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_HARDLOCKUP_DETECTOR_SPARC64)
33 # define WATCHDOG_HARDLOCKUP_DEFAULT	1
34 #else
35 # define WATCHDOG_HARDLOCKUP_DEFAULT	0
36 #endif
37 
38 unsigned long __read_mostly watchdog_enabled;
39 int __read_mostly watchdog_user_enabled = 1;
40 static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT;
41 static int __read_mostly watchdog_softlockup_user_enabled = 1;
42 int __read_mostly watchdog_thresh = 10;
43 static int __read_mostly watchdog_hardlockup_available;
44 
45 struct cpumask watchdog_cpumask __read_mostly;
46 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
47 
48 #ifdef CONFIG_HARDLOCKUP_DETECTOR
49 
50 # ifdef CONFIG_SMP
51 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
52 # endif /* CONFIG_SMP */
53 
54 /*
55  * Should we panic when a soft-lockup or hard-lockup occurs:
56  */
57 unsigned int __read_mostly hardlockup_panic =
58 			IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
59 /*
60  * We may not want to enable hard lockup detection by default in all cases,
61  * for example when running the kernel as a guest on a hypervisor. In these
62  * cases this function can be called to disable hard lockup detection. This
63  * function should only be executed once by the boot processor before the
64  * kernel command line parameters are parsed, because otherwise it is not
65  * possible to override this in hardlockup_panic_setup().
66  */
hardlockup_detector_disable(void)67 void __init hardlockup_detector_disable(void)
68 {
69 	watchdog_hardlockup_user_enabled = 0;
70 }
71 
hardlockup_panic_setup(char * str)72 static int __init hardlockup_panic_setup(char *str)
73 {
74 	if (!strncmp(str, "panic", 5))
75 		hardlockup_panic = 1;
76 	else if (!strncmp(str, "nopanic", 7))
77 		hardlockup_panic = 0;
78 	else if (!strncmp(str, "0", 1))
79 		watchdog_hardlockup_user_enabled = 0;
80 	else if (!strncmp(str, "1", 1))
81 		watchdog_hardlockup_user_enabled = 1;
82 	return 1;
83 }
84 __setup("nmi_watchdog=", hardlockup_panic_setup);
85 
86 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
87 
88 #if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER)
89 
90 static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts);
91 static DEFINE_PER_CPU(int, hrtimer_interrupts_saved);
92 static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned);
93 static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched);
94 static unsigned long watchdog_hardlockup_all_cpu_dumped;
95 
arch_touch_nmi_watchdog(void)96 notrace void arch_touch_nmi_watchdog(void)
97 {
98 	/*
99 	 * Using __raw here because some code paths have
100 	 * preemption enabled.  If preemption is enabled
101 	 * then interrupts should be enabled too, in which
102 	 * case we shouldn't have to worry about the watchdog
103 	 * going off.
104 	 */
105 	raw_cpu_write(watchdog_hardlockup_touched, true);
106 }
107 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
108 
watchdog_hardlockup_touch_cpu(unsigned int cpu)109 void watchdog_hardlockup_touch_cpu(unsigned int cpu)
110 {
111 	per_cpu(watchdog_hardlockup_touched, cpu) = true;
112 }
113 
is_hardlockup(unsigned int cpu)114 static bool is_hardlockup(unsigned int cpu)
115 {
116 	int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu));
117 
118 	if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
119 		return true;
120 
121 	/*
122 	 * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE
123 	 * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is
124 	 * written/read by a single CPU.
