Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
7 * Note: Most of this code is borrowed heavily from the original softlockup
8 * detector, so thanks to Ingo for the initial implementation.
9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
10 * to those contributors as well.
11 */
12
13#define pr_fmt(fmt) "watchdog: " fmt
14
15#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/module.h>
20#include <linux/sysctl.h>
21#include <linux/tick.h>
22#include <linux/sched/clock.h>
23#include <linux/sched/debug.h>
24#include <linux/sched/isolation.h>
25#include <linux/stop_machine.h>
26
27#include <asm/irq_regs.h>
28#include <linux/kvm_para.h>
29
30static DEFINE_MUTEX(watchdog_mutex);
31
32#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
33# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
34# define NMI_WATCHDOG_DEFAULT 1
35#else
36# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
37# define NMI_WATCHDOG_DEFAULT 0
38#endif
39
40unsigned long __read_mostly watchdog_enabled;
41int __read_mostly watchdog_user_enabled = 1;
42int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
43int __read_mostly soft_watchdog_user_enabled = 1;
44int __read_mostly watchdog_thresh = 10;
45static int __read_mostly nmi_watchdog_available;
46
47static struct cpumask watchdog_allowed_mask __read_mostly;
48
49struct cpumask watchdog_cpumask __read_mostly;
50unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
51
52#ifdef CONFIG_HARDLOCKUP_DETECTOR
53/*
54 * Should we panic when a soft-lockup or hard-lockup occurs:
55 */
56unsigned int __read_mostly hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
58/*
59 * We may not want to enable hard lockup detection by default in all cases,
60 * for example when running the kernel as a guest on a hypervisor. In these
61 * cases this function can be called to disable hard lockup detection. This
62 * function should only be executed once by the boot processor before the
63 * kernel command line parameters are parsed, because otherwise it is not
64 * possible to override this in hardlockup_panic_setup().
65 */
66void __init hardlockup_detector_disable(void)
67{
68 nmi_watchdog_user_enabled = 0;
69}
70
71static int __init hardlockup_panic_setup(char *str)
72{
73 if (!strncmp(str, "panic", 5))
74 hardlockup_panic = 1;
75 else if (!strncmp(str, "nopanic", 7))
76 hardlockup_panic = 0;
77 else if (!strncmp(str, "0", 1))
78 nmi_watchdog_user_enabled = 0;
79 else if (!strncmp(str, "1", 1))
80 nmi_watchdog_user_enabled = 1;
81 return 1;
82}
83__setup("nmi_watchdog=", hardlockup_panic_setup);
84
85# ifdef CONFIG_SMP
86int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
87
88static int __init hardlockup_all_cpu_backtrace_setup(char *str)
89{
90 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
91 return 1;
92}
93__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
94# endif /* CONFIG_SMP */
95#endif /* CONFIG_HARDLOCKUP_DETECTOR */
96
97/*
98 * These functions can be overridden if an architecture implements its
99 * own hardlockup detector.
100 *
101 * watchdog_nmi_enable/disable can be implemented to start and stop when
102 * softlockup watchdog threads start and stop. The arch must select the
103 * SOFTLOCKUP_DETECTOR Kconfig.
104 */
105int __weak watchdog_nmi_enable(unsigned int cpu)
106{
107 hardlockup_detector_perf_enable();
108 return 0;
109}
110
111void __weak watchdog_nmi_disable(unsigned int cpu)
112{
113 hardlockup_detector_perf_disable();
114}
115
116/* Return 0, if a NMI watchdog is available. Error code otherwise */
117int __weak __init watchdog_nmi_probe(void)
118{
119 return hardlockup_detector_perf_init();
120}
121
122/**
123 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
124 *
125 * The reconfiguration steps are:
126 * watchdog_nmi_stop();
127 * update_variables();
128 * watchdog_nmi_start();
129 */
130void __weak watchdog_nmi_stop(void) { }
131
132/**
133 * watchdog_nmi_start - Start the watchdog after reconfiguration
134 *
135 * Counterpart to watchdog_nmi_stop().
136 *
137 * The following variables have been updated in update_variables() and
138 * contain the currently valid configuration:
139 * - watchdog_enabled
140 * - watchdog_thresh
141 * - watchdog_cpumask
142 */
143void __weak watchdog_nmi_start(void) { }
144
145/**
146 * lockup_detector_update_enable - Update the sysctl enable bit
147 *
148 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
149 * can't race with watchdog_nmi_disable().
