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
23#include <asm/irq_regs.h>
24#include <linux/kvm_para.h>
25#include <linux/perf_event.h>
26
27/*
28 * The run state of the lockup detectors is controlled by the content of the
29 * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit -
30 * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector.
31 *
32 * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled'
33 * are variables that are only used as an 'interface' between the parameters
34 * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The
35 * 'watchdog_thresh' variable is handled differently because its value is not
36 * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh'
37 * is equal zero.
38 */
39#define NMI_WATCHDOG_ENABLED_BIT   0
40#define SOFT_WATCHDOG_ENABLED_BIT  1
41#define NMI_WATCHDOG_ENABLED      (1 << NMI_WATCHDOG_ENABLED_BIT)
42#define SOFT_WATCHDOG_ENABLED     (1 << SOFT_WATCHDOG_ENABLED_BIT)
43
44static DEFINE_MUTEX(watchdog_proc_mutex);
45
46#ifdef CONFIG_HARDLOCKUP_DETECTOR
47static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED;
48#else
49static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED;
50#endif
51int __read_mostly nmi_watchdog_enabled;
52int __read_mostly soft_watchdog_enabled;
53int __read_mostly watchdog_user_enabled;
54int __read_mostly watchdog_thresh = 10;
55
56#ifdef CONFIG_SMP
57int __read_mostly sysctl_softlockup_all_cpu_backtrace;
58#else
59#define sysctl_softlockup_all_cpu_backtrace 0
60#endif
61
62static int __read_mostly watchdog_running;
63static u64 __read_mostly sample_period;
64
65static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
66static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
67static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
68static DEFINE_PER_CPU(bool, softlockup_touch_sync);
69static DEFINE_PER_CPU(bool, soft_watchdog_warn);
70static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
71static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
72static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
73#ifdef CONFIG_HARDLOCKUP_DETECTOR
74static DEFINE_PER_CPU(bool, hard_watchdog_warn);
75static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
76static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
77static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
78#endif
79static unsigned long soft_lockup_nmi_warn;
80
81/* boot commands */
82/*
83 * Should we panic when a soft-lockup or hard-lockup occurs:
84 */
85#ifdef CONFIG_HARDLOCKUP_DETECTOR
86static int hardlockup_panic =
87			CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
88/*
89 * We may not want to enable hard lockup detection by default in all cases,
90 * for example when running the kernel as a guest on a hypervisor. In these
91 * cases this function can be called to disable hard lockup detection. This
92 * function should only be executed once by the boot processor before the
93 * kernel command line parameters are parsed, because otherwise it is not
94 * possible to override this in hardlockup_panic_setup().
95 */
96void hardlockup_detector_disable(void)
97{
98	watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
99}
100
101static int __init hardlockup_panic_setup(char *str)
102{
103	if (!strncmp(str, "panic", 5))
104		hardlockup_panic = 1;
105	else if (!strncmp(str, "nopanic", 7))
106		hardlockup_panic = 0;
107	else if (!strncmp(str, "0", 1))
108		watchdog_enabled &= ~NMI_WATCHDOG_ENABLED;
109	else if (!strncmp(str, "1", 1))
110		watchdog_enabled |= NMI_WATCHDOG_ENABLED;
111	return 1;
112}
113__setup("nmi_watchdog=", hardlockup_panic_setup);
114#endif
115
116unsigned int __read_mostly softlockup_panic =
117			CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
118
119static int __init softlockup_panic_setup(char *str)
120{
121	softlockup_panic = simple_strtoul(str, NULL, 0);
122
123	return 1;
124}
125__setup("softlockup_panic=", softlockup_panic_setup);
126
127static int __init nowatchdog_setup(char *str)
128{
129	watchdog_enabled = 0;
130	return 1;
131}
132__setup("nowatchdog", nowatchdog_setup);
133
134static int __init nosoftlockup_setup(char *str)
135{
136	watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED;
137	return 1;
138}
139__setup("nosoftlockup", nosoftlockup_setup);
140
141#ifdef CONFIG_SMP
142static int __init softlockup_all_cpu_backtrace_setup(char *str)
143{
144	sysctl_softlockup_all_cpu_backtrace =
145		!!simple_strtol(str, NULL, 0);
146	return 1;
147}
148__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
149#endif
150
151/*
152 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
153 * lockups can have false positives under extreme conditions. So we generally
154 * want a higher threshold for soft lockups than for hard lockups. So we couple
155 * the thresholds with a factor: we make the soft threshold twice the amount of
156 * time the hard threshold is.
