537
538 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
539 &watchdog_allowed_mask);
540}
541
542/* Temporarily park all watchdog threads */
543static void softlockup_park_all_threads(void)
544{
545 cpumask_clear(&watchdog_allowed_mask);
546 softlockup_update_smpboot_threads();
547}
548
549/* Unpark enabled threads */
550static void softlockup_unpark_threads(void)
551{
552 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
553 softlockup_update_smpboot_threads();
554}
555
556static void lockup_detector_reconfigure(void)
557{
558 cpus_read_lock();
559 watchdog_nmi_stop();
560 softlockup_park_all_threads();
561 set_sample_period();
562 lockup_detector_update_enable();
563 if (watchdog_enabled && watchdog_thresh)
564 softlockup_unpark_threads();
565 watchdog_nmi_start();
566 cpus_read_unlock();
567 /*
568 * Must be called outside the cpus locked section to prevent
569 * recursive locking in the perf code.
570 */
571 __lockup_detector_cleanup();
572}
573
574/*
575 * Create the watchdog thread infrastructure and configure the detector(s).
576 *
577 * The threads are not unparked as watchdog_allowed_mask is empty. When
578 * the threads are sucessfully initialized, take the proper locks and
579 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
580 */
581static __init void lockup_detector_setup(void)
582{
583 int ret;
584
585 /*
586 * If sysctl is off and watchdog got disabled on the command line,
587 * nothing to do here.
588 */
589 lockup_detector_update_enable();
590
591 if (!IS_ENABLED(CONFIG_SYSCTL) &&
592 !(watchdog_enabled && watchdog_thresh))
593 return;
594
595 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
596 &watchdog_allowed_mask);
597 if (ret) {
598 pr_err("Failed to initialize soft lockup detector threads\n");
599 return;
600 }
601
602 mutex_lock(&watchdog_mutex);
603 softlockup_threads_initialized = true;
604 lockup_detector_reconfigure();
605 mutex_unlock(&watchdog_mutex);
606}
607
608#else /* CONFIG_SOFTLOCKUP_DETECTOR */
609static inline int watchdog_park_threads(void) { return 0; }
610static inline void watchdog_unpark_threads(void) { }
611static inline int watchdog_enable_all_cpus(void) { return 0; }
612static inline void watchdog_disable_all_cpus(void) { }
613static void lockup_detector_reconfigure(void)
614{
615 cpus_read_lock();
616 watchdog_nmi_stop();
617 lockup_detector_update_enable();
618 watchdog_nmi_start();
619 cpus_read_unlock();
620}
621static inline void lockup_detector_setup(void)
622{
623 lockup_detector_reconfigure();
624}
625#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
626
627static void __lockup_detector_cleanup(void)
628{
629 lockdep_assert_held(&watchdog_mutex);
630 hardlockup_detector_perf_cleanup();
631}
632
633/**
634 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
635 *
636 * Caller must not hold the cpu hotplug rwsem.
637 */
638void lockup_detector_cleanup(void)
639{
640 mutex_lock(&watchdog_mutex);
641 __lockup_detector_cleanup();
642 mutex_unlock(&watchdog_mutex);
643}
644
645/**
646 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
647 *
648 * Special interface for parisc. It prevents lockup detector warnings from
649 * the default pm_poweroff() function which busy loops forever.
650 */
651void lockup_detector_soft_poweroff(void)
652{
653 watchdog_enabled = 0;
654}
655
656#ifdef CONFIG_SYSCTL
657
658/* Propagate any changes to the watchdog threads */
659static void proc_watchdog_update(void)
660{
661 /* Remove impossible cpus to keep sysctl output clean. */
662 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
663 lockup_detector_reconfigure();
664}
665
666/*
667 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
668 *
669 * caller | table->data points to | 'which'
670 * -------------------|----------------------------|--------------------------
671 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
672 * | | SOFT_WATCHDOG_ENABLED
673 * -------------------|----------------------------|--------------------------
674 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
675 * -------------------|----------------------------|--------------------------
676 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
677 */
678static int proc_watchdog_common(int which, struct ctl_table *table, int write,
679 void __user *buffer, size_t *lenp, loff_t *ppos)
680{
681 int err, old, *param = table->data;
682
683 mutex_lock(&watchdog_mutex);
684
685 if (!write) {
686 /*
687 * On read synchronize the userspace interface. This is a
688 * racy snapshot.
