Loading...
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}
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_HARDLOCKUP_DETECTOR_SPARC64)
33# define WATCHDOG_HARDLOCKUP_DEFAULT 1
34#else
35# define WATCHDOG_HARDLOCKUP_DEFAULT 0
36#endif
37
38unsigned long __read_mostly watchdog_enabled;
39int __read_mostly watchdog_user_enabled = 1;
40static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT;
41static int __read_mostly watchdog_softlockup_user_enabled = 1;
42int __read_mostly watchdog_thresh = 10;
43static int __read_mostly watchdog_hardlockup_available;
44
45struct cpumask watchdog_cpumask __read_mostly;
46unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
47
48#ifdef CONFIG_HARDLOCKUP_DETECTOR
49
50# ifdef CONFIG_SMP
51int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
52# endif /* CONFIG_SMP */
53
54/*
55 * Should we panic when a soft-lockup or hard-lockup occurs:
56 */
57unsigned int __read_mostly hardlockup_panic =
58 IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
59/*
60 * We may not want to enable hard lockup detection by default in all cases,
61 * for example when running the kernel as a guest on a hypervisor. In these
62 * cases this function can be called to disable hard lockup detection. This
63 * function should only be executed once by the boot processor before the
64 * kernel command line parameters are parsed, because otherwise it is not
65 * possible to override this in hardlockup_panic_setup().
66 */
67void __init hardlockup_detector_disable(void)
68{
69 watchdog_hardlockup_user_enabled = 0;
70}
71
72static int __init hardlockup_panic_setup(char *str)
73{
74 if (!strncmp(str, "panic", 5))
75 hardlockup_panic = 1;
76 else if (!strncmp(str, "nopanic", 7))
77 hardlockup_panic = 0;
78 else if (!strncmp(str, "0", 1))
79 watchdog_hardlockup_user_enabled = 0;
80 else if (!strncmp(str, "1", 1))
81 watchdog_hardlockup_user_enabled = 1;
82 return 1;
83}
84__setup("nmi_watchdog=", hardlockup_panic_setup);
85
86#endif /* CONFIG_HARDLOCKUP_DETECTOR */
87
88#if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER)
89
90static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts);
91static DEFINE_PER_CPU(int, hrtimer_interrupts_saved);
92static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned);
93static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched);
94static unsigned long hard_lockup_nmi_warn;
95
96notrace void arch_touch_nmi_watchdog(void)
97{
98 /*
99 * Using __raw here because some code paths have
100 * preemption enabled. If preemption is enabled
101 * then interrupts should be enabled too, in which
102 * case we shouldn't have to worry about the watchdog
103 * going off.
104 */
105 raw_cpu_write(watchdog_hardlockup_touched, true);
106}
107EXPORT_SYMBOL(arch_touch_nmi_watchdog);
108
109void watchdog_hardlockup_touch_cpu(unsigned int cpu)
110{
111 per_cpu(watchdog_hardlockup_touched, cpu) = true;
112}
113
114static bool is_hardlockup(unsigned int cpu)
115{
116 int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu));
117
118 if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
119 return true;
120
121 /*
122 * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE
123 * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is
124 * written/read by a single CPU.
125 */
126 per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
127
128 return false;
129}
130
131static void watchdog_hardlockup_kick(void)
132{
133 int new_interrupts;
134
135 new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts));
136 watchdog_buddy_check_hardlockup(new_interrupts);
137}
138
139void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
140{
141 if (per_cpu(watchdog_hardlockup_touched, cpu)) {
142 per_cpu(watchdog_hardlockup_touched, cpu) = false;
143 return;
144 }
145
146 /*
147 * Check for a hardlockup by making sure the CPU's timer
148 * interrupt is incrementing. The timer interrupt should have
149 * fired multiple times before we overflow'd. If it hasn't
150 * then this is a good indication the cpu is stuck
151 */
152 if (is_hardlockup(cpu)) {
153 unsigned int this_cpu = smp_processor_id();
154 unsigned long flags;
155
156 /* Only print hardlockups once. */
157 if (per_cpu(watchdog_hardlockup_warned, cpu))
158 return;
159
160 /*
161 * Prevent multiple hard-lockup reports if one cpu is already
162 * engaged in dumping all cpu back traces.