125 	 */
126 	per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
127 
128 	return false;
129 }
130 
watchdog_hardlockup_kick(void)131 static void watchdog_hardlockup_kick(void)
132 {
133 	int new_interrupts;
134 
135 	new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts));
136 	watchdog_buddy_check_hardlockup(new_interrupts);
137 }
138 
watchdog_hardlockup_check(unsigned int cpu,struct pt_regs * regs)139 void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
140 {
141 	if (per_cpu(watchdog_hardlockup_touched, cpu)) {
142 		per_cpu(watchdog_hardlockup_touched, cpu) = false;
143 		return;
144 	}
145 
146 	/*
147 	 * Check for a hardlockup by making sure the CPU's timer
148 	 * interrupt is incrementing. The timer interrupt should have
149 	 * fired multiple times before we overflow'd. If it hasn't
150 	 * then this is a good indication the cpu is stuck
151 	 */
152 	if (is_hardlockup(cpu)) {
153 		unsigned int this_cpu = smp_processor_id();
154 
155 		/* Only print hardlockups once. */
156 		if (per_cpu(watchdog_hardlockup_warned, cpu))
157 			return;
158 
159 		pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
160 		print_modules();
161 		print_irqtrace_events(current);
162 		if (cpu == this_cpu) {
163 			if (regs)
164 				show_regs(regs);
165 			else
166 				dump_stack();
167 		} else {
168 			trigger_single_cpu_backtrace(cpu);
169 		}
170 
171 		/*
172 		 * Perform multi-CPU dump only once to avoid multiple
173 		 * hardlockups generating interleaving traces
174 		 */
175 		if (sysctl_hardlockup_all_cpu_backtrace &&
176 		    !test_and_set_bit(0, &watchdog_hardlockup_all_cpu_dumped))
177 			trigger_allbutcpu_cpu_backtrace(cpu);
178 
179 		if (hardlockup_panic)
180 			nmi_panic(regs, "Hard LOCKUP");
181 
182 		per_cpu(watchdog_hardlockup_warned, cpu) = true;
183 	} else {
184 		per_cpu(watchdog_hardlockup_warned, cpu) = false;
185 	}
186 }
187 
188 #else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
189 
watchdog_hardlockup_kick(void)190 static inline void watchdog_hardlockup_kick(void) { }
191 
192 #endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
193 
194 /*
195  * These functions can be overridden based on the configured hardlockdup detector.
196  *
197  * watchdog_hardlockup_enable/disable can be implemented to start and stop when
198  * softlockup watchdog start and stop. The detector must select the
199  * SOFTLOCKUP_DETECTOR Kconfig.
200  */
watchdog_hardlockup_enable(unsigned int cpu)201 void __weak watchdog_hardlockup_enable(unsigned int cpu) { }
202 
watchdog_hardlockup_disable(unsigned int cpu)203 void __weak watchdog_hardlockup_disable(unsigned int cpu) { }
204 
205 /*
206  * Watchdog-detector specific API.
207  *
208  * Return 0 when hardlockup watchdog is available, negative value otherwise.
209  * Note that the negative value means that a delayed probe might
210  * succeed later.
211  */
watchdog_hardlockup_probe(void)212 int __weak __init watchdog_hardlockup_probe(void)
213 {
214 	return -ENODEV;
215 }
216 
217 /**
218  * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration
219  *
220  * The reconfiguration steps are:
221  * watchdog_hardlockup_stop();
222  * update_variables();
223  * watchdog_hardlockup_start();
224  */
watchdog_hardlockup_stop(void)225 void __weak watchdog_hardlockup_stop(void) { }
226 
227 /**
228  * watchdog_hardlockup_start - Start the watchdog after reconfiguration
229  *
230  * Counterpart to watchdog_hardlockup_stop().
231  *
232  * The following variables have been updated in update_variables() and
233  * contain the currently valid configuration:
234  * - watchdog_enabled
235  * - watchdog_thresh
236  * - watchdog_cpumask
237  */
watchdog_hardlockup_start(void)238 void __weak watchdog_hardlockup_start(void) { }
239 
240 /**
241  * lockup_detector_update_enable - Update the sysctl enable bit
242  *
243  * Caller needs to make sure that the hard watchdogs are off, so this
244  * can't race with watchdog_hardlockup_disable().
245  */
lockup_detector_update_enable(void)246 static void lockup_detector_update_enable(void)
247 {
248 	watchdog_enabled = 0;
249 	if (!watchdog_user_enabled)
250 		return;
251 	if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled)
252 		watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED;
253 	if (watchdog_softlockup_user_enabled)
254 		watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED;
255 }
256 
257 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
258 
259 /*
260  * Delay the soflockup report when running a known slow code.
