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/*
2 * Detect hard and soft lockups on a system
3 *
4 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
5 *
6 * this code detects hard lockups: incidents in where on a CPU
7 * the kernel does not respond to anything except NMI.
8 *
9 * Note: Most of this code is borrowed heavily from softlockup.c,
10 * so thanks to Ingo for the initial implementation.
11 * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
12 * to those contributors as well.
13 */
14
15#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/delay.h>
20#include <linux/freezer.h>
21#include <linux/kthread.h>
22#include <linux/lockdep.h>
23#include <linux/notifier.h>
24#include <linux/module.h>
25#include <linux/sysctl.h>
26
27#include <asm/irq_regs.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 5 chances to
121 * increment before the hardlockup detector generates
122 * a warning
123 */
124 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
125}
126
127/* Commands for resetting the watchdog */
128static void __touch_watchdog(void)
129{
130 int this_cpu = smp_processor_id();
131
132 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
133}
134
135void touch_softlockup_watchdog(void)
136{
137 __this_cpu_write(watchdog_touch_ts, 0);
138}
139EXPORT_SYMBOL(touch_softlockup_watchdog);
140
141void touch_all_softlockup_watchdogs(void)
142{
143 int cpu;
144
145 /*
146 * this is done lockless
147 * do we care if a 0 races with a timestamp?
148 * all it means is the softlock check starts one cycle later
149 */
150 for_each_online_cpu(cpu)
151 per_cpu(watchdog_touch_ts, cpu) = 0;
152}
153
154#ifdef CONFIG_HARDLOCKUP_DETECTOR
155void touch_nmi_watchdog(void)
156{
157 if (watchdog_enabled) {
158 unsigned cpu;
159
160 for_each_present_cpu(cpu) {
161 if (per_cpu(watchdog_nmi_touch, cpu) != true)
162 per_cpu(watchdog_nmi_touch, cpu) = true;
163 }
164 }
165 touch_softlockup_watchdog();
166}
167EXPORT_SYMBOL(touch_nmi_watchdog);
168
169#endif
170
171void touch_softlockup_watchdog_sync(void)
172{
173 __raw_get_cpu_var(softlockup_touch_sync) = true;
174 __raw_get_cpu_var(watchdog_touch_ts) = 0;
175}
176
177#ifdef CONFIG_HARDLOCKUP_DETECTOR
178/* watchdog detector functions */
179static int is_hardlockup(void)
180{
181 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
182
183 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
184 return 1;
185
186 __this_cpu_write(hrtimer_interrupts_saved, hrint);
187 return 0;
188}
189#endif
190
191static int is_softlockup(unsigned long touch_ts)
192{
193 unsigned long now = get_timestamp(smp_processor_id());
194
195 /* Warn about unreasonable delays: */
196 if (time_after(now, touch_ts + get_softlockup_thresh()))
197 return now - touch_ts;
198
199 return 0;
200}
201
202#ifdef CONFIG_HARDLOCKUP_DETECTOR
203
204static struct perf_event_attr wd_hw_attr = {
205 .type = PERF_TYPE_HARDWARE,
206 .config = PERF_COUNT_HW_CPU_CYCLES,
207 .size = sizeof(struct perf_event_attr),
208 .pinned = 1,
209 .disabled = 1,
210};
211
212/* Callback function for perf event subsystem */
213static void watchdog_overflow_callback(struct perf_event *event,
214 struct perf_sample_data *data,
215 struct pt_regs *regs)
216{
217 /* Ensure the watchdog never gets throttled */
218 event->hw.interrupts = 0;
219
220 if (__this_cpu_read(watchdog_nmi_touch) == true) {
221 __this_cpu_write(watchdog_nmi_touch, false);
222 return;
223 }
224
225 /* check for a hardlockup
226 * This is done by making sure our timer interrupt
227 * is incrementing. The timer interrupt should have
228 * fired multiple times before we overflow'd. If it hasn't
229 * then this is a good indication the cpu is stuck
230 */
231 if (is_hardlockup()) {
232 int this_cpu = smp_processor_id();
233
234 /* only print hardlockups once */
235 if (__this_cpu_read(hard_watchdog_warn) == true)
236 return;
237
238 if (hardlockup_panic)
239 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
240 else
241 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
242
243 __this_cpu_write(hard_watchdog_warn, true);
244 return;
245 }
246
247 __this_cpu_write(hard_watchdog_warn, false);
248 return;
249}
250static void watchdog_interrupt_count(void)
251{
252 __this_cpu_inc(hrtimer_interrupts);
253}
254#else
255static inline void watchdog_interrupt_count(void) { return; }
256#endif /* CONFIG_HARDLOCKUP_DETECTOR */
257
258/* watchdog kicker functions */
259static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
260{
261 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
262 struct pt_regs *regs = get_irq_regs();
263 int duration;
264
265 /* kick the hardlockup detector */
266 watchdog_interrupt_count();
267
268 /* kick the softlockup detector */
269 wake_up_process(__this_cpu_read(softlockup_watchdog));
270
271 /* .. and repeat */
272 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
273
274 if (touch_ts == 0) {
275 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
276 /*
277 * If the time stamp was touched atomically
278 * make sure the scheduler tick is up to date.