163 */
164 if (sysctl_hardlockup_all_cpu_backtrace) {
165 if (test_and_set_bit_lock(0, &hard_lockup_nmi_warn))
166 return;
167 }
168
169 /*
170 * NOTE: we call printk_cpu_sync_get_irqsave() after printing
171 * the lockup message. While it would be nice to serialize
172 * that printout, we really want to make sure that if some
173 * other CPU somehow locked up while holding the lock associated
174 * with printk_cpu_sync_get_irqsave() that we can still at least
175 * get the message about the lockup out.
176 */
177 pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
178 printk_cpu_sync_get_irqsave(flags);
179
180 print_modules();
181 print_irqtrace_events(current);
182 if (cpu == this_cpu) {
183 if (regs)
184 show_regs(regs);
185 else
186 dump_stack();
187 printk_cpu_sync_put_irqrestore(flags);
188 } else {
189 printk_cpu_sync_put_irqrestore(flags);
190 trigger_single_cpu_backtrace(cpu);
191 }
192
193 if (sysctl_hardlockup_all_cpu_backtrace) {
194 trigger_allbutcpu_cpu_backtrace(cpu);
195 if (!hardlockup_panic)
196 clear_bit_unlock(0, &hard_lockup_nmi_warn);
197 }
198
199 if (hardlockup_panic)
200 nmi_panic(regs, "Hard LOCKUP");
201
202 per_cpu(watchdog_hardlockup_warned, cpu) = true;
203 } else {
204 per_cpu(watchdog_hardlockup_warned, cpu) = false;
205 }
206}
207
208#else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
209
210static inline void watchdog_hardlockup_kick(void) { }
211
212#endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
213
214/*
215 * These functions can be overridden based on the configured hardlockdup detector.
216 *
217 * watchdog_hardlockup_enable/disable can be implemented to start and stop when
218 * softlockup watchdog start and stop. The detector must select the
219 * SOFTLOCKUP_DETECTOR Kconfig.
220 */
221void __weak watchdog_hardlockup_enable(unsigned int cpu) { }
222
223void __weak watchdog_hardlockup_disable(unsigned int cpu) { }
224
225/*
226 * Watchdog-detector specific API.
227 *
228 * Return 0 when hardlockup watchdog is available, negative value otherwise.
229 * Note that the negative value means that a delayed probe might
230 * succeed later.
231 */
232int __weak __init watchdog_hardlockup_probe(void)
233{
234 return -ENODEV;
235}
236
237/**
238 * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration
239 *
240 * The reconfiguration steps are:
241 * watchdog_hardlockup_stop();
242 * update_variables();
243 * watchdog_hardlockup_start();
244 */
245void __weak watchdog_hardlockup_stop(void) { }
246
247/**
248 * watchdog_hardlockup_start - Start the watchdog after reconfiguration
249 *
250 * Counterpart to watchdog_hardlockup_stop().
251 *
252 * The following variables have been updated in update_variables() and
253 * contain the currently valid configuration:
254 * - watchdog_enabled
255 * - watchdog_thresh
256 * - watchdog_cpumask
257 */
258void __weak watchdog_hardlockup_start(void) { }
259
260/**
261 * lockup_detector_update_enable - Update the sysctl enable bit
262 *
263 * Caller needs to make sure that the hard watchdogs are off, so this
264 * can't race with watchdog_hardlockup_disable().
265 */
266static void lockup_detector_update_enable(void)
267{
268 watchdog_enabled = 0;
269 if (!watchdog_user_enabled)
270 return;
271 if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled)
272 watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED;
273 if (watchdog_softlockup_user_enabled)
274 watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED;
275}
276
277#ifdef CONFIG_SOFTLOCKUP_DETECTOR
278
279/*
280 * Delay the soflockup report when running a known slow code.