150 */
151static void lockup_detector_update_enable(void)
152{
153 watchdog_enabled = 0;
154 if (!watchdog_user_enabled)
155 return;
156 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
157 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
158 if (soft_watchdog_user_enabled)
159 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
160}
161
162#ifdef CONFIG_SOFTLOCKUP_DETECTOR
163
164/* Global variables, exported for sysctl */
165unsigned int __read_mostly softlockup_panic =
166 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
167
168static bool softlockup_initialized __read_mostly;
169static u64 __read_mostly sample_period;
170
171static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
172static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
173static DEFINE_PER_CPU(bool, softlockup_touch_sync);
174static DEFINE_PER_CPU(bool, soft_watchdog_warn);
175static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
176static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
177static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
178static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
179static unsigned long soft_lockup_nmi_warn;
180
181static int __init softlockup_panic_setup(char *str)
182{
183 softlockup_panic = simple_strtoul(str, NULL, 0);
184 return 1;
185}
186__setup("softlockup_panic=", softlockup_panic_setup);
187
188static int __init nowatchdog_setup(char *str)
189{
190 watchdog_user_enabled = 0;
191 return 1;
192}
193__setup("nowatchdog", nowatchdog_setup);
194
195static int __init nosoftlockup_setup(char *str)
196{
197 soft_watchdog_user_enabled = 0;
198 return 1;
199}
200__setup("nosoftlockup", nosoftlockup_setup);
201
202static int __init watchdog_thresh_setup(char *str)
203{
204 get_option(&str, &watchdog_thresh);
205 return 1;
206}
207__setup("watchdog_thresh=", watchdog_thresh_setup);
208
209#ifdef CONFIG_SMP
210int __read_mostly sysctl_softlockup_all_cpu_backtrace;
211
212static int __init softlockup_all_cpu_backtrace_setup(char *str)
213{
214 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
215 return 1;
216}
217__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
218#endif
219
220static void __lockup_detector_cleanup(void);
221
222/*
223 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
224 * lockups can have false positives under extreme conditions. So we generally
225 * want a higher threshold for soft lockups than for hard lockups. So we couple
226 * the thresholds with a factor: we make the soft threshold twice the amount of
227 * time the hard threshold is.
228 */
229static int get_softlockup_thresh(void)
230{
231 return watchdog_thresh * 2;
232}
233
234/*
235 * Returns seconds, approximately. We don't need nanosecond
236 * resolution, and we don't need to waste time with a big divide when
237 * 2^30ns == 1.074s.
238 */
239static unsigned long get_timestamp(void)
240{
241 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
242}
243
244static void set_sample_period(void)
245{
246 /*
247 * convert watchdog_thresh from seconds to ns
248 * the divide by 5 is to give hrtimer several chances (two
249 * or three with the current relation between the soft
250 * and hard thresholds) to increment before the
251 * hardlockup detector generates a warning
252 */
253 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
254 watchdog_update_hrtimer_threshold(sample_period);
255}
256
257/* Commands for resetting the watchdog */
258static void __touch_watchdog(void)
259{
260 __this_cpu_write(watchdog_touch_ts, get_timestamp());
261}
262
263/**
264 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
265 *
266 * Call when the scheduler may have stalled for legitimate reasons
267 * preventing the watchdog task from executing - e.g. the scheduler
268 * entering idle state. This should only be used for scheduler events.
269 * Use touch_softlockup_watchdog() for everything else.
270 */
271notrace void touch_softlockup_watchdog_sched(void)
272{
273 /*
274 * Preemption can be enabled. It doesn't matter which CPU's timestamp
275 * gets zeroed here, so use the raw_ operation.
276 */
277 raw_cpu_write(watchdog_touch_ts, 0);
278}
279
280notrace void touch_softlockup_watchdog(void)
281{
282 touch_softlockup_watchdog_sched();
283 wq_watchdog_touch(raw_smp_processor_id());
284}
285EXPORT_SYMBOL(touch_softlockup_watchdog);
286
287void touch_all_softlockup_watchdogs(void)
288{
289 int cpu;
290
291 /*
292 * watchdog_mutex cannpt be taken here, as this might be called
293 * from (soft)interrupt context, so the access to
294 * watchdog_allowed_cpumask might race with a concurrent update.
295 *
296 * The watchdog time stamp can race against a concurrent real
297 * update as well, the only side effect might be a cycle delay for
298 * the softlockup check.
299 */
300 for_each_cpu(cpu, &watchdog_allowed_mask)
301 per_cpu(watchdog_touch_ts, cpu) = 0;
302 wq_watchdog_touch(-1);
303}
304
305void touch_softlockup_watchdog_sync(void)
306{
307 __this_cpu_write(softlockup_touch_sync, true);
308 __this_cpu_write(watchdog_touch_ts, 0);
309}
310
311static int is_softlockup(unsigned long touch_ts)
312{
313 unsigned long now = get_timestamp();
314
315 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
316 /* Warn about unreasonable delays. */
317 if (time_after(now, touch_ts + get_softlockup_thresh()))
318 return now - touch_ts;
319 }
320 return 0;
321}
322
323/* watchdog detector functions */
324bool is_hardlockup(void)
325{
326 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
327
328 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
329 return true;
330
331 __this_cpu_write(hrtimer_interrupts_saved, hrint);
332 return false;
333}
334
335static void watchdog_interrupt_count(void)
336{
337 __this_cpu_inc(hrtimer_interrupts);
338}
339
340static DEFINE_PER_CPU(struct completion, softlockup_completion);
341static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
342
343/*
344 * The watchdog thread function - touches the timestamp.