261  * It does _not_ affect the timestamp of the last successdul reschedule.
262  */
263 #define SOFTLOCKUP_DELAY_REPORT	ULONG_MAX
264 
265 #ifdef CONFIG_SMP
266 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
267 #endif
268 
269 static struct cpumask watchdog_allowed_mask __read_mostly;
270 
271 /* Global variables, exported for sysctl */
272 unsigned int __read_mostly softlockup_panic =
273 			IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
274 
275 static bool softlockup_initialized __read_mostly;
276 static u64 __read_mostly sample_period;
277 
278 /* Timestamp taken after the last successful reschedule. */
279 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
280 /* Timestamp of the last softlockup report. */
281 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
282 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
283 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
284 static unsigned long soft_lockup_nmi_warn;
285 
softlockup_panic_setup(char * str)286 static int __init softlockup_panic_setup(char *str)
287 {
288 	softlockup_panic = simple_strtoul(str, NULL, 0);
289 	return 1;
290 }
291 __setup("softlockup_panic=", softlockup_panic_setup);
292 
nowatchdog_setup(char * str)293 static int __init nowatchdog_setup(char *str)
294 {
295 	watchdog_user_enabled = 0;
296 	return 1;
297 }
298 __setup("nowatchdog", nowatchdog_setup);
299 
nosoftlockup_setup(char * str)300 static int __init nosoftlockup_setup(char *str)
301 {
302 	watchdog_softlockup_user_enabled = 0;
303 	return 1;
304 }
305 __setup("nosoftlockup", nosoftlockup_setup);
306 
watchdog_thresh_setup(char * str)307 static int __init watchdog_thresh_setup(char *str)
308 {
309 	get_option(&str, &watchdog_thresh);
310 	return 1;
311 }
312 __setup("watchdog_thresh=", watchdog_thresh_setup);
313 
314 static void __lockup_detector_cleanup(void);
315 
316 /*
317  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
318  * lockups can have false positives under extreme conditions. So we generally
319  * want a higher threshold for soft lockups than for hard lockups. So we couple
320  * the thresholds with a factor: we make the soft threshold twice the amount of
321  * time the hard threshold is.
322  */
get_softlockup_thresh(void)323 static int get_softlockup_thresh(void)
324 {
325 	return watchdog_thresh * 2;
326 }
327 
328 /*
329  * Returns seconds, approximately.  We don't need nanosecond
330  * resolution, and we don't need to waste time with a big divide when
331  * 2^30ns == 1.074s.
332  */
get_timestamp(void)333 static unsigned long get_timestamp(void)
334 {
335 	return running_clock() >> 30LL;  /* 2^30 ~= 10^9 */
336 }
337 
set_sample_period(void)338 static void set_sample_period(void)
339 {
340 	/*
341 	 * convert watchdog_thresh from seconds to ns
342 	 * the divide by 5 is to give hrtimer several chances (two
343 	 * or three with the current relation between the soft
344 	 * and hard thresholds) to increment before the
345 	 * hardlockup detector generates a warning
346 	 */
347 	sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
348 	watchdog_update_hrtimer_threshold(sample_period);
349 }
350 
update_report_ts(void)351 static void update_report_ts(void)
352 {
353 	__this_cpu_write(watchdog_report_ts, get_timestamp());
354 }
355 
356 /* Commands for resetting the watchdog */
update_touch_ts(void)357 static void update_touch_ts(void)
358 {
359 	__this_cpu_write(watchdog_touch_ts, get_timestamp());
360 	update_report_ts();
361 }
362 
363 /**
364  * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
365  *
366  * Call when the scheduler may have stalled for legitimate reasons
367  * preventing the watchdog task from executing - e.g. the scheduler
368  * entering idle state.  This should only be used for scheduler events.
369  * Use touch_softlockup_watchdog() for everything else.