281 * It does _not_ affect the timestamp of the last successdul reschedule.
282 */
283#define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
284
285#ifdef CONFIG_SMP
286int __read_mostly sysctl_softlockup_all_cpu_backtrace;
287#endif
288
289static struct cpumask watchdog_allowed_mask __read_mostly;
290
291/* Global variables, exported for sysctl */
292unsigned int __read_mostly softlockup_panic =
293 IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
294
295static bool softlockup_initialized __read_mostly;
296static u64 __read_mostly sample_period;
297
298/* Timestamp taken after the last successful reschedule. */
299static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
300/* Timestamp of the last softlockup report. */
301static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
302static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
303static DEFINE_PER_CPU(bool, softlockup_touch_sync);
304static unsigned long soft_lockup_nmi_warn;
305
306static int __init softlockup_panic_setup(char *str)
307{
308 softlockup_panic = simple_strtoul(str, NULL, 0);
309 return 1;
310}
311__setup("softlockup_panic=", softlockup_panic_setup);
312
313static int __init nowatchdog_setup(char *str)
314{
315 watchdog_user_enabled = 0;
316 return 1;
317}
318__setup("nowatchdog", nowatchdog_setup);
319
320static int __init nosoftlockup_setup(char *str)
321{
322 watchdog_softlockup_user_enabled = 0;
323 return 1;
324}
325__setup("nosoftlockup", nosoftlockup_setup);
326
327static int __init watchdog_thresh_setup(char *str)
328{
329 get_option(&str, &watchdog_thresh);
330 return 1;
331}
332__setup("watchdog_thresh=", watchdog_thresh_setup);
333
334static void __lockup_detector_cleanup(void);
335
336/*
337 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
338 * lockups can have false positives under extreme conditions. So we generally
339 * want a higher threshold for soft lockups than for hard lockups. So we couple
340 * the thresholds with a factor: we make the soft threshold twice the amount of
341 * time the hard threshold is.
342 */
343static int get_softlockup_thresh(void)
344{
345 return watchdog_thresh * 2;
346}
347
348/*
349 * Returns seconds, approximately. We don't need nanosecond
350 * resolution, and we don't need to waste time with a big divide when
351 * 2^30ns == 1.074s.
352 */
353static unsigned long get_timestamp(void)
354{
355 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
356}
357
358static void set_sample_period(void)
359{
360 /*
361 * convert watchdog_thresh from seconds to ns
362 * the divide by 5 is to give hrtimer several chances (two
363 * or three with the current relation between the soft
364 * and hard thresholds) to increment before the
365 * hardlockup detector generates a warning
366 */
367 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
368 watchdog_update_hrtimer_threshold(sample_period);
369}
370
371static void update_report_ts(void)
372{
373 __this_cpu_write(watchdog_report_ts, get_timestamp());
374}
375
376/* Commands for resetting the watchdog */
377static void update_touch_ts(void)
378{
379 __this_cpu_write(watchdog_touch_ts, get_timestamp());
380 update_report_ts();
381}
382
383/**
384 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
385 *
386 * Call when the scheduler may have stalled for legitimate reasons
387 * preventing the watchdog task from executing - e.g. the scheduler
388 * entering idle state. This should only be used for scheduler events.
389 * Use touch_softlockup_watchdog() for everything else.
390 */
391notrace void touch_softlockup_watchdog_sched(void)
392{
393 /*
394 * Preemption can be enabled. It doesn't matter which CPU's watchdog
395 * report period gets restarted here, so use the raw_ operation.