157 */
158static int get_softlockup_thresh(void)
159{
160	return watchdog_thresh * 2;
161}
162
163/*
164 * Returns seconds, approximately.  We don't need nanosecond
165 * resolution, and we don't need to waste time with a big divide when
166 * 2^30ns == 1.074s.
167 */
168static unsigned long get_timestamp(void)
169{
170	return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
171}
172
173static void set_sample_period(void)
174{
175	/*
176	 * convert watchdog_thresh from seconds to ns
177	 * the divide by 5 is to give hrtimer several chances (two
178	 * or three with the current relation between the soft
179	 * and hard thresholds) to increment before the
180	 * hardlockup detector generates a warning
181	 */
182	sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
183}
184
185/* Commands for resetting the watchdog */
186static void __touch_watchdog(void)
187{
188	__this_cpu_write(watchdog_touch_ts, get_timestamp());
189}
190
191void touch_softlockup_watchdog(void)
192{
193	/*
194	 * Preemption can be enabled.  It doesn't matter which CPU's timestamp
195	 * gets zeroed here, so use the raw_ operation.
196	 */
197	raw_cpu_write(watchdog_touch_ts, 0);
198}
199EXPORT_SYMBOL(touch_softlockup_watchdog);
200
201void touch_all_softlockup_watchdogs(void)
202{
203	int cpu;
204
205	/*
206	 * this is done lockless
207	 * do we care if a 0 races with a timestamp?
208	 * all it means is the softlock check starts one cycle later
209	 */
210	for_each_online_cpu(cpu)
211		per_cpu(watchdog_touch_ts, cpu) = 0;
212}
213
214#ifdef CONFIG_HARDLOCKUP_DETECTOR
215void touch_nmi_watchdog(void)
216{
217	/*
218	 * Using __raw here because some code paths have
219	 * preemption enabled.  If preemption is enabled
220	 * then interrupts should be enabled too, in which
221	 * case we shouldn't have to worry about the watchdog
222	 * going off.
223	 */
224	raw_cpu_write(watchdog_nmi_touch, true);
225	touch_softlockup_watchdog();
226}
227EXPORT_SYMBOL(touch_nmi_watchdog);
228
229#endif
230
231void touch_softlockup_watchdog_sync(void)
232{
233	__this_cpu_write(softlockup_touch_sync, true);
234	__this_cpu_write(watchdog_touch_ts, 0);
235}
236
237#ifdef CONFIG_HARDLOCKUP_DETECTOR
238/* watchdog detector functions */
239static int is_hardlockup(void)
240{
241	unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
242
243	if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
244		return 1;
245
246	__this_cpu_write(hrtimer_interrupts_saved, hrint);
247	return 0;
248}
249#endif
250
251static int is_softlockup(unsigned long touch_ts)
252{
253	unsigned long now = get_timestamp();
254
255	if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) {
256		/* Warn about unreasonable delays. */
257		if (time_after(now, touch_ts + get_softlockup_thresh()))
258			return now - touch_ts;
259	}
260	return 0;
261}
262
263#ifdef CONFIG_HARDLOCKUP_DETECTOR
264
265static struct perf_event_attr wd_hw_attr = {
266	.type		= PERF_TYPE_HARDWARE,
267	.config		= PERF_COUNT_HW_CPU_CYCLES,
268	.size		= sizeof(struct perf_event_attr),
269	.pinned		= 1,
270	.disabled	= 1,
271};
272
273/* Callback function for perf event subsystem */
274static void watchdog_overflow_callback(struct perf_event *event,
275		 struct perf_sample_data *data,
276		 struct pt_regs *regs)
277{