345 *
346 * It only runs once every sample_period seconds (4 seconds by
347 * default) to reset the softlockup timestamp. If this gets delayed
348 * for more than 2*watchdog_thresh seconds then the debug-printout
349 * triggers in watchdog_timer_fn().
350 */
351static int softlockup_fn(void *data)
352{
353 __this_cpu_write(soft_lockup_hrtimer_cnt,
354 __this_cpu_read(hrtimer_interrupts));
355 __touch_watchdog();
356 complete(this_cpu_ptr(&softlockup_completion));
357
358 return 0;
359}
360
361/* watchdog kicker functions */
362static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
363{
364 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
365 struct pt_regs *regs = get_irq_regs();
366 int duration;
367 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
368
369 if (!watchdog_enabled)
370 return HRTIMER_NORESTART;
371
372 /* kick the hardlockup detector */
373 watchdog_interrupt_count();
374
375 /* kick the softlockup detector */
376 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
377 reinit_completion(this_cpu_ptr(&softlockup_completion));
378 stop_one_cpu_nowait(smp_processor_id(),
379 softlockup_fn, NULL,
380 this_cpu_ptr(&softlockup_stop_work));
381 }
382
383 /* .. and repeat */
384 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
385
386 if (touch_ts == 0) {
387 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
388 /*
389 * If the time stamp was touched atomically
390 * make sure the scheduler tick is up to date.
391 */
392 __this_cpu_write(softlockup_touch_sync, false);
393 sched_clock_tick();
394 }
395
396 /* Clear the guest paused flag on watchdog reset */
397 kvm_check_and_clear_guest_paused();
398 __touch_watchdog();
399 return HRTIMER_RESTART;
400 }
401
402 /* check for a softlockup
403 * This is done by making sure a high priority task is
404 * being scheduled. The task touches the watchdog to
405 * indicate it is getting cpu time. If it hasn't then
406 * this is a good indication some task is hogging the cpu
407 */
408 duration = is_softlockup(touch_ts);
409 if (unlikely(duration)) {
410 /*
411 * If a virtual machine is stopped by the host it can look to
412 * the watchdog like a soft lockup, check to see if the host
413 * stopped the vm before we issue the warning
414 */
415 if (kvm_check_and_clear_guest_paused())
416 return HRTIMER_RESTART;
417
418 /* only warn once */
419 if (__this_cpu_read(soft_watchdog_warn) == true) {
420 /*
421 * When multiple processes are causing softlockups the
422 * softlockup detector only warns on the first one
423 * because the code relies on a full quiet cycle to
424 * re-arm. The second process prevents the quiet cycle
425 * and never gets reported. Use task pointers to detect
426 * this.
427 */
428 if (__this_cpu_read(softlockup_task_ptr_saved) !=
429 current) {
430 __this_cpu_write(soft_watchdog_warn, false);
431 __touch_watchdog();
432 }
433 return HRTIMER_RESTART;
434 }
435
436 if (softlockup_all_cpu_backtrace) {
437 /* Prevent multiple soft-lockup reports if one cpu is already
438 * engaged in dumping cpu back traces
439 */
440 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
441 /* Someone else will report us. Let's give up */
442 __this_cpu_write(soft_watchdog_warn, true);
443 return HRTIMER_RESTART;
444 }
445 }
446
447 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
448 smp_processor_id(), duration,
449 current->comm, task_pid_nr(current));
450 __this_cpu_write(softlockup_task_ptr_saved, current);
451 print_modules();
452 print_irqtrace_events(current);
453 if (regs)
454 show_regs(regs);
455 else
456 dump_stack();
457
458 if (softlockup_all_cpu_backtrace) {
459 /* Avoid generating two back traces for current
460 * given that one is already made above
461 */
462 trigger_allbutself_cpu_backtrace();
463
464 clear_bit(0, &soft_lockup_nmi_warn);
465 /* Barrier to sync with other cpus */
466 smp_mb__after_atomic();
467 }
468
469 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
470 if (softlockup_panic)
471 panic("softlockup: hung tasks");
472 __this_cpu_write(soft_watchdog_warn, true);
473 } else
474 __this_cpu_write(soft_watchdog_warn, false);
475
476 return HRTIMER_RESTART;
477}
478
479static void watchdog_enable(unsigned int cpu)
480{
481 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
482 struct completion *done = this_cpu_ptr(&softlockup_completion);
483
484 WARN_ON_ONCE(cpu != smp_processor_id());
485
486 init_completion(done);
487 complete(done);
488
489 /*
490 * Start the timer first to prevent the NMI watchdog triggering
491 * before the timer has a chance to fire.
492 */
493 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
494 hrtimer->function = watchdog_timer_fn;
495 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
496 HRTIMER_MODE_REL_PINNED_HARD);
497
498 /* Initialize timestamp */
499 __touch_watchdog();
500 /* Enable the perf event */
501 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
502 watchdog_nmi_enable(cpu);
503}
504
505static void watchdog_disable(unsigned int cpu)
506{
507 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
508
509 WARN_ON_ONCE(cpu != smp_processor_id());
510
511 /*
512 * Disable the perf event first. That prevents that a large delay
513 * between disabling the timer and disabling the perf event causes
514 * the perf NMI to detect a false positive.