396 */
397 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
398}
399
400notrace void touch_softlockup_watchdog(void)
401{
402 touch_softlockup_watchdog_sched();
403 wq_watchdog_touch(raw_smp_processor_id());
404}
405EXPORT_SYMBOL(touch_softlockup_watchdog);
406
407void touch_all_softlockup_watchdogs(void)
408{
409 int cpu;
410
411 /*
412 * watchdog_mutex cannpt be taken here, as this might be called
413 * from (soft)interrupt context, so the access to
414 * watchdog_allowed_cpumask might race with a concurrent update.
415 *
416 * The watchdog time stamp can race against a concurrent real
417 * update as well, the only side effect might be a cycle delay for
418 * the softlockup check.
419 */
420 for_each_cpu(cpu, &watchdog_allowed_mask) {
421 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
422 wq_watchdog_touch(cpu);
423 }
424}
425
426void touch_softlockup_watchdog_sync(void)
427{
428 __this_cpu_write(softlockup_touch_sync, true);
429 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
430}
431
432static int is_softlockup(unsigned long touch_ts,
433 unsigned long period_ts,
434 unsigned long now)
435{
436 if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
437 /* Warn about unreasonable delays. */
438 if (time_after(now, period_ts + get_softlockup_thresh()))
439 return now - touch_ts;
440 }
441 return 0;
442}
443
444/* watchdog detector functions */
445static DEFINE_PER_CPU(struct completion, softlockup_completion);
446static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
447
448/*
449 * The watchdog feed function - touches the timestamp.
450 *
451 * It only runs once every sample_period seconds (4 seconds by
452 * default) to reset the softlockup timestamp. If this gets delayed
453 * for more than 2*watchdog_thresh seconds then the debug-printout
454 * triggers in watchdog_timer_fn().
455 */
456static int softlockup_fn(void *data)
457{
458 update_touch_ts();
459 complete(this_cpu_ptr(&softlockup_completion));
460
461 return 0;
462}
463
464/* watchdog kicker functions */
465static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
466{
467 unsigned long touch_ts, period_ts, now;
468 struct pt_regs *regs = get_irq_regs();
469 int duration;
470 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
471 unsigned long flags;
472
473 if (!watchdog_enabled)
474 return HRTIMER_NORESTART;
475
476 watchdog_hardlockup_kick();
477
478 /* kick the softlockup detector */
479 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
480 reinit_completion(this_cpu_ptr(&softlockup_completion));
481 stop_one_cpu_nowait(smp_processor_id(),
482 softlockup_fn, NULL,
483 this_cpu_ptr(&softlockup_stop_work));
484 }
485
486 /* .. and repeat */
487 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
488
489 /*
490 * Read the current timestamp first. It might become invalid anytime
491 * when a virtual machine is stopped by the host or when the watchog
492 * is touched from NMI.
493 */
494 now = get_timestamp();
495 /*
496 * If a virtual machine is stopped by the host it can look to
497 * the watchdog like a soft lockup. This function touches the watchdog.
498 */
499 kvm_check_and_clear_guest_paused();
500 /*
501 * The stored timestamp is comparable with @now only when not touched.
502 * It might get touched anytime from NMI. Make sure that is_softlockup()
503 * uses the same (valid) value.
504 */
505 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
506
507 /* Reset the interval when touched by known problematic code. */
508 if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
509 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
510 /*
511 * If the time stamp was touched atomically
512 * make sure the scheduler tick is up to date.
513 */
514 __this_cpu_write(softlockup_touch_sync, false);
515 sched_clock_tick();
516 }
517
518 update_report_ts();
519 return HRTIMER_RESTART;
520 }
521
522 /* Check for a softlockup. */
523 touch_ts = __this_cpu_read(watchdog_touch_ts);
524 duration = is_softlockup(touch_ts, period_ts, now);
525 if (unlikely(duration)) {
526 /*
527 * Prevent multiple soft-lockup reports if one cpu is already
528 * engaged in dumping all cpu back traces.