689 */
690 *param = (watchdog_enabled & which) != 0;
691 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
692 } else {
693 old = READ_ONCE(*param);
694 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
695 if (!err && old != READ_ONCE(*param))
696 proc_watchdog_update();
697 }
698 mutex_unlock(&watchdog_mutex);
699 return err;
700}
701
702/*
703 * /proc/sys/kernel/watchdog
704 */
705int proc_watchdog(struct ctl_table *table, int write,
706 void __user *buffer, size_t *lenp, loff_t *ppos)
707{
708 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
709 table, write, buffer, lenp, ppos);
710}
711
712/*
713 * /proc/sys/kernel/nmi_watchdog
714 */
715int proc_nmi_watchdog(struct ctl_table *table, int write,
716 void __user *buffer, size_t *lenp, loff_t *ppos)
717{
718 if (!nmi_watchdog_available && write)
719 return -ENOTSUPP;
720 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
721 table, write, buffer, lenp, ppos);
722}
723
724/*
725 * /proc/sys/kernel/soft_watchdog
726 */
727int proc_soft_watchdog(struct ctl_table *table, int write,
728 void __user *buffer, size_t *lenp, loff_t *ppos)
729{
730 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
731 table, write, buffer, lenp, ppos);
732}
733
734/*
735 * /proc/sys/kernel/watchdog_thresh
736 */
737int proc_watchdog_thresh(struct ctl_table *table, int write,
738 void __user *buffer, size_t *lenp, loff_t *ppos)
739{
740 int err, old;
741
742 mutex_lock(&watchdog_mutex);
743
744 old = READ_ONCE(watchdog_thresh);
745 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
746
747 if (!err && write && old != READ_ONCE(watchdog_thresh))
748 proc_watchdog_update();
749
750 mutex_unlock(&watchdog_mutex);
751 return err;
752}
753
754/*
755 * The cpumask is the mask of possible cpus that the watchdog can run
756 * on, not the mask of cpus it is actually running on. This allows the
757 * user to specify a mask that will include cpus that have not yet
758 * been brought online, if desired.
759 */
760int proc_watchdog_cpumask(struct ctl_table *table, int write,
761 void __user *buffer, size_t *lenp, loff_t *ppos)
762{
763 int err;
764
765 mutex_lock(&watchdog_mutex);
766
767 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
768 if (!err && write)
769 proc_watchdog_update();
770
771 mutex_unlock(&watchdog_mutex);
772 return err;
773}
774#endif /* CONFIG_SYSCTL */
775
776void __init lockup_detector_init(void)
777{
778 if (tick_nohz_full_enabled())
779 pr_info("Disabling watchdog on nohz_full cores by default\n");
780
781 cpumask_copy(&watchdog_cpumask,
782 housekeeping_cpumask(HK_FLAG_TIMER));
783
784 if (!watchdog_nmi_probe())
785 nmi_watchdog_available = true;
786 lockup_detector_setup();
787}
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# ifdef CONFIG_SMP
55int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
56# endif /* CONFIG_SMP */
57
58/*
59 * Should we panic when a soft-lockup or hard-lockup occurs:
60 */
61unsigned int __read_mostly hardlockup_panic =
62 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
63/*
64 * We may not want to enable hard lockup detection by default in all cases,
65 * for example when running the kernel as a guest on a hypervisor. In these
66 * cases this function can be called to disable hard lockup detection. This
67 * function should only be executed once by the boot processor before the
68 * kernel command line parameters are parsed, because otherwise it is not
69 * possible to override this in hardlockup_panic_setup().