278	/* Ensure the watchdog never gets throttled */
279	event->hw.interrupts = 0;
280
281	if (__this_cpu_read(watchdog_nmi_touch) == true) {
282		__this_cpu_write(watchdog_nmi_touch, false);
283		return;
284	}
285
286	/* check for a hardlockup
287	 * This is done by making sure our timer interrupt
288	 * is incrementing.  The timer interrupt should have
289	 * fired multiple times before we overflow'd.  If it hasn't
290	 * then this is a good indication the cpu is stuck
291	 */
292	if (is_hardlockup()) {
293		int this_cpu = smp_processor_id();
294
295		/* only print hardlockups once */
296		if (__this_cpu_read(hard_watchdog_warn) == true)
297			return;
298
299		if (hardlockup_panic)
300			panic("Watchdog detected hard LOCKUP on cpu %d",
301			      this_cpu);
302		else
303			WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
304			     this_cpu);
305
306		__this_cpu_write(hard_watchdog_warn, true);
307		return;
308	}
309
310	__this_cpu_write(hard_watchdog_warn, false);
311	return;
312}
313#endif /* CONFIG_HARDLOCKUP_DETECTOR */
314
315static void watchdog_interrupt_count(void)
316{
317	__this_cpu_inc(hrtimer_interrupts);
318}
319
320static int watchdog_nmi_enable(unsigned int cpu);
321static void watchdog_nmi_disable(unsigned int cpu);
322
323/* watchdog kicker functions */
324static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
325{
326	unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
327	struct pt_regs *regs = get_irq_regs();
328	int duration;
329	int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
330
331	/* kick the hardlockup detector */
332	watchdog_interrupt_count();
333
334	/* kick the softlockup detector */
335	wake_up_process(__this_cpu_read(softlockup_watchdog));
336
337	/* .. and repeat */
338	hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
339
340	if (touch_ts == 0) {
341		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
342			/*
343			 * If the time stamp was touched atomically
344			 * make sure the scheduler tick is up to date.
345			 */
346			__this_cpu_write(softlockup_touch_sync, false);
347			sched_clock_tick();
348		}
349
350		/* Clear the guest paused flag on watchdog reset */
351		kvm_check_and_clear_guest_paused();
352		__touch_watchdog();
353		return HRTIMER_RESTART;
354	}
355
356	/* check for a softlockup
357	 * This is done by making sure a high priority task is
358	 * being scheduled.  The task touches the watchdog to
359	 * indicate it is getting cpu time.  If it hasn't then
360	 * this is a good indication some task is hogging the cpu
361	 */
362	duration = is_softlockup(touch_ts);
363	if (unlikely(duration)) {
364		/*
365		 * If a virtual machine is stopped by the host it can look to
366		 * the watchdog like a soft lockup, check to see if the host
367		 * stopped the vm before we issue the warning
368		 */
369		if (kvm_check_and_clear_guest_paused())
370			return HRTIMER_RESTART;
371
372		/* only warn once */
373		if (__this_cpu_read(soft_watchdog_warn) == true) {
374			/*
375			 * When multiple processes are causing softlockups the
376			 * softlockup detector only warns on the first one
377			 * because the code relies on a full quiet cycle to
378			 * re-arm.  The second process prevents the quiet cycle
379			 * and never gets reported.  Use task pointers to detect
380			 * this.
