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