70 */
71void __init hardlockup_detector_disable(void)
72{
73 nmi_watchdog_user_enabled = 0;
74}
75
76static int __init hardlockup_panic_setup(char *str)
77{
78 if (!strncmp(str, "panic", 5))
79 hardlockup_panic = 1;
80 else if (!strncmp(str, "nopanic", 7))
81 hardlockup_panic = 0;
82 else if (!strncmp(str, "0", 1))
83 nmi_watchdog_user_enabled = 0;
84 else if (!strncmp(str, "1", 1))
85 nmi_watchdog_user_enabled = 1;
86 return 1;
87}
88__setup("nmi_watchdog=", hardlockup_panic_setup);
89
90#endif /* CONFIG_HARDLOCKUP_DETECTOR */
91
92/*
93 * These functions can be overridden if an architecture implements its
94 * own hardlockup detector.
95 *
96 * watchdog_nmi_enable/disable can be implemented to start and stop when
97 * softlockup watchdog threads start and stop. The arch must select the
98 * SOFTLOCKUP_DETECTOR Kconfig.
99 */
100int __weak watchdog_nmi_enable(unsigned int cpu)
101{
102 hardlockup_detector_perf_enable();
103 return 0;
104}
105
106void __weak watchdog_nmi_disable(unsigned int cpu)
107{
108 hardlockup_detector_perf_disable();
109}
110
111/* Return 0, if a NMI watchdog is available. Error code otherwise */
112int __weak __init watchdog_nmi_probe(void)
113{
114 return hardlockup_detector_perf_init();
115}
116
117/**
118 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
119 *
120 * The reconfiguration steps are:
121 * watchdog_nmi_stop();
122 * update_variables();
123 * watchdog_nmi_start();
124 */
125void __weak watchdog_nmi_stop(void) { }
126
127/**
128 * watchdog_nmi_start - Start the watchdog after reconfiguration
129 *
130 * Counterpart to watchdog_nmi_stop().
131 *
132 * The following variables have been updated in update_variables() and
133 * contain the currently valid configuration:
134 * - watchdog_enabled
135 * - watchdog_thresh
136 * - watchdog_cpumask
137 */
138void __weak watchdog_nmi_start(void) { }
139
140/**
141 * lockup_detector_update_enable - Update the sysctl enable bit
142 *
143 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
144 * can't race with watchdog_nmi_disable().
145 */
146static void lockup_detector_update_enable(void)
147{
148 watchdog_enabled = 0;
149 if (!watchdog_user_enabled)
150 return;
151 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
152 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
153 if (soft_watchdog_user_enabled)
154 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
155}
156
157#ifdef CONFIG_SOFTLOCKUP_DETECTOR
158
159#define SOFTLOCKUP_RESET ULONG_MAX
160
161#ifdef CONFIG_SMP
162int __read_mostly sysctl_softlockup_all_cpu_backtrace;
163#endif
164
165/* Global variables, exported for sysctl */
166unsigned int __read_mostly softlockup_panic =
167 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
168
169static bool softlockup_initialized __read_mostly;
170static u64 __read_mostly sample_period;
171
172static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
173static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
174static DEFINE_PER_CPU(bool, softlockup_touch_sync);
175static DEFINE_PER_CPU(bool, soft_watchdog_warn);
176static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
177static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
178static unsigned long soft_lockup_nmi_warn;
179
180static int __init nowatchdog_setup(char *str)
181{
182 watchdog_user_enabled = 0;
183 return 1;
184}
185__setup("nowatchdog", nowatchdog_setup);
186
187static int __init nosoftlockup_setup(char *str)
188{
189 soft_watchdog_user_enabled = 0;
190 return 1;
191}
192__setup("nosoftlockup", nosoftlockup_setup);
193
194static int __init watchdog_thresh_setup(char *str)
195{
196 get_option(&str, &watchdog_thresh);
197 return 1;
198}
199__setup("watchdog_thresh=", watchdog_thresh_setup);
200
201static void __lockup_detector_cleanup(void);
202
203/*
204 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
205 * lockups can have false positives under extreme conditions. So we generally
206 * want a higher threshold for soft lockups than for hard lockups. So we couple
207 * the thresholds with a factor: we make the soft threshold twice the amount of
208 * time the hard threshold is.