381			 */
382			if (__this_cpu_read(softlockup_task_ptr_saved) !=
383			    current) {
384				__this_cpu_write(soft_watchdog_warn, false);
385				__touch_watchdog();
386			}
387			return HRTIMER_RESTART;
388		}
389
390		if (softlockup_all_cpu_backtrace) {
391			/* Prevent multiple soft-lockup reports if one cpu is already
392			 * engaged in dumping cpu back traces
393			 */
394			if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
395				/* Someone else will report us. Let's give up */
396				__this_cpu_write(soft_watchdog_warn, true);
397				return HRTIMER_RESTART;
398			}
399		}
400
401		pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
402			smp_processor_id(), duration,
403			current->comm, task_pid_nr(current));
404		__this_cpu_write(softlockup_task_ptr_saved, current);
405		print_modules();
406		print_irqtrace_events(current);
407		if (regs)
408			show_regs(regs);
409		else
410			dump_stack();
411
412		if (softlockup_all_cpu_backtrace) {
413			/* Avoid generating two back traces for current
414			 * given that one is already made above
415			 */
416			trigger_allbutself_cpu_backtrace();
417
418			clear_bit(0, &soft_lockup_nmi_warn);
419			/* Barrier to sync with other cpus */
420			smp_mb__after_atomic();
421		}
422
423		add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
424		if (softlockup_panic)
425			panic("softlockup: hung tasks");
426		__this_cpu_write(soft_watchdog_warn, true);
427	} else
428		__this_cpu_write(soft_watchdog_warn, false);
429
430	return HRTIMER_RESTART;
431}
432
433static void watchdog_set_prio(unsigned int policy, unsigned int prio)
434{
435	struct sched_param param = { .sched_priority = prio };
436
437	sched_setscheduler(current, policy, &param);
438}
439
440static void watchdog_enable(unsigned int cpu)
441{
442	struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
443
444	/* kick off the timer for the hardlockup detector */
445	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
446	hrtimer->function = watchdog_timer_fn;
447
448	/* Enable the perf event */
449	watchdog_nmi_enable(cpu);
450
451	/* done here because hrtimer_start can only pin to smp_processor_id() */
452	hrtimer_start(hrtimer, ns_to_ktime(sample_period),
453		      HRTIMER_MODE_REL_PINNED);
454
455	/* initialize timestamp */
456	watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
457	__touch_watchdog();
458}
459
460static void watchdog_disable(unsigned int cpu)
461{
462	struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
463
464	watchdog_set_prio(SCHED_NORMAL, 0);
465	hrtimer_cancel(hrtimer);
466	/* disable the perf event */
467	watchdog_nmi_disable(cpu);
468}
469
470static void watchdog_cleanup(unsigned int cpu, bool online)
471{
472	watchdog_disable(cpu);
473}
474
475static int watchdog_should_run(unsigned int cpu)
476{
477	return __this_cpu_read(hrtimer_interrupts) !=
478		__this_cpu_read(soft_lockup_hrtimer_cnt);
479}
480
481/*
482 * The watchdog thread function - touches the timestamp.
483 *
484 * It only runs once every sample_period seconds (4 seconds by
485 * default) to reset the softlockup timestamp. If this gets delayed
486 * for more than 2*watchdog_thresh seconds then the debug-printout
487 * triggers in watchdog_timer_fn().
488 */
489static void watchdog(unsigned int cpu)
490{
491	__this_cpu_write(soft_lockup_hrtimer_cnt,
492			 __this_cpu_read(hrtimer_interrupts));
493	__touch_watchdog();
494
495	/*
496	 * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the
497	 * failure path. Check for failures that can occur asynchronously -
498	 * for example, when CPUs are on-lined - and shut down the hardware
499	 * perf event on each CPU accordingly.
500	 *
501	 * The only non-obvious place this bit can be cleared is through
502	 * watchdog_nmi_enable(), so a pr_info() is placed there.  Placing a
503	 * pr_info here would be too noisy as it would result in a message
504	 * every few seconds if the hardlockup was disabled but the softlockup
505	 * enabled.