515 */
516 watchdog_nmi_disable(cpu);
517 hrtimer_cancel(hrtimer);
518 wait_for_completion(this_cpu_ptr(&softlockup_completion));
519}
520
521static int softlockup_stop_fn(void *data)
522{
523 watchdog_disable(smp_processor_id());
524 return 0;
525}
526
527static void softlockup_stop_all(void)
528{
529 int cpu;
530
531 if (!softlockup_initialized)
532 return;
533
534 for_each_cpu(cpu, &watchdog_allowed_mask)
535 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
536
537 cpumask_clear(&watchdog_allowed_mask);
538}
539
540static int softlockup_start_fn(void *data)
541{
542 watchdog_enable(smp_processor_id());
543 return 0;
544}
545
546static void softlockup_start_all(void)
547{
548 int cpu;
549
550 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
551 for_each_cpu(cpu, &watchdog_allowed_mask)
552 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
553}
554
555int lockup_detector_online_cpu(unsigned int cpu)
556{
557 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
558 watchdog_enable(cpu);
559 return 0;
560}
561
562int lockup_detector_offline_cpu(unsigned int cpu)
563{
564 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
565 watchdog_disable(cpu);
566 return 0;
567}
568
569static void lockup_detector_reconfigure(void)
570{
571 cpus_read_lock();
572 watchdog_nmi_stop();
573
574 softlockup_stop_all();
575 set_sample_period();
576 lockup_detector_update_enable();
577 if (watchdog_enabled && watchdog_thresh)
578 softlockup_start_all();
579
580 watchdog_nmi_start();
581 cpus_read_unlock();
582 /*
583 * Must be called outside the cpus locked section to prevent
584 * recursive locking in the perf code.
585 */
586 __lockup_detector_cleanup();
587}
588
589/*
590 * Create the watchdog thread infrastructure and configure the detector(s).
591 *
592 * The threads are not unparked as watchdog_allowed_mask is empty. When
593 * the threads are successfully initialized, take the proper locks and
594 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
595 */
596static __init void lockup_detector_setup(void)
597{
598 /*
599 * If sysctl is off and watchdog got disabled on the command line,
600 * nothing to do here.
601 */
602 lockup_detector_update_enable();
603
604 if (!IS_ENABLED(CONFIG_SYSCTL) &&
605 !(watchdog_enabled && watchdog_thresh))
606 return;
607
608 mutex_lock(&watchdog_mutex);
609 lockup_detector_reconfigure();
610 softlockup_initialized = true;
611 mutex_unlock(&watchdog_mutex);
612}
613
614#else /* CONFIG_SOFTLOCKUP_DETECTOR */
615static void lockup_detector_reconfigure(void)
616{
617 cpus_read_lock();
618 watchdog_nmi_stop();
619 lockup_detector_update_enable();
620 watchdog_nmi_start();
621 cpus_read_unlock();
622}
623static inline void lockup_detector_setup(void)
624{
625 lockup_detector_reconfigure();
626}
627#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
628
629static void __lockup_detector_cleanup(void)
630{
631 lockdep_assert_held(&watchdog_mutex);
632 hardlockup_detector_perf_cleanup();
633}
634
635/**
636 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
637 *
638 * Caller must not hold the cpu hotplug rwsem.
639 */
640void lockup_detector_cleanup(void)
641{
642 mutex_lock(&watchdog_mutex);
643 __lockup_detector_cleanup();
644 mutex_unlock(&watchdog_mutex);
645}
646
647/**
648 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
649 *
650 * Special interface for parisc. It prevents lockup detector warnings from
651 * the default pm_poweroff() function which busy loops forever.
652 */
653void lockup_detector_soft_poweroff(void)
654{
655 watchdog_enabled = 0;
656}
657
658#ifdef CONFIG_SYSCTL
659
660/* Propagate any changes to the watchdog threads */
661static void proc_watchdog_update(void)
662{
663 /* Remove impossible cpus to keep sysctl output clean. */
664 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
665 lockup_detector_reconfigure();
666}
667
668/*
669 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
670 *
671 * caller | table->data points to | 'which'
672 * -------------------|----------------------------|--------------------------
673 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
674 * | | SOFT_WATCHDOG_ENABLED
675 * -------------------|----------------------------|--------------------------
676 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
677 * -------------------|----------------------------|--------------------------
678 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
679 */
680static int proc_watchdog_common(int which, struct ctl_table *table, int write,
681 void __user *buffer, size_t *lenp, loff_t *ppos)
682{
683 int err, old, *param = table->data;
684
685 mutex_lock(&watchdog_mutex);
686
687 if (!write) {
688 /*
689 * On read synchronize the userspace interface. This is a
690 * racy snapshot.