209 */
210static int get_softlockup_thresh(void)
211{
212 return watchdog_thresh * 2;
213}
214
215/*
216 * Returns seconds, approximately. We don't need nanosecond
217 * resolution, and we don't need to waste time with a big divide when
218 * 2^30ns == 1.074s.
219 */
220static unsigned long get_timestamp(void)
221{
222 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
223}
224
225static void set_sample_period(void)
226{
227 /*
228 * convert watchdog_thresh from seconds to ns
229 * the divide by 5 is to give hrtimer several chances (two
230 * or three with the current relation between the soft
231 * and hard thresholds) to increment before the
232 * hardlockup detector generates a warning
233 */
234 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
235 watchdog_update_hrtimer_threshold(sample_period);
236}
237
238/* Commands for resetting the watchdog */
239static void __touch_watchdog(void)
240{
241 __this_cpu_write(watchdog_touch_ts, get_timestamp());
242}
243
244/**
245 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
246 *
247 * Call when the scheduler may have stalled for legitimate reasons
248 * preventing the watchdog task from executing - e.g. the scheduler
249 * entering idle state. This should only be used for scheduler events.
250 * Use touch_softlockup_watchdog() for everything else.
251 */
252notrace void touch_softlockup_watchdog_sched(void)
253{
254 /*
255 * Preemption can be enabled. It doesn't matter which CPU's timestamp
256 * gets zeroed here, so use the raw_ operation.
257 */
258 raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
259}
260
261notrace void touch_softlockup_watchdog(void)
262{
263 touch_softlockup_watchdog_sched();
264 wq_watchdog_touch(raw_smp_processor_id());
265}
266EXPORT_SYMBOL(touch_softlockup_watchdog);
267
268void touch_all_softlockup_watchdogs(void)
269{
270 int cpu;
271
272 /*
273 * watchdog_mutex cannpt be taken here, as this might be called
274 * from (soft)interrupt context, so the access to
275 * watchdog_allowed_cpumask might race with a concurrent update.
276 *
277 * The watchdog time stamp can race against a concurrent real
278 * update as well, the only side effect might be a cycle delay for
279 * the softlockup check.
280 */
281 for_each_cpu(cpu, &watchdog_allowed_mask)
282 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
283 wq_watchdog_touch(-1);
284}
285
286void touch_softlockup_watchdog_sync(void)
287{
288 __this_cpu_write(softlockup_touch_sync, true);
289 __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
290}
291
292static int is_softlockup(unsigned long touch_ts)
293{
294 unsigned long now = get_timestamp();
295
296 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
297 /* Warn about unreasonable delays. */
298 if (time_after(now, touch_ts + get_softlockup_thresh()))
299 return now - touch_ts;
300 }
301 return 0;
302}
303
304/* watchdog detector functions */
305bool is_hardlockup(void)
306{
307 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
308
309 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
310 return true;
311
312 __this_cpu_write(hrtimer_interrupts_saved, hrint);
313 return false;
314}
315
316static void watchdog_interrupt_count(void)
317{
318 __this_cpu_inc(hrtimer_interrupts);
319}
320
321static DEFINE_PER_CPU(struct completion, softlockup_completion);
322static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
323
324/*
325 * The watchdog thread function - touches the timestamp.
326 *
327 * It only runs once every sample_period seconds (4 seconds by
328 * default) to reset the softlockup timestamp. If this gets delayed
329 * for more than 2*watchdog_thresh seconds then the debug-printout
330 * triggers in watchdog_timer_fn().