529 */
530 if (softlockup_all_cpu_backtrace) {
531 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
532 return HRTIMER_RESTART;
533 }
534
535 /* Start period for the next softlockup warning. */
536 update_report_ts();
537
538 printk_cpu_sync_get_irqsave(flags);
539 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
540 smp_processor_id(), duration,
541 current->comm, task_pid_nr(current));
542 print_modules();
543 print_irqtrace_events(current);
544 if (regs)
545 show_regs(regs);
546 else
547 dump_stack();
548 printk_cpu_sync_put_irqrestore(flags);
549
550 if (softlockup_all_cpu_backtrace) {
551 trigger_allbutcpu_cpu_backtrace(smp_processor_id());
552 if (!softlockup_panic)
553 clear_bit_unlock(0, &soft_lockup_nmi_warn);
554 }
555
556 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
557 if (softlockup_panic)
558 panic("softlockup: hung tasks");
559 }
560
561 return HRTIMER_RESTART;
562}
563
564static void watchdog_enable(unsigned int cpu)
565{
566 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
567 struct completion *done = this_cpu_ptr(&softlockup_completion);
568
569 WARN_ON_ONCE(cpu != smp_processor_id());
570
571 init_completion(done);
572 complete(done);
573
574 /*
575 * Start the timer first to prevent the hardlockup watchdog triggering
576 * before the timer has a chance to fire.
577 */
578 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
579 hrtimer->function = watchdog_timer_fn;
580 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
581 HRTIMER_MODE_REL_PINNED_HARD);
582
583 /* Initialize timestamp */
584 update_touch_ts();
585 /* Enable the hardlockup detector */
586 if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)
587 watchdog_hardlockup_enable(cpu);
588}
589
590static void watchdog_disable(unsigned int cpu)
591{
592 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
593
594 WARN_ON_ONCE(cpu != smp_processor_id());
595
596 /*
597 * Disable the hardlockup detector first. That prevents that a large
598 * delay between disabling the timer and disabling the hardlockup
599 * detector causes a false positive.
600 */
601 watchdog_hardlockup_disable(cpu);
602 hrtimer_cancel(hrtimer);
603 wait_for_completion(this_cpu_ptr(&softlockup_completion));
604}
605
606static int softlockup_stop_fn(void *data)
607{
608 watchdog_disable(smp_processor_id());
609 return 0;
610}
611
612static void softlockup_stop_all(void)
613{
614 int cpu;
615
616 if (!softlockup_initialized)
617 return;
618
619 for_each_cpu(cpu, &watchdog_allowed_mask)
620 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
621
622 cpumask_clear(&watchdog_allowed_mask);
623}
624
625static int softlockup_start_fn(void *data)
626{
627 watchdog_enable(smp_processor_id());
628 return 0;
629}
630
631static void softlockup_start_all(void)
632{
633 int cpu;
634
635 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
636 for_each_cpu(cpu, &watchdog_allowed_mask)
637 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
638}
639
640int lockup_detector_online_cpu(unsigned int cpu)
641{
642 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
643 watchdog_enable(cpu);
644 return 0;
645}
646
647int lockup_detector_offline_cpu(unsigned int cpu)
648{
649 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
650 watchdog_disable(cpu);
651 return 0;
652}
653
654static void __lockup_detector_reconfigure(void)
655{
656 cpus_read_lock();
657 watchdog_hardlockup_stop();
658
659 softlockup_stop_all();
660 set_sample_period();
661 lockup_detector_update_enable();
662 if (watchdog_enabled && watchdog_thresh)
663 softlockup_start_all();
664
665 watchdog_hardlockup_start();
666 cpus_read_unlock();
667 /*
668 * Must be called outside the cpus locked section to prevent
669 * recursive locking in the perf code.
670 */
671 __lockup_detector_cleanup();
672}
673
674void lockup_detector_reconfigure(void)
675{
676 mutex_lock(&watchdog_mutex);
677 __lockup_detector_reconfigure();
678 mutex_unlock(&watchdog_mutex);
679}
680
681/*
682 * Create the watchdog infrastructure and configure the detector(s).