370  */
touch_softlockup_watchdog_sched(void)371 notrace void touch_softlockup_watchdog_sched(void)
372 {
373 	/*
374 	 * Preemption can be enabled.  It doesn't matter which CPU's watchdog
375 	 * report period gets restarted here, so use the raw_ operation.
376 	 */
377 	raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
378 }
379 
touch_softlockup_watchdog(void)380 notrace void touch_softlockup_watchdog(void)
381 {
382 	touch_softlockup_watchdog_sched();
383 	wq_watchdog_touch(raw_smp_processor_id());
384 }
385 EXPORT_SYMBOL(touch_softlockup_watchdog);
386 
touch_all_softlockup_watchdogs(void)387 void touch_all_softlockup_watchdogs(void)
388 {
389 	int cpu;
390 
391 	/*
392 	 * watchdog_mutex cannpt be taken here, as this might be called
393 	 * from (soft)interrupt context, so the access to
394 	 * watchdog_allowed_cpumask might race with a concurrent update.
395 	 *
396 	 * The watchdog time stamp can race against a concurrent real
397 	 * update as well, the only side effect might be a cycle delay for
398 	 * the softlockup check.
399 	 */
400 	for_each_cpu(cpu, &watchdog_allowed_mask) {
401 		per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
402 		wq_watchdog_touch(cpu);
403 	}
404 }
405 
touch_softlockup_watchdog_sync(void)406 void touch_softlockup_watchdog_sync(void)
407 {
408 	__this_cpu_write(softlockup_touch_sync, true);
409 	__this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
410 }
411 
is_softlockup(unsigned long touch_ts,unsigned long period_ts,unsigned long now)412 static int is_softlockup(unsigned long touch_ts,
413 			 unsigned long period_ts,
414 			 unsigned long now)
415 {
416 	if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
417 		/* Warn about unreasonable delays. */
418 		if (time_after(now, period_ts + get_softlockup_thresh()))
419 			return now - touch_ts;
420 	}
421 	return 0;
422 }
423 
424 /* watchdog detector functions */
425 static DEFINE_PER_CPU(struct completion, softlockup_completion);
426 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
427 
428 /*
429  * The watchdog feed function - touches the timestamp.
430  *
431  * It only runs once every sample_period seconds (4 seconds by
432  * default) to reset the softlockup timestamp. If this gets delayed
433  * for more than 2*watchdog_thresh seconds then the debug-printout
434  * triggers in watchdog_timer_fn().
435  */
softlockup_fn(void * data)436 static int softlockup_fn(void *data)
437 {
438 	update_touch_ts();
439 	complete(this_cpu_ptr(&softlockup_completion));
440 
441 	return 0;
442 }
443 
444 /* watchdog kicker functions */
watchdog_timer_fn(struct hrtimer * hrtimer)445 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
446 {
447 	unsigned long touch_ts, period_ts, now;
448 	struct pt_regs *regs = get_irq_regs();
449 	int duration;
450 	int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
451 
452 	if (!watchdog_enabled)
453 		return HRTIMER_NORESTART;
454 
455 	watchdog_hardlockup_kick();
456 
457 	/* kick the softlockup detector */
458 	if (completion_done(this_cpu_ptr(&softlockup_completion))) {
459 		reinit_completion(this_cpu_ptr(&softlockup_completion));
460 		stop_one_cpu_nowait(smp_processor_id(),
461 				softlockup_fn, NULL,
462 				this_cpu_ptr(&softlockup_stop_work));
463 	}
464 
465 	/* .. and repeat */
466 	hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
467 
468 	/*
469 	 * Read the current timestamp first. It might become invalid anytime
470 	 * when a virtual machine is stopped by the host or when the watchog
471 	 * is touched from NMI.
472 	 */
473 	now = get_timestamp();
474 	/*
475 	 * If a virtual machine is stopped by the host it can look to
476 	 * the watchdog like a soft lockup. This function touches the watchdog.
477 	 */
478 	kvm_check_and_clear_guest_paused();
479 	/*
480 	 * The stored timestamp is comparable with @now only when not touched.