331 */
332static int softlockup_fn(void *data)
333{
334 __touch_watchdog();
335 complete(this_cpu_ptr(&softlockup_completion));
336
337 return 0;
338}
339
340/* watchdog kicker functions */
341static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
342{
343 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
344 struct pt_regs *regs = get_irq_regs();
345 int duration;
346 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
347
348 if (!watchdog_enabled)
349 return HRTIMER_NORESTART;
350
351 /* kick the hardlockup detector */
352 watchdog_interrupt_count();
353
354 /* kick the softlockup detector */
355 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
356 reinit_completion(this_cpu_ptr(&softlockup_completion));
357 stop_one_cpu_nowait(smp_processor_id(),
358 softlockup_fn, NULL,
359 this_cpu_ptr(&softlockup_stop_work));
360 }
361
362 /* .. and repeat */
363 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
364
365 if (touch_ts == SOFTLOCKUP_RESET) {
366 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
367 /*
368 * If the time stamp was touched atomically
369 * make sure the scheduler tick is up to date.
370 */
371 __this_cpu_write(softlockup_touch_sync, false);
372 sched_clock_tick();
373 }
374
375 /* Clear the guest paused flag on watchdog reset */
376 kvm_check_and_clear_guest_paused();
377 __touch_watchdog();
378 return HRTIMER_RESTART;
379 }
380
381 /* check for a softlockup
382 * This is done by making sure a high priority task is
383 * being scheduled. The task touches the watchdog to
384 * indicate it is getting cpu time. If it hasn't then
385 * this is a good indication some task is hogging the cpu
386 */
387 duration = is_softlockup(touch_ts);
388 if (unlikely(duration)) {
389 /*
390 * If a virtual machine is stopped by the host it can look to
391 * the watchdog like a soft lockup, check to see if the host
392 * stopped the vm before we issue the warning
393 */
394 if (kvm_check_and_clear_guest_paused())
395 return HRTIMER_RESTART;
396
397 /* only warn once */
398 if (__this_cpu_read(soft_watchdog_warn) == true)
399 return HRTIMER_RESTART;
400
401 if (softlockup_all_cpu_backtrace) {
402 /* Prevent multiple soft-lockup reports if one cpu is already
403 * engaged in dumping cpu back traces
404 */
405 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
406 /* Someone else will report us. Let's give up */
407 __this_cpu_write(soft_watchdog_warn, true);
408 return HRTIMER_RESTART;
409 }
410 }
411
412 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
413 smp_processor_id(), duration,
414 current->comm, task_pid_nr(current));
415 print_modules();
416 print_irqtrace_events(current);
417 if (regs)
418 show_regs(regs);
419 else
420 dump_stack();
421
422 if (softlockup_all_cpu_backtrace) {
423 /* Avoid generating two back traces for current
424 * given that one is already made above
425 */
426 trigger_allbutself_cpu_backtrace();
427
428 clear_bit(0, &soft_lockup_nmi_warn);
429 /* Barrier to sync with other cpus */
430 smp_mb__after_atomic();
431 }
432
433 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
434 if (softlockup_panic)
435 panic("softlockup: hung tasks");
436 __this_cpu_write(soft_watchdog_warn, true);
437 } else
438 __this_cpu_write(soft_watchdog_warn, false);
439
440 return HRTIMER_RESTART;
441}
442
443static void watchdog_enable(unsigned int cpu)
444{
445 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
446 struct completion *done = this_cpu_ptr(&softlockup_completion);
447
448 WARN_ON_ONCE(cpu != smp_processor_id());
449
450 init_completion(done);
451 complete(done);
452
453 /*
454 * Start the timer first to prevent the NMI watchdog triggering
455 * before the timer has a chance to fire.
456 */
457 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
458 hrtimer->function = watchdog_timer_fn;
459 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
460 HRTIMER_MODE_REL_PINNED_HARD);
461
462 /* Initialize timestamp */
463 __touch_watchdog();
464 /* Enable the perf event */
465 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
466 watchdog_nmi_enable(cpu);
467}
468
469static void watchdog_disable(unsigned int cpu)
470{
471 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
472
473 WARN_ON_ONCE(cpu != smp_processor_id());
474
475 /*
476 * Disable the perf event first. That prevents that a large delay
477 * between disabling the timer and disabling the perf event causes
478 * the perf NMI to detect a false positive.