683 */
684static __init void lockup_detector_setup(void)
685{
686 /*
687 * If sysctl is off and watchdog got disabled on the command line,
688 * nothing to do here.
689 */
690 lockup_detector_update_enable();
691
692 if (!IS_ENABLED(CONFIG_SYSCTL) &&
693 !(watchdog_enabled && watchdog_thresh))
694 return;
695
696 mutex_lock(&watchdog_mutex);
697 __lockup_detector_reconfigure();
698 softlockup_initialized = true;
699 mutex_unlock(&watchdog_mutex);
700}
701
702#else /* CONFIG_SOFTLOCKUP_DETECTOR */
703static void __lockup_detector_reconfigure(void)
704{
705 cpus_read_lock();
706 watchdog_hardlockup_stop();
707 lockup_detector_update_enable();
708 watchdog_hardlockup_start();
709 cpus_read_unlock();
710}
711void lockup_detector_reconfigure(void)
712{
713 __lockup_detector_reconfigure();
714}
715static inline void lockup_detector_setup(void)
716{
717 __lockup_detector_reconfigure();
718}
719#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
720
721static void __lockup_detector_cleanup(void)
722{
723 lockdep_assert_held(&watchdog_mutex);
724 hardlockup_detector_perf_cleanup();
725}
726
727/**
728 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
729 *
730 * Caller must not hold the cpu hotplug rwsem.
731 */
732void lockup_detector_cleanup(void)
733{
734 mutex_lock(&watchdog_mutex);
735 __lockup_detector_cleanup();
736 mutex_unlock(&watchdog_mutex);
737}
738
739/**
740 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
741 *
742 * Special interface for parisc. It prevents lockup detector warnings from
743 * the default pm_poweroff() function which busy loops forever.
744 */
745void lockup_detector_soft_poweroff(void)
746{
747 watchdog_enabled = 0;
748}
749
750#ifdef CONFIG_SYSCTL
751
752/* Propagate any changes to the watchdog infrastructure */
753static void proc_watchdog_update(void)
754{
755 /* Remove impossible cpus to keep sysctl output clean. */
756 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
757 __lockup_detector_reconfigure();
758}
759
760/*
761 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
762 *
763 * caller | table->data points to | 'which'
764 * -------------------|----------------------------------|-------------------------------
765 * proc_watchdog | watchdog_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED |
766 * | | WATCHDOG_SOFTOCKUP_ENABLED
767 * -------------------|----------------------------------|-------------------------------
768 * proc_nmi_watchdog | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED
769 * -------------------|----------------------------------|-------------------------------
770 * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED
771 */
772static int proc_watchdog_common(int which, struct ctl_table *table, int write,
773 void *buffer, size_t *lenp, loff_t *ppos)
774{
775 int err, old, *param = table->data;
776
777 mutex_lock(&watchdog_mutex);
778
779 if (!write) {
780 /*
781 * On read synchronize the userspace interface. This is a
782 * racy snapshot.