279 */
280 __this_cpu_write(softlockup_touch_sync, false);
281 sched_clock_tick();
282 }
283 __touch_watchdog();
284 return HRTIMER_RESTART;
285 }
286
287 /* check for a softlockup
288 * This is done by making sure a high priority task is
289 * being scheduled. The task touches the watchdog to
290 * indicate it is getting cpu time. If it hasn't then
291 * this is a good indication some task is hogging the cpu
292 */
293 duration = is_softlockup(touch_ts);
294 if (unlikely(duration)) {
295 /* only warn once */
296 if (__this_cpu_read(soft_watchdog_warn) == true)
297 return HRTIMER_RESTART;
298
299 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
300 smp_processor_id(), duration,
301 current->comm, task_pid_nr(current));
302 print_modules();
303 print_irqtrace_events(current);
304 if (regs)
305 show_regs(regs);
306 else
307 dump_stack();
308
309 if (softlockup_panic)
310 panic("softlockup: hung tasks");
311 __this_cpu_write(soft_watchdog_warn, true);
312 } else
313 __this_cpu_write(soft_watchdog_warn, false);
314
315 return HRTIMER_RESTART;
316}
317
318
319/*
320 * The watchdog thread - touches the timestamp.
321 */
322static int watchdog(void *unused)
323{
324 static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
325 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
326
327 sched_setscheduler(current, SCHED_FIFO, ¶m);
328
329 /* initialize timestamp */
330 __touch_watchdog();
331
332 /* kick off the timer for the hardlockup detector */
333 /* done here because hrtimer_start can only pin to smp_processor_id() */
334 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
335 HRTIMER_MODE_REL_PINNED);
336
337 set_current_state(TASK_INTERRUPTIBLE);
338 /*
339 * Run briefly once per second to reset the softlockup timestamp.
340 * If this gets delayed for more than 60 seconds then the
341 * debug-printout triggers in watchdog_timer_fn().
342 */
343 while (!kthread_should_stop()) {
344 __touch_watchdog();
345 schedule();
346
347 if (kthread_should_stop())
348 break;
349
350 set_current_state(TASK_INTERRUPTIBLE);
351 }
352 __set_current_state(TASK_RUNNING);
353
354 return 0;
355}
356
357
358#ifdef CONFIG_HARDLOCKUP_DETECTOR
359static int watchdog_nmi_enable(int cpu)
360{
361 struct perf_event_attr *wd_attr;
362 struct perf_event *event = per_cpu(watchdog_ev, cpu);
363
364 /* is it already setup and enabled? */
365 if (event && event->state > PERF_EVENT_STATE_OFF)
366 goto out;
367
368 /* it is setup but not enabled */
369 if (event != NULL)
370 goto out_enable;
371
372 wd_attr = &wd_hw_attr;
373 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
374
375 /* Try to register using hardware perf events */
376 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
377 if (!IS_ERR(event)) {
378 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
379 goto out_save;
380 }
381
382
383 /* vary the KERN level based on the returned errno */
384 if (PTR_ERR(event) == -EOPNOTSUPP)
385 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
386 else if (PTR_ERR(event) == -ENOENT)
387 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
388 else
389 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
390 return PTR_ERR(event);
391
392 /* success path */
393out_save:
394 per_cpu(watchdog_ev, cpu) = event;
395out_enable:
396 perf_event_enable(per_cpu(watchdog_ev, cpu));
397out:
398 return 0;
399}
400
401static void watchdog_nmi_disable(int cpu)
402{
403 struct perf_event *event = per_cpu(watchdog_ev, cpu);
404
405 if (event) {
406 perf_event_disable(event);
407 per_cpu(watchdog_ev, cpu) = NULL;
408
409 /* should be in cleanup, but blocks oprofile */
410 perf_event_release_kernel(event);
411 }
412 return;
413}
414#else
415static int watchdog_nmi_enable(int cpu) { return 0; }
416static void watchdog_nmi_disable(int cpu) { return; }
417#endif /* CONFIG_HARDLOCKUP_DETECTOR */
418
419/* prepare/enable/disable routines */
420static void watchdog_prepare_cpu(int cpu)
421{
422 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