479 */
480 watchdog_nmi_disable(cpu);
481 hrtimer_cancel(hrtimer);
482 wait_for_completion(this_cpu_ptr(&softlockup_completion));
483}
484
485static int softlockup_stop_fn(void *data)
486{
487 watchdog_disable(smp_processor_id());
488 return 0;
489}
490
491static void softlockup_stop_all(void)
492{
493 int cpu;
494
495 if (!softlockup_initialized)
496 return;
497
498 for_each_cpu(cpu, &watchdog_allowed_mask)
499 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
500
501 cpumask_clear(&watchdog_allowed_mask);
502}
503
504static int softlockup_start_fn(void *data)
505{
506 watchdog_enable(smp_processor_id());
507 return 0;
508}
509
510static void softlockup_start_all(void)
511{
512 int cpu;
513
514 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
515 for_each_cpu(cpu, &watchdog_allowed_mask)
516 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
517}
518
519int lockup_detector_online_cpu(unsigned int cpu)
520{
521 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
522 watchdog_enable(cpu);
523 return 0;
524}
525
526int lockup_detector_offline_cpu(unsigned int cpu)
527{
528 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
529 watchdog_disable(cpu);
530 return 0;
531}
532
533static void lockup_detector_reconfigure(void)
534{
535 cpus_read_lock();
536 watchdog_nmi_stop();
537
538 softlockup_stop_all();
539 set_sample_period();
540 lockup_detector_update_enable();
541 if (watchdog_enabled && watchdog_thresh)
542 softlockup_start_all();
543
544 watchdog_nmi_start();
545 cpus_read_unlock();
546 /*
547 * Must be called outside the cpus locked section to prevent
548 * recursive locking in the perf code.
549 */
550 __lockup_detector_cleanup();
551}
552
553/*
554 * Create the watchdog thread infrastructure and configure the detector(s).
555 *
556 * The threads are not unparked as watchdog_allowed_mask is empty. When
557 * the threads are successfully initialized, take the proper locks and
558 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
559 */
560static __init void lockup_detector_setup(void)
561{
562 /*
563 * If sysctl is off and watchdog got disabled on the command line,
564 * nothing to do here.
565 */
566 lockup_detector_update_enable();
567
568 if (!IS_ENABLED(CONFIG_SYSCTL) &&
569 !(watchdog_enabled && watchdog_thresh))
570 return;
571
572 mutex_lock(&watchdog_mutex);
573 lockup_detector_reconfigure();
574 softlockup_initialized = true;
575 mutex_unlock(&watchdog_mutex);
576}
577
578#else /* CONFIG_SOFTLOCKUP_DETECTOR */
579static void lockup_detector_reconfigure(void)
580{
581 cpus_read_lock();
582 watchdog_nmi_stop();
583 lockup_detector_update_enable();
584 watchdog_nmi_start();
585 cpus_read_unlock();
586}
587static inline void lockup_detector_setup(void)
588{
589 lockup_detector_reconfigure();
590}
591#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
592
593static void __lockup_detector_cleanup(void)
594{
595 lockdep_assert_held(&watchdog_mutex);
596 hardlockup_detector_perf_cleanup();
597}
598
599/**
600 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
601 *
602 * Caller must not hold the cpu hotplug rwsem.
603 */
604void lockup_detector_cleanup(void)
605{
606 mutex_lock(&watchdog_mutex);
607 __lockup_detector_cleanup();
608 mutex_unlock(&watchdog_mutex);
609}
610
611/**
612 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
613 *
614 * Special interface for parisc. It prevents lockup detector warnings from
615 * the default pm_poweroff() function which busy loops forever.