691 */
692 *param = (watchdog_enabled & which) != 0;
693 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
694 } else {
695 old = READ_ONCE(*param);
696 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
697 if (!err && old != READ_ONCE(*param))
698 proc_watchdog_update();
699 }
700 mutex_unlock(&watchdog_mutex);
701 return err;
702}
703
704/*
705 * /proc/sys/kernel/watchdog
706 */
707int proc_watchdog(struct ctl_table *table, int write,
708 void __user *buffer, size_t *lenp, loff_t *ppos)
709{
710 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
711 table, write, buffer, lenp, ppos);
712}
713
714/*
715 * /proc/sys/kernel/nmi_watchdog
716 */
717int proc_nmi_watchdog(struct ctl_table *table, int write,
718 void __user *buffer, size_t *lenp, loff_t *ppos)
719{
720 if (!nmi_watchdog_available && write)
721 return -ENOTSUPP;
722 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
723 table, write, buffer, lenp, ppos);
724}
725
726/*
727 * /proc/sys/kernel/soft_watchdog
728 */
729int proc_soft_watchdog(struct ctl_table *table, int write,
730 void __user *buffer, size_t *lenp, loff_t *ppos)
731{
732 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
733 table, write, buffer, lenp, ppos);
734}
735
736/*
737 * /proc/sys/kernel/watchdog_thresh
738 */
739int proc_watchdog_thresh(struct ctl_table *table, int write,
740 void __user *buffer, size_t *lenp, loff_t *ppos)
741{
742 int err, old;
743
744 mutex_lock(&watchdog_mutex);
745
746 old = READ_ONCE(watchdog_thresh);
747 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
748
749 if (!err && write && old != READ_ONCE(watchdog_thresh))
750 proc_watchdog_update();
751
752 mutex_unlock(&watchdog_mutex);
753 return err;
754}
755
756/*
757 * The cpumask is the mask of possible cpus that the watchdog can run
758 * on, not the mask of cpus it is actually running on. This allows the
759 * user to specify a mask that will include cpus that have not yet
760 * been brought online, if desired.
761 */
762int proc_watchdog_cpumask(struct ctl_table *table, int write,
763 void __user *buffer, size_t *lenp, loff_t *ppos)
764{
765 int err;
766
767 mutex_lock(&watchdog_mutex);
768
769 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
770 if (!err && write)
771 proc_watchdog_update();
772
773 mutex_unlock(&watchdog_mutex);
774 return err;
775}
776#endif /* CONFIG_SYSCTL */
777
778void __init lockup_detector_init(void)
779{
780 if (tick_nohz_full_enabled())
781 pr_info("Disabling watchdog on nohz_full cores by default\n");
782
783 cpumask_copy(&watchdog_cpumask,
784 housekeeping_cpumask(HK_FLAG_TIMER));
785
786 if (!watchdog_nmi_probe())
787 nmi_watchdog_available = true;
788 lockup_detector_setup();
789}
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/delay.h>
19#include <linux/freezer.h>
20#include <linux/kthread.h>
21#include <linux/lockdep.h>
22#include <linux/notifier.h>
23#include <linux/module.h>
24#include <linux/sysctl.h>
25
26#include <asm/irq_regs.h>
27#include <linux/kvm_para.h>
28#include <linux/perf_event.h>
29
30int watchdog_enabled = 1;
31int __read_mostly watchdog_thresh = 10;
32
33static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
34static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
35static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
36static DEFINE_PER_CPU(bool, softlockup_touch_sync);
37static DEFINE_PER_CPU(bool, soft_watchdog_warn);
38#ifdef CONFIG_HARDLOCKUP_DETECTOR
39static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
41static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
42static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
43static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
44#endif
45
46/* boot commands */
47/*
48 * Should we panic when a soft-lockup or hard-lockup occurs:
49 */
50#ifdef CONFIG_HARDLOCKUP_DETECTOR
51static int hardlockup_panic =
52 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
53
54static int __init hardlockup_panic_setup(char *str)
55{
56 if (!strncmp(str, "panic", 5))
57 hardlockup_panic = 1;
58 else if (!strncmp(str, "nopanic", 7))
59 hardlockup_panic = 0;
60 else if (!strncmp(str, "0", 1))
61 watchdog_enabled = 0;
62 return 1;
63}
64__setup("nmi_watchdog=", hardlockup_panic_setup);
65#endif
66
67unsigned int __read_mostly softlockup_panic =
68 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
69
70static int __init softlockup_panic_setup(char *str)
71{
72 softlockup_panic = simple_strtoul(str, NULL, 0);
73
74 return 1;
75}
76__setup("softlockup_panic=", softlockup_panic_setup);
77
78static int __init nowatchdog_setup(char *str)
79{
80 watchdog_enabled = 0;
81 return 1;
82}
83__setup("nowatchdog", nowatchdog_setup);
84
85/* deprecated */
86static int __init nosoftlockup_setup(char *str)
87{
88 watchdog_enabled = 0;
89 return 1;
90}
91__setup("nosoftlockup", nosoftlockup_setup);
92/* */
93
94/*
95 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
96 * lockups can have false positives under extreme conditions. So we generally
97 * want a higher threshold for soft lockups than for hard lockups. So we couple
98 * the thresholds with a factor: we make the soft threshold twice the amount of
99 * time the hard threshold is.
