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