506	 */
507	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
508		watchdog_nmi_disable(cpu);
509}
510
511#ifdef CONFIG_HARDLOCKUP_DETECTOR
512/*
513 * People like the simple clean cpu node info on boot.
514 * Reduce the watchdog noise by only printing messages
515 * that are different from what cpu0 displayed.
516 */
517static unsigned long cpu0_err;
518
519static int watchdog_nmi_enable(unsigned int cpu)
520{
521	struct perf_event_attr *wd_attr;
522	struct perf_event *event = per_cpu(watchdog_ev, cpu);
523
524	/* nothing to do if the hard lockup detector is disabled */
525	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
526		goto out;
527
528	/* is it already setup and enabled? */
529	if (event && event->state > PERF_EVENT_STATE_OFF)
530		goto out;
531
532	/* it is setup but not enabled */
533	if (event != NULL)
534		goto out_enable;
535
536	wd_attr = &wd_hw_attr;
537	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
538
539	/* Try to register using hardware perf events */
540	event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
541
542	/* save cpu0 error for future comparision */
543	if (cpu == 0 && IS_ERR(event))
544		cpu0_err = PTR_ERR(event);
545
546	if (!IS_ERR(event)) {
547		/* only print for cpu0 or different than cpu0 */
548		if (cpu == 0 || cpu0_err)
549			pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
550		goto out_save;
551	}
552
553	/*
554	 * Disable the hard lockup detector if _any_ CPU fails to set up
555	 * set up the hardware perf event. The watchdog() function checks
556	 * the NMI_WATCHDOG_ENABLED bit periodically.
557	 *
558	 * The barriers are for syncing up watchdog_enabled across all the
559	 * cpus, as clear_bit() does not use barriers.
560	 */
561	smp_mb__before_atomic();
562	clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled);
563	smp_mb__after_atomic();
564
565	/* skip displaying the same error again */
566	if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
567		return PTR_ERR(event);
568
569	/* vary the KERN level based on the returned errno */
570	if (PTR_ERR(event) == -EOPNOTSUPP)
571		pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
572	else if (PTR_ERR(event) == -ENOENT)
573		pr_warn("disabled (cpu%i): hardware events not enabled\n",
574			 cpu);
575	else
576		pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
577			cpu, PTR_ERR(event));
578
579	pr_info("Shutting down hard lockup detector on all cpus\n");
580
581	return PTR_ERR(event);
582
583	/* success path */
584out_save:
585	per_cpu(watchdog_ev, cpu) = event;
586out_enable:
587	perf_event_enable(per_cpu(watchdog_ev, cpu));
588out:
589	return 0;
590}
591
592static void watchdog_nmi_disable(unsigned int cpu)
593{
594	struct perf_event *event = per_cpu(watchdog_ev, cpu);
595
596	if (event) {
597		perf_event_disable(event);
598		per_cpu(watchdog_ev, cpu) = NULL;
599
600		/* should be in cleanup, but blocks oprofile */
601		perf_event_release_kernel(event);
602	}
603	if (cpu == 0) {
604		/* watchdog_nmi_enable() expects this to be zero initially. */
605		cpu0_err = 0;
606	}
607}
608
609void watchdog_nmi_enable_all(void)
610{
611	int cpu;
612
613	mutex_lock(&watchdog_proc_mutex);
614
615	if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED))
616		goto unlock;
617
618	get_online_cpus();
619	for_each_online_cpu(cpu)
620		watchdog_nmi_enable(cpu);
621	put_online_cpus();
622
623unlock:
624	mutex_unlock(&watchdog_proc_mutex);
625}
626
627void watchdog_nmi_disable_all(void)
628{
629	int cpu;
630
631	mutex_lock(&watchdog_proc_mutex);
632
633	if (!watchdog_running)
634		goto unlock;
635
636	get_online_cpus();
637	for_each_online_cpu(cpu)
638		watchdog_nmi_disable(cpu);
639	put_online_cpus();
640
641unlock:
642	mutex_unlock(&watchdog_proc_mutex);
643}
644#else
645static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
646static void watchdog_nmi_disable(unsigned int cpu) { return; }
647void watchdog_nmi_enable_all(void) {}
648void watchdog_nmi_disable_all(void) {}
649#endif /* CONFIG_HARDLOCKUP_DETECTOR */
650
651static struct smp_hotplug_thread watchdog_threads = {
652	.store			= &softlockup_watchdog,
653	.thread_should_run	= watchdog_should_run,
654	.thread_fn		= watchdog,
655	.thread_comm		= "watchdog/%u",
656	.setup			= watchdog_enable,
657	.cleanup		= watchdog_cleanup,
658	.park			= watchdog_disable,
659	.unpark			= watchdog_enable,
660};
661
662static void restart_watchdog_hrtimer(void *info)
663{
664	struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
665	int ret;
666
667	/*
668	 * No need to cancel and restart hrtimer if it is currently executing
669	 * because it will reprogram itself with the new period now.