783 */
784 *param = (watchdog_enabled & which) != 0;
785 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
786 } else {
787 old = READ_ONCE(*param);
788 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
789 if (!err && old != READ_ONCE(*param))
790 proc_watchdog_update();
791 }
792 mutex_unlock(&watchdog_mutex);
793 return err;
794}
795
796/*
797 * /proc/sys/kernel/watchdog
798 */
799static int proc_watchdog(struct ctl_table *table, int write,
800 void *buffer, size_t *lenp, loff_t *ppos)
801{
802 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED |
803 WATCHDOG_SOFTOCKUP_ENABLED,
804 table, write, buffer, lenp, ppos);
805}
806
807/*
808 * /proc/sys/kernel/nmi_watchdog
809 */
810static int proc_nmi_watchdog(struct ctl_table *table, int write,
811 void *buffer, size_t *lenp, loff_t *ppos)
812{
813 if (!watchdog_hardlockup_available && write)
814 return -ENOTSUPP;
815 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED,
816 table, write, buffer, lenp, ppos);
817}
818
819#ifdef CONFIG_SOFTLOCKUP_DETECTOR
820/*
821 * /proc/sys/kernel/soft_watchdog
822 */
823static int proc_soft_watchdog(struct ctl_table *table, int write,
824 void *buffer, size_t *lenp, loff_t *ppos)
825{
826 return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED,
827 table, write, buffer, lenp, ppos);
828}
829#endif
830
831/*
832 * /proc/sys/kernel/watchdog_thresh
833 */
834static int proc_watchdog_thresh(struct ctl_table *table, int write,
835 void *buffer, size_t *lenp, loff_t *ppos)
836{
837 int err, old;
838
839 mutex_lock(&watchdog_mutex);
840
841 old = READ_ONCE(watchdog_thresh);
842 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
843
844 if (!err && write && old != READ_ONCE(watchdog_thresh))
845 proc_watchdog_update();
846
847 mutex_unlock(&watchdog_mutex);
848 return err;
849}
850
851/*
852 * The cpumask is the mask of possible cpus that the watchdog can run
853 * on, not the mask of cpus it is actually running on. This allows the
854 * user to specify a mask that will include cpus that have not yet
855 * been brought online, if desired.
856 */
857static int proc_watchdog_cpumask(struct ctl_table *table, int write,
858 void *buffer, size_t *lenp, loff_t *ppos)
859{
860 int err;
861
862 mutex_lock(&watchdog_mutex);
863
864 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
865 if (!err && write)
866 proc_watchdog_update();
867
868 mutex_unlock(&watchdog_mutex);
869 return err;
870}
871
872static const int sixty = 60;
873
874static struct ctl_table watchdog_sysctls[] = {
875 {
876 .procname = "watchdog",
877 .data = &watchdog_user_enabled,
878 .maxlen = sizeof(int),
879 .mode = 0644,
880 .proc_handler = proc_watchdog,
881 .extra1 = SYSCTL_ZERO,
882 .extra2 = SYSCTL_ONE,
883 },
884 {
885 .procname = "watchdog_thresh",
886 .data = &watchdog_thresh,
887 .maxlen = sizeof(int),
888 .mode = 0644,
889 .proc_handler = proc_watchdog_thresh,
890 .extra1 = SYSCTL_ZERO,
891 .extra2 = (void *)&sixty,
892 },
893 {
894 .procname = "watchdog_cpumask",
895 .data = &watchdog_cpumask_bits,
896 .maxlen = NR_CPUS,
897 .mode = 0644,
898 .proc_handler = proc_watchdog_cpumask,
899 },
900#ifdef CONFIG_SOFTLOCKUP_DETECTOR
901 {
902 .procname = "soft_watchdog",
903 .data = &watchdog_softlockup_user_enabled,
904 .maxlen = sizeof(int),
905 .mode = 0644,
906 .proc_handler = proc_soft_watchdog,
907 .extra1 = SYSCTL_ZERO,
908 .extra2 = SYSCTL_ONE,
909 },
910 {
911 .procname = "softlockup_panic",
912 .data = &softlockup_panic,
913 .maxlen = sizeof(int),
914 .mode = 0644,
915 .proc_handler = proc_dointvec_minmax,
916 .extra1 = SYSCTL_ZERO,
917 .extra2 = SYSCTL_ONE,
918 },