100 */
101static int get_softlockup_thresh(void)
102{
103 return watchdog_thresh * 2;
104}
105
106/*
107 * Returns seconds, approximately. We don't need nanosecond
108 * resolution, and we don't need to waste time with a big divide when
109 * 2^30ns == 1.074s.
110 */
111static unsigned long get_timestamp(int this_cpu)
112{
113 return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */
114}
115
116static unsigned long get_sample_period(void)
117{
118 /*
119 * convert watchdog_thresh from seconds to ns
120 * the divide by 5 is to give hrtimer several chances (two
121 * or three with the current relation between the soft
122 * and hard thresholds) to increment before the
123 * hardlockup detector generates a warning
124 */
125 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
126}
127
128/* Commands for resetting the watchdog */
129static void __touch_watchdog(void)
130{
131 int this_cpu = smp_processor_id();
132
133 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
134}
135
136void touch_softlockup_watchdog(void)
137{
138 __this_cpu_write(watchdog_touch_ts, 0);
139}
140EXPORT_SYMBOL(touch_softlockup_watchdog);
141
142void touch_all_softlockup_watchdogs(void)
143{
144 int cpu;
145
146 /*
147 * this is done lockless
148 * do we care if a 0 races with a timestamp?
149 * all it means is the softlock check starts one cycle later
150 */
151 for_each_online_cpu(cpu)
152 per_cpu(watchdog_touch_ts, cpu) = 0;
153}
154
155#ifdef CONFIG_HARDLOCKUP_DETECTOR
156void touch_nmi_watchdog(void)
157{
158 if (watchdog_enabled) {
159 unsigned cpu;
160
161 for_each_present_cpu(cpu) {
162 if (per_cpu(watchdog_nmi_touch, cpu) != true)
163 per_cpu(watchdog_nmi_touch, cpu) = true;
164 }
165 }
166 touch_softlockup_watchdog();
167}
168EXPORT_SYMBOL(touch_nmi_watchdog);
169
170#endif
171
172void touch_softlockup_watchdog_sync(void)
173{
174 __raw_get_cpu_var(softlockup_touch_sync) = true;
175 __raw_get_cpu_var(watchdog_touch_ts) = 0;
176}
177
178#ifdef CONFIG_HARDLOCKUP_DETECTOR
179/* watchdog detector functions */
180static int is_hardlockup(void)
181{
182 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
183
184 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
185 return 1;
186
187 __this_cpu_write(hrtimer_interrupts_saved, hrint);
188 return 0;
189}
190#endif
191
192static int is_softlockup(unsigned long touch_ts)
193{
194 unsigned long now = get_timestamp(smp_processor_id());
195
196 /* Warn about unreasonable delays: */
197 if (time_after(now, touch_ts + get_softlockup_thresh()))
198 return now - touch_ts;
199
200 return 0;
201}
202
203#ifdef CONFIG_HARDLOCKUP_DETECTOR
204
205static struct perf_event_attr wd_hw_attr = {
206 .type = PERF_TYPE_HARDWARE,
207 .config = PERF_COUNT_HW_CPU_CYCLES,
208 .size = sizeof(struct perf_event_attr),
209 .pinned = 1,
210 .disabled = 1,
211};
212
213/* Callback function for perf event subsystem */
214static void watchdog_overflow_callback(struct perf_event *event,
215 struct perf_sample_data *data,
216 struct pt_regs *regs)
217{
218 /* Ensure the watchdog never gets throttled */
219 event->hw.interrupts = 0;
220
221 if (__this_cpu_read(watchdog_nmi_touch) == true) {
222 __this_cpu_write(watchdog_nmi_touch, false);
223 return;
224 }
225
226 /* check for a hardlockup
227 * This is done by making sure our timer interrupt
228 * is incrementing. The timer interrupt should have
229 * fired multiple times before we overflow'd. If it hasn't
230 * then this is a good indication the cpu is stuck
231 */
232 if (is_hardlockup()) {
233 int this_cpu = smp_processor_id();
234
235 /* only print hardlockups once */
236 if (__this_cpu_read(hard_watchdog_warn) == true)
237 return;
238
239 if (hardlockup_panic)
240 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
241 else
242 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
243
244 __this_cpu_write(hard_watchdog_warn, true);
245 return;
246 }
247
248 __this_cpu_write(hard_watchdog_warn, false);
249 return;
250}
251static void watchdog_interrupt_count(void)
252{
253 __this_cpu_inc(hrtimer_interrupts);
254}
255#else
256static inline void watchdog_interrupt_count(void) { return; }
257#endif /* CONFIG_HARDLOCKUP_DETECTOR */
258
259/* watchdog kicker functions */
260static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
261{
262 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
263 struct pt_regs *regs = get_irq_regs();
264 int duration;
265
266 /* kick the hardlockup detector */
267 watchdog_interrupt_count();
268
269 /* kick the softlockup detector */
270 wake_up_process(__this_cpu_read(softlockup_watchdog));
271
272 /* .. and repeat */
273 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
274
275 if (touch_ts == 0) {
276 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
277 /*
278 * If the time stamp was touched atomically
279 * make sure the scheduler tick is up to date.