670	 * We should never see it unqueued here because we are running per-cpu
671	 * with interrupts disabled.
672	 */
673	ret = hrtimer_try_to_cancel(hrtimer);
674	if (ret == 1)
675		hrtimer_start(hrtimer, ns_to_ktime(sample_period),
676				HRTIMER_MODE_REL_PINNED);
677}
678
679static void update_watchdog(int cpu)
680{
681	/*
682	 * Make sure that perf event counter will adopt to a new
683	 * sampling period. Updating the sampling period directly would
684	 * be much nicer but we do not have an API for that now so
685	 * let's use a big hammer.
686	 * Hrtimer will adopt the new period on the next tick but this
687	 * might be late already so we have to restart the timer as well.
688	 */
689	watchdog_nmi_disable(cpu);
690	smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
691	watchdog_nmi_enable(cpu);
692}
693
694static void update_watchdog_all_cpus(void)
695{
696	int cpu;
697
698	get_online_cpus();
699	for_each_online_cpu(cpu)
700		update_watchdog(cpu);
701	put_online_cpus();
702}
703
704static int watchdog_enable_all_cpus(void)
705{
706	int err = 0;
707
708	if (!watchdog_running) {
709		err = smpboot_register_percpu_thread(&watchdog_threads);
710		if (err)
711			pr_err("Failed to create watchdog threads, disabled\n");
712		else
713			watchdog_running = 1;
714	} else {
715		/*
716		 * Enable/disable the lockup detectors or
717		 * change the sample period 'on the fly'.
718		 */
719		update_watchdog_all_cpus();
720	}
721
722	return err;
723}
724
725/* prepare/enable/disable routines */
726/* sysctl functions */
727#ifdef CONFIG_SYSCTL
728static void watchdog_disable_all_cpus(void)
729{
730	if (watchdog_running) {
731		watchdog_running = 0;
732		smpboot_unregister_percpu_thread(&watchdog_threads);
733	}
734}
735
736/*
737 * Update the run state of the lockup detectors.
738 */
739static int proc_watchdog_update(void)
740{
741	int err = 0;
742
743	/*
744	 * Watchdog threads won't be started if they are already active.
745	 * The 'watchdog_running' variable in watchdog_*_all_cpus() takes
746	 * care of this. If those threads are already active, the sample
747	 * period will be updated and the lockup detectors will be enabled
748	 * or disabled 'on the fly'.