919#ifdef CONFIG_SMP
920 {
921 .procname = "softlockup_all_cpu_backtrace",
922 .data = &sysctl_softlockup_all_cpu_backtrace,
923 .maxlen = sizeof(int),
924 .mode = 0644,
925 .proc_handler = proc_dointvec_minmax,
926 .extra1 = SYSCTL_ZERO,
927 .extra2 = SYSCTL_ONE,
928 },
929#endif /* CONFIG_SMP */
930#endif
931#ifdef CONFIG_HARDLOCKUP_DETECTOR
932 {
933 .procname = "hardlockup_panic",
934 .data = &hardlockup_panic,
935 .maxlen = sizeof(int),
936 .mode = 0644,
937 .proc_handler = proc_dointvec_minmax,
938 .extra1 = SYSCTL_ZERO,
939 .extra2 = SYSCTL_ONE,
940 },
941#ifdef CONFIG_SMP
942 {
943 .procname = "hardlockup_all_cpu_backtrace",
944 .data = &sysctl_hardlockup_all_cpu_backtrace,
945 .maxlen = sizeof(int),
946 .mode = 0644,
947 .proc_handler = proc_dointvec_minmax,
948 .extra1 = SYSCTL_ZERO,
949 .extra2 = SYSCTL_ONE,
950 },
951#endif /* CONFIG_SMP */
952#endif
953 {}
954};
955
956static struct ctl_table watchdog_hardlockup_sysctl[] = {
957 {
958 .procname = "nmi_watchdog",
959 .data = &watchdog_hardlockup_user_enabled,
960 .maxlen = sizeof(int),
961 .mode = 0444,
962 .proc_handler = proc_nmi_watchdog,
963 .extra1 = SYSCTL_ZERO,
964 .extra2 = SYSCTL_ONE,
965 },
966 {}
967};
968
969static void __init watchdog_sysctl_init(void)
970{
971 register_sysctl_init("kernel", watchdog_sysctls);
972
973 if (watchdog_hardlockup_available)
974 watchdog_hardlockup_sysctl[0].mode = 0644;
975 register_sysctl_init("kernel", watchdog_hardlockup_sysctl);
976}
977
978#else
979#define watchdog_sysctl_init() do { } while (0)
980#endif /* CONFIG_SYSCTL */
981
982static void __init lockup_detector_delay_init(struct work_struct *work);
983static bool allow_lockup_detector_init_retry __initdata;
984
985static struct work_struct detector_work __initdata =
986 __WORK_INITIALIZER(detector_work, lockup_detector_delay_init);
987
988static void __init lockup_detector_delay_init(struct work_struct *work)
989{
990 int ret;
991
992 ret = watchdog_hardlockup_probe();
993 if (ret) {
994 pr_info("Delayed init of the lockup detector failed: %d\n", ret);
995 pr_info("Hard watchdog permanently disabled\n");
996 return;
997 }
998
999 allow_lockup_detector_init_retry = false;
1000
1001 watchdog_hardlockup_available = true;
1002 lockup_detector_setup();
1003}
1004
1005/*
1006 * lockup_detector_retry_init - retry init lockup detector if possible.
1007 *
1008 * Retry hardlockup detector init. It is useful when it requires some
1009 * functionality that has to be initialized later on a particular
1010 * platform.
1011 */
1012void __init lockup_detector_retry_init(void)
1013{
1014 /* Must be called before late init calls */
1015 if (!allow_lockup_detector_init_retry)
1016 return;
1017
1018 schedule_work(&detector_work);
1019}
1020
1021/*
1022 * Ensure that optional delayed hardlockup init is proceed before
1023 * the init code and memory is freed.
1024 */
1025static int __init lockup_detector_check(void)
1026{
1027 /* Prevent any later retry. */
1028 allow_lockup_detector_init_retry = false;
1029
1030 /* Make sure no work is pending. */
1031 flush_work(&detector_work);
1032
1033 watchdog_sysctl_init();
1034
1035 return 0;
1036
1037}
1038late_initcall_sync(lockup_detector_check);
1039
1040void __init lockup_detector_init(void)
1041{
1042 if (tick_nohz_full_enabled())
1043 pr_info("Disabling watchdog on nohz_full cores by default\n");
1044
1045 cpumask_copy(&watchdog_cpumask,
1046 housekeeping_cpumask(HK_TYPE_TIMER));
1047
1048 if (!watchdog_hardlockup_probe())
1049 watchdog_hardlockup_available = true;
1050 else
1051 allow_lockup_detector_init_retry = true;
1052
1053 lockup_detector_setup();
1054}