481 	 * It might get touched anytime from NMI. Make sure that is_softlockup()
482 	 * uses the same (valid) value.
483 	 */
484 	period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
485 
486 	/* Reset the interval when touched by known problematic code. */
487 	if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
488 		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
489 			/*
490 			 * If the time stamp was touched atomically
491 			 * make sure the scheduler tick is up to date.
492 			 */
493 			__this_cpu_write(softlockup_touch_sync, false);
494 			sched_clock_tick();
495 		}
496 
497 		update_report_ts();
498 		return HRTIMER_RESTART;
499 	}
500 
501 	/* Check for a softlockup. */
502 	touch_ts = __this_cpu_read(watchdog_touch_ts);
503 	duration = is_softlockup(touch_ts, period_ts, now);
504 	if (unlikely(duration)) {
505 		/*
506 		 * Prevent multiple soft-lockup reports if one cpu is already
507 		 * engaged in dumping all cpu back traces.
508 		 */
509 		if (softlockup_all_cpu_backtrace) {
510 			if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
511 				return HRTIMER_RESTART;
512 		}
513 
514 		/* Start period for the next softlockup warning. */
515 		update_report_ts();
516 
517 		pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
518 			smp_processor_id(), duration,
519 			current->comm, task_pid_nr(current));
520 		print_modules();
521 		print_irqtrace_events(current);
522 		if (regs)
523 			show_regs(regs);
524 		else
525 			dump_stack();
526 
527 		if (softlockup_all_cpu_backtrace) {
528 			trigger_allbutcpu_cpu_backtrace(smp_processor_id());
529 			clear_bit_unlock(0, &soft_lockup_nmi_warn);
530 		}
531 
532 		add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
533 		if (softlockup_panic)
534 			panic("softlockup: hung tasks");
535 	}
536 
537 	return HRTIMER_RESTART;
538 }
539 
watchdog_enable(unsigned int cpu)540 static void watchdog_enable(unsigned int cpu)
541 {
542 	struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
543 	struct completion *done = this_cpu_ptr(&softlockup_completion);
544 
545 	WARN_ON_ONCE(cpu != smp_processor_id());
546 
547 	init_completion(done);
548 	complete(done);
549 
550 	/*
551 	 * Start the timer first to prevent the hardlockup watchdog triggering
552 	 * before the timer has a chance to fire.
553 	 */
554 	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
555 	hrtimer->function = watchdog_timer_fn;
556 	hrtimer_start(hrtimer, ns_to_ktime(sample_period),
557 		      HRTIMER_MODE_REL_PINNED_HARD);
558 
559 	/* Initialize timestamp */
560 	update_touch_ts();
561 	/* Enable the hardlockup detector */
562 	if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)
563 		watchdog_hardlockup_enable(cpu);
564 }
565 
watchdog_disable(unsigned int cpu)566 static void watchdog_disable(unsigned int cpu)
567 {
568 	struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
569 
570 	WARN_ON_ONCE(cpu != smp_processor_id());
571 
572 	/*
573 	 * Disable the hardlockup detector first. That prevents that a large
574 	 * delay between disabling the timer and disabling the hardlockup
575 	 * detector causes a false positive.