749	 */
750	if (watchdog_enabled && watchdog_thresh)
751		err = watchdog_enable_all_cpus();
752	else
753		watchdog_disable_all_cpus();
754
755	return err;
756
757}
758
759/*
760 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
761 *
762 * caller             | table->data points to | 'which' contains the flag(s)
763 * -------------------|-----------------------|-----------------------------
764 * proc_watchdog      | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed
765 *                    |                       | with SOFT_WATCHDOG_ENABLED
766 * -------------------|-----------------------|-----------------------------
767 * proc_nmi_watchdog  | nmi_watchdog_enabled  | NMI_WATCHDOG_ENABLED
768 * -------------------|-----------------------|-----------------------------
769 * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED
770 */
771static int proc_watchdog_common(int which, struct ctl_table *table, int write,
772				void __user *buffer, size_t *lenp, loff_t *ppos)
773{
774	int err, old, new;
775	int *watchdog_param = (int *)table->data;
776
777	mutex_lock(&watchdog_proc_mutex);
778
779	/*
780	 * If the parameter is being read return the state of the corresponding
781	 * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the
782	 * run state of the lockup detectors.
783	 */
784	if (!write) {
785		*watchdog_param = (watchdog_enabled & which) != 0;
786		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
787	} else {
788		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
789		if (err)
790			goto out;
791
792		/*
793		 * There is a race window between fetching the current value
794		 * from 'watchdog_enabled' and storing the new value. During
795		 * this race window, watchdog_nmi_enable() can sneak in and
796		 * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'.
797		 * The 'cmpxchg' detects this race and the loop retries.
798		 */
799		do {
800			old = watchdog_enabled;
801			/*
802			 * If the parameter value is not zero set the
803			 * corresponding bit(s), else clear it(them).
804			 */
805			if (*watchdog_param)
806				new = old | which;
807			else
808				new = old & ~which;
809		} while (cmpxchg(&watchdog_enabled, old, new) != old);
810
811		/*
812		 * Update the run state of the lockup detectors.
813		 * Restore 'watchdog_enabled' on failure.
814		 */
815		if (old == new)
816			goto out;
817
818		err = proc_watchdog_update();
819		if (err)
820			watchdog_enabled = old;
821	}
822out:
823	mutex_unlock(&watchdog_proc_mutex);
824	return err;
825}
826
827/*
828 * /proc/sys/kernel/watchdog
829 */
830int proc_watchdog(struct ctl_table *table, int write,
831		  void __user *buffer, size_t *lenp, loff_t *ppos)
832{
833	return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
834				    table, write, buffer, lenp, ppos);
835}
836
837/*
838 * /proc/sys/kernel/nmi_watchdog
839 */
840int proc_nmi_watchdog(struct ctl_table *table, int write,
841		      void __user *buffer, size_t *lenp, loff_t *ppos)
842{
843	return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
844				    table, write, buffer, lenp, ppos);
845}
846
847/*
848 * /proc/sys/kernel/soft_watchdog
849 */
850int proc_soft_watchdog(struct ctl_table *table, int write,
851			void __user *buffer, size_t *lenp, loff_t *ppos)
852{
853	return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
854				    table, write, buffer, lenp, ppos);
855}
856
857/*
858 * /proc/sys/kernel/watchdog_thresh
859 */
860int proc_watchdog_thresh(struct ctl_table *table, int write,
861			 void __user *buffer, size_t *lenp, loff_t *ppos)
862{
863	int err, old, new;
864
865	mutex_lock(&watchdog_proc_mutex);
866
867	old = ACCESS_ONCE(watchdog_thresh);
868	err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
869
870	if (err || !write)
871		goto out;
872
873	/*
874	 * Update the sample period.
875	 * Restore 'watchdog_thresh' on failure.
876	 */
877	new = ACCESS_ONCE(watchdog_thresh);
878	if (old == new)
879		goto out;
880
881	set_sample_period();
882	err = proc_watchdog_update();
883	if (err)
884		watchdog_thresh = old;
885out:
886	mutex_unlock(&watchdog_proc_mutex);
887	return err;
888}
889#endif /* CONFIG_SYSCTL */
890
891void __init lockup_detector_init(void)
892{
893	set_sample_period();
894
895	if (watchdog_enabled)
896		watchdog_enable_all_cpus();
897}
898