280 */
281 __this_cpu_write(softlockup_touch_sync, false);
282 sched_clock_tick();
283 }
284
285 /* Clear the guest paused flag on watchdog reset */
286 kvm_check_and_clear_guest_paused();
287 __touch_watchdog();
288 return HRTIMER_RESTART;
289 }
290
291 /* check for a softlockup
292 * This is done by making sure a high priority task is
293 * being scheduled. The task touches the watchdog to
294 * indicate it is getting cpu time. If it hasn't then
295 * this is a good indication some task is hogging the cpu
296 */
297 duration = is_softlockup(touch_ts);
298 if (unlikely(duration)) {
299 /*
300 * If a virtual machine is stopped by the host it can look to
301 * the watchdog like a soft lockup, check to see if the host
302 * stopped the vm before we issue the warning
303 */
304 if (kvm_check_and_clear_guest_paused())
305 return HRTIMER_RESTART;
306
307 /* only warn once */
308 if (__this_cpu_read(soft_watchdog_warn) == true)
309 return HRTIMER_RESTART;
310
311 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
312 smp_processor_id(), duration,
313 current->comm, task_pid_nr(current));
314 print_modules();
315 print_irqtrace_events(current);
316 if (regs)
317 show_regs(regs);
318 else
319 dump_stack();
320
321 if (softlockup_panic)
322 panic("softlockup: hung tasks");
323 __this_cpu_write(soft_watchdog_warn, true);
324 } else
325 __this_cpu_write(soft_watchdog_warn, false);
326
327 return HRTIMER_RESTART;
328}
329
330
331/*
332 * The watchdog thread - touches the timestamp.
333 */
334static int watchdog(void *unused)
335{
336 struct sched_param param = { .sched_priority = 0 };
337 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
338
339 /* initialize timestamp */
340 __touch_watchdog();
341
342 /* kick off the timer for the hardlockup detector */
343 /* done here because hrtimer_start can only pin to smp_processor_id() */
344 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
345 HRTIMER_MODE_REL_PINNED);
346
347 set_current_state(TASK_INTERRUPTIBLE);
348 /*
349 * Run briefly (kicked by the hrtimer callback function) once every
350 * get_sample_period() seconds (4 seconds by default) to reset the
351 * softlockup timestamp. If this gets delayed for more than
352 * 2*watchdog_thresh seconds then the debug-printout triggers in
353 * watchdog_timer_fn().
354 */
355 while (!kthread_should_stop()) {
356 __touch_watchdog();
357 schedule();
358
359 if (kthread_should_stop())
360 break;
361
362 set_current_state(TASK_INTERRUPTIBLE);
363 }
364 /*
365 * Drop the policy/priority elevation during thread exit to avoid a
366 * scheduling latency spike.
367 */
368 __set_current_state(TASK_RUNNING);
369 sched_setscheduler(current, SCHED_NORMAL, ¶m);
370 return 0;
371}
372
373
374#ifdef CONFIG_HARDLOCKUP_DETECTOR
375/*
376 * People like the simple clean cpu node info on boot.
377 * Reduce the watchdog noise by only printing messages
378 * that are different from what cpu0 displayed.
379 */
380static unsigned long cpu0_err;
381
382static int watchdog_nmi_enable(int cpu)
383{
384 struct perf_event_attr *wd_attr;
385 struct perf_event *event = per_cpu(watchdog_ev, cpu);
386
387 /* is it already setup and enabled? */
388 if (event && event->state > PERF_EVENT_STATE_OFF)
389 goto out;
390
391 /* it is setup but not enabled */
392 if (event != NULL)
393 goto out_enable;
394
395 wd_attr = &wd_hw_attr;
396 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
397
398 /* Try to register using hardware perf events */
399 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
400
401 /* save cpu0 error for future comparision */
402 if (cpu == 0 && IS_ERR(event))
403 cpu0_err = PTR_ERR(event);
404
405 if (!IS_ERR(event)) {
406 /* only print for cpu0 or different than cpu0 */
407 if (cpu == 0 || cpu0_err)
408 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
409 goto out_save;
410 }
411
412 /* skip displaying the same error again */
413 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
414 return PTR_ERR(event);
415
416 /* vary the KERN level based on the returned errno */
417 if (PTR_ERR(event) == -EOPNOTSUPP)
418 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
419 else if (PTR_ERR(event) == -ENOENT)
420 pr_warning("disabled (cpu%i): hardware events not enabled\n",
421 cpu);
422 else
423 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
424 cpu, PTR_ERR(event));
425 return PTR_ERR(event);
426
427 /* success path */
428out_save:
429 per_cpu(watchdog_ev, cpu) = event;
430out_enable:
431 perf_event_enable(per_cpu(watchdog_ev, cpu));
432out:
433 return 0;
434}
435
436static void watchdog_nmi_disable(int cpu)
437{