423
424 WARN_ON(per_cpu(softlockup_watchdog, cpu));
425 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
426 hrtimer->function = watchdog_timer_fn;
427}
428
429static int watchdog_enable(int cpu)
430{
431 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
432 int err = 0;
433
434 /* enable the perf event */
435 err = watchdog_nmi_enable(cpu);
436
437 /* Regardless of err above, fall through and start softlockup */
438
439 /* create the watchdog thread */
440 if (!p) {
441 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
442 if (IS_ERR(p)) {
443 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
444 if (!err) {
445 /* if hardlockup hasn't already set this */
446 err = PTR_ERR(p);
447 /* and disable the perf event */
448 watchdog_nmi_disable(cpu);
449 }
450 goto out;
451 }
452 kthread_bind(p, cpu);
453 per_cpu(watchdog_touch_ts, cpu) = 0;
454 per_cpu(softlockup_watchdog, cpu) = p;
455 wake_up_process(p);
456 }
457
458out:
459 return err;
460}
461
462static void watchdog_disable(int cpu)
463{
464 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
465 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
466
467 /*
468 * cancel the timer first to stop incrementing the stats
469 * and waking up the kthread
470 */
471 hrtimer_cancel(hrtimer);
472
473 /* disable the perf event */
474 watchdog_nmi_disable(cpu);
475
476 /* stop the watchdog thread */
477 if (p) {
478 per_cpu(softlockup_watchdog, cpu) = NULL;
479 kthread_stop(p);
480 }
481}
482
483static void watchdog_enable_all_cpus(void)
484{
485 int cpu;
486
487 watchdog_enabled = 0;
488
489 for_each_online_cpu(cpu)
490 if (!watchdog_enable(cpu))
491 /* if any cpu succeeds, watchdog is considered
492 enabled for the system */
493 watchdog_enabled = 1;
494
495 if (!watchdog_enabled)
496 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
497
498}
499
500static void watchdog_disable_all_cpus(void)
501{
502 int cpu;
503
504 for_each_online_cpu(cpu)
505 watchdog_disable(cpu);
506
507 /* if all watchdogs are disabled, then they are disabled for the system */
508 watchdog_enabled = 0;
509}
510
511
512/* sysctl functions */
513#ifdef CONFIG_SYSCTL
514/*
515 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
516 */
517
518int proc_dowatchdog(struct ctl_table *table, int write,
519 void __user *buffer, size_t *lenp, loff_t *ppos)
520{
521 int ret;
522
523 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524 if (ret || !write)
525 goto out;
526
527 if (watchdog_enabled && watchdog_thresh)
528 watchdog_enable_all_cpus();
529 else
530 watchdog_disable_all_cpus();
531
532out:
533 return ret;
534}
535#endif /* CONFIG_SYSCTL */
536
537
538/*
539 * Create/destroy watchdog threads as CPUs come and go:
540 */
541static int __cpuinit
542cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
543{
544 int hotcpu = (unsigned long)hcpu;
545
546 switch (action) {
547 case CPU_UP_PREPARE:
548 case CPU_UP_PREPARE_FROZEN:
549 watchdog_prepare_cpu(hotcpu);
550 break;
551 case CPU_ONLINE:
552 case CPU_ONLINE_FROZEN:
553 if (watchdog_enabled)
554 watchdog_enable(hotcpu);
555 break;
556#ifdef CONFIG_HOTPLUG_CPU
557 case CPU_UP_CANCELED:
558 case CPU_UP_CANCELED_FROZEN:
559 watchdog_disable(hotcpu);
560 break;
561 case CPU_DEAD:
562 case CPU_DEAD_FROZEN:
563 watchdog_disable(hotcpu);
564 break;
565#endif /* CONFIG_HOTPLUG_CPU */
566 }
567
568 /*
569 * hardlockup and softlockup are not important enough
570 * to block cpu bring up. Just always succeed and
571 * rely on printk output to flag problems.
572 */
573 return NOTIFY_OK;
574}
575
576static struct notifier_block __cpuinitdata cpu_nfb = {
577 .notifier_call = cpu_callback
578};
579
580void __init lockup_detector_init(void)
581{
582 void *cpu = (void *)(long)smp_processor_id();
583 int err;
584
585 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
586 WARN_ON(notifier_to_errno(err));
587
588 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
589 register_cpu_notifier(&cpu_nfb);
590
591 return;
592}