616 */
617void lockup_detector_soft_poweroff(void)
618{
619 watchdog_enabled = 0;
620}
621
622#ifdef CONFIG_SYSCTL
623
624/* Propagate any changes to the watchdog threads */
625static void proc_watchdog_update(void)
626{
627 /* Remove impossible cpus to keep sysctl output clean. */
628 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
629 lockup_detector_reconfigure();
630}
631
632/*
633 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
634 *
635 * caller | table->data points to | 'which'
636 * -------------------|----------------------------|--------------------------
637 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
638 * | | SOFT_WATCHDOG_ENABLED
639 * -------------------|----------------------------|--------------------------
640 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
641 * -------------------|----------------------------|--------------------------
642 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
643 */
644static int proc_watchdog_common(int which, struct ctl_table *table, int write,
645 void *buffer, size_t *lenp, loff_t *ppos)
646{
647 int err, old, *param = table->data;
648
649 mutex_lock(&watchdog_mutex);
650
651 if (!write) {
652 /*
653 * On read synchronize the userspace interface. This is a
654 * racy snapshot.
655 */
656 *param = (watchdog_enabled & which) != 0;
657 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
658 } else {
659 old = READ_ONCE(*param);
660 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
661 if (!err && old != READ_ONCE(*param))
662 proc_watchdog_update();
663 }
664 mutex_unlock(&watchdog_mutex);
665 return err;
666}
667
668/*
669 * /proc/sys/kernel/watchdog
670 */
671int proc_watchdog(struct ctl_table *table, int write,
672 void *buffer, size_t *lenp, loff_t *ppos)
673{
674 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
675 table, write, buffer, lenp, ppos);
676}
677
678/*
679 * /proc/sys/kernel/nmi_watchdog
680 */
681int proc_nmi_watchdog(struct ctl_table *table, int write,
682 void *buffer, size_t *lenp, loff_t *ppos)
683{
684 if (!nmi_watchdog_available && write)
685 return -ENOTSUPP;
686 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
687 table, write, buffer, lenp, ppos);
688}
689
690/*
691 * /proc/sys/kernel/soft_watchdog
692 */
693int proc_soft_watchdog(struct ctl_table *table, int write,
694 void *buffer, size_t *lenp, loff_t *ppos)
695{
696 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
697 table, write, buffer, lenp, ppos);
698}
699
700/*
701 * /proc/sys/kernel/watchdog_thresh
702 */
703int proc_watchdog_thresh(struct ctl_table *table, int write,
704 void *buffer, size_t *lenp, loff_t *ppos)
705{
706 int err, old;
707
708 mutex_lock(&watchdog_mutex);
709
710 old = READ_ONCE(watchdog_thresh);
711 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
712
713 if (!err && write && old != READ_ONCE(watchdog_thresh))
714 proc_watchdog_update();
715
716 mutex_unlock(&watchdog_mutex);
717 return err;
718}
719
720/*
721 * The cpumask is the mask of possible cpus that the watchdog can run
722 * on, not the mask of cpus it is actually running on. This allows the
723 * user to specify a mask that will include cpus that have not yet
724 * been brought online, if desired.
725 */
726int proc_watchdog_cpumask(struct ctl_table *table, int write,
727 void *buffer, size_t *lenp, loff_t *ppos)
728{
729 int err;
730
731 mutex_lock(&watchdog_mutex);
732
733 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
734 if (!err && write)
735 proc_watchdog_update();
736
737 mutex_unlock(&watchdog_mutex);
738 return err;
739}
740#endif /* CONFIG_SYSCTL */
741
742void __init lockup_detector_init(void)
743{
744 if (tick_nohz_full_enabled())
745 pr_info("Disabling watchdog on nohz_full cores by default\n");
746
747 cpumask_copy(&watchdog_cpumask,
748 housekeeping_cpumask(HK_FLAG_TIMER));
749
750 if (!watchdog_nmi_probe())
751 nmi_watchdog_available = true;
752 lockup_detector_setup();
753}