438 struct perf_event *event = per_cpu(watchdog_ev, cpu);
439
440 if (event) {
441 perf_event_disable(event);
442 per_cpu(watchdog_ev, cpu) = NULL;
443
444 /* should be in cleanup, but blocks oprofile */
445 perf_event_release_kernel(event);
446 }
447 return;
448}
449#else
450static int watchdog_nmi_enable(int cpu) { return 0; }
451static void watchdog_nmi_disable(int cpu) { return; }
452#endif /* CONFIG_HARDLOCKUP_DETECTOR */
453
454/* prepare/enable/disable routines */
455static void watchdog_prepare_cpu(int cpu)
456{
457 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
458
459 WARN_ON(per_cpu(softlockup_watchdog, cpu));
460 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
461 hrtimer->function = watchdog_timer_fn;
462}
463
464static int watchdog_enable(int cpu)
465{
466 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
467 int err = 0;
468
469 /* enable the perf event */
470 err = watchdog_nmi_enable(cpu);
471
472 /* Regardless of err above, fall through and start softlockup */
473
474 /* create the watchdog thread */
475 if (!p) {
476 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
477 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
478 if (IS_ERR(p)) {
479 pr_err("softlockup watchdog for %i failed\n", cpu);
480 if (!err) {
481 /* if hardlockup hasn't already set this */
482 err = PTR_ERR(p);
483 /* and disable the perf event */
484 watchdog_nmi_disable(cpu);
485 }
486 goto out;
487 }
488 sched_setscheduler(p, SCHED_FIFO, ¶m);
489 kthread_bind(p, cpu);
490 per_cpu(watchdog_touch_ts, cpu) = 0;
491 per_cpu(softlockup_watchdog, cpu) = p;
492 wake_up_process(p);
493 }
494
495out:
496 return err;
497}
498
499static void watchdog_disable(int cpu)
500{
501 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
502 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
503
504 /*
505 * cancel the timer first to stop incrementing the stats
506 * and waking up the kthread
507 */
508 hrtimer_cancel(hrtimer);
509
510 /* disable the perf event */
511 watchdog_nmi_disable(cpu);
512
513 /* stop the watchdog thread */
514 if (p) {
515 per_cpu(softlockup_watchdog, cpu) = NULL;
516 kthread_stop(p);
517 }
518}
519
520/* sysctl functions */
521#ifdef CONFIG_SYSCTL
522static void watchdog_enable_all_cpus(void)
523{
524 int cpu;
525
526 watchdog_enabled = 0;
527
528 for_each_online_cpu(cpu)
529 if (!watchdog_enable(cpu))
530 /* if any cpu succeeds, watchdog is considered
531 enabled for the system */
532 watchdog_enabled = 1;
533
534 if (!watchdog_enabled)
535 pr_err("failed to be enabled on some cpus\n");
536
537}
538
539static void watchdog_disable_all_cpus(void)
540{
541 int cpu;
542
543 for_each_online_cpu(cpu)
544 watchdog_disable(cpu);
545
546 /* if all watchdogs are disabled, then they are disabled for the system */
547 watchdog_enabled = 0;
548}
549
550
551/*
552 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
553 */
554
555int proc_dowatchdog(struct ctl_table *table, int write,
556 void __user *buffer, size_t *lenp, loff_t *ppos)
557{
558 int ret;
559
560 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
561 if (ret || !write)
562 goto out;
563
564 if (watchdog_enabled && watchdog_thresh)
565 watchdog_enable_all_cpus();
566 else
567 watchdog_disable_all_cpus();
568
569out:
570 return ret;
571}
572#endif /* CONFIG_SYSCTL */
573
574
575/*
576 * Create/destroy watchdog threads as CPUs come and go:
577 */
578static int __cpuinit
579cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
580{
581 int hotcpu = (unsigned long)hcpu;
582
583 switch (action) {
584 case CPU_UP_PREPARE:
585 case CPU_UP_PREPARE_FROZEN:
586 watchdog_prepare_cpu(hotcpu);
587 break;
588 case CPU_ONLINE:
589 case CPU_ONLINE_FROZEN:
590 if (watchdog_enabled)
591 watchdog_enable(hotcpu);
592 break;
593#ifdef CONFIG_HOTPLUG_CPU
594 case CPU_UP_CANCELED:
595 case CPU_UP_CANCELED_FROZEN:
596 watchdog_disable(hotcpu);
597 break;
598 case CPU_DEAD:
599 case CPU_DEAD_FROZEN:
600 watchdog_disable(hotcpu);
601 break;
602#endif /* CONFIG_HOTPLUG_CPU */
603 }
604
605 /*
606 * hardlockup and softlockup are not important enough
607 * to block cpu bring up. Just always succeed and
608 * rely on printk output to flag problems.
609 */
610 return NOTIFY_OK;
611}
612
613static struct notifier_block __cpuinitdata cpu_nfb = {
614 .notifier_call = cpu_callback
615};
616
617void __init lockup_detector_init(void)
618{
619 void *cpu = (void *)(long)smp_processor_id();
620 int err;
621
622 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
623 WARN_ON(notifier_to_errno(err));
624
625 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
626 register_cpu_notifier(&cpu_nfb);
627
628 return;
629}