576 	 */
577 	watchdog_hardlockup_disable(cpu);
578 	hrtimer_cancel(hrtimer);
579 	wait_for_completion(this_cpu_ptr(&softlockup_completion));
580 }
581 
softlockup_stop_fn(void * data)582 static int softlockup_stop_fn(void *data)
583 {
584 	watchdog_disable(smp_processor_id());
585 	return 0;
586 }
587 
softlockup_stop_all(void)588 static void softlockup_stop_all(void)
589 {
590 	int cpu;
591 
592 	if (!softlockup_initialized)
593 		return;
594 
595 	for_each_cpu(cpu, &watchdog_allowed_mask)
596 		smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
597 
598 	cpumask_clear(&watchdog_allowed_mask);
599 }
600 
softlockup_start_fn(void * data)601 static int softlockup_start_fn(void *data)
602 {
603 	watchdog_enable(smp_processor_id());
604 	return 0;
605 }
606 
softlockup_start_all(void)607 static void softlockup_start_all(void)
608 {
609 	int cpu;
610 
611 	cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
612 	for_each_cpu(cpu, &watchdog_allowed_mask)
613 		smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
614 }
615 
lockup_detector_online_cpu(unsigned int cpu)616 int lockup_detector_online_cpu(unsigned int cpu)
617 {
618 	if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
619 		watchdog_enable(cpu);
620 	return 0;
621 }
622 
lockup_detector_offline_cpu(unsigned int cpu)623 int lockup_detector_offline_cpu(unsigned int cpu)
624 {
625 	if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
626 		watchdog_disable(cpu);
627 	return 0;
628 }
629 
__lockup_detector_reconfigure(void)630 static void __lockup_detector_reconfigure(void)
631 {
632 	cpus_read_lock();
633 	watchdog_hardlockup_stop();
634 
635 	softlockup_stop_all();
636 	set_sample_period();
637 	lockup_detector_update_enable();
638 	if (watchdog_enabled && watchdog_thresh)
639 		softlockup_start_all();
640 
641 	watchdog_hardlockup_start();
642 	cpus_read_unlock();
643 	/*
644 	 * Must be called outside the cpus locked section to prevent
645 	 * recursive locking in the perf code.
646 	 */
647 	__lockup_detector_cleanup();
648 }
649 
lockup_detector_reconfigure(void)650 void lockup_detector_reconfigure(void)
651 {
652 	mutex_lock(&watchdog_mutex);
653 	__lockup_detector_reconfigure();
654 	mutex_unlock(&watchdog_mutex);
655 }
656 
657 /*
658  * Create the watchdog infrastructure and configure the detector(s).
659  */
lockup_detector_setup(void)660 static __init void lockup_detector_setup(void)
661 {
662 	/*
663 	 * If sysctl is off and watchdog got disabled on the command line,
664 	 * nothing to do here.
665 	 */
666 	lockup_detector_update_enable();
667 
668 	if (!IS_ENABLED(CONFIG_SYSCTL) &&
669 	    !(watchdog_enabled && watchdog_thresh))
670 		return;
671 
672 	mutex_lock(&watchdog_mutex);
673 	__lockup_detector_reconfigure();
674 	softlockup_initialized = true;
675 	mutex_unlock(&watchdog_mutex);
676 }
677 
678 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
__lockup_detector_reconfigure(void)679 static void __lockup_detector_reconfigure(void)
680 {
681 	cpus_read_lock();
682 	watchdog_hardlockup_stop();
683 	lockup_detector_update_enable();
684 	watchdog_hardlockup_start();
685 	cpus_read_unlock();
686 }
lockup_detector_reconfigure(void)687 void lockup_detector_reconfigure(void)
688 {
689 	__lockup_detector_reconfigure();
690 }
lockup_detector_setup(void)691 static inline void lockup_detector_setup(void)
692 {
693 	__lockup_detector_reconfigure();
694 }
695 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
696 
__lockup_detector_cleanup(void)697 static void __lockup_detector_cleanup(void)
698 {
699 	lockdep_assert_held(&watchdog_mutex);
700 	hardlockup_detector_perf_cleanup();
701 }
702 
703 /**
704  * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
705  *
706  * Caller must not hold the cpu hotplug rwsem.
707  */
lockup_detector_cleanup(void)708 void lockup_detector_cleanup(void)
709 {
710 	mutex_lock(&watchdog_mutex);
711 	__lockup_detector_cleanup();
712 	mutex_unlock(&watchdog_mutex);
713 }
714 
715 /**
716  * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
717  *
718  * Special interface for parisc. It prevents lockup detector warnings from
719  * the default pm_poweroff() function which busy loops forever.
720  */
lockup_detector_soft_poweroff(void)721 void lockup_detector_soft_poweroff(void)
722 {
723 	watchdog_enabled = 0;
724 }
725 
726 #ifdef CONFIG_SYSCTL
727 
728 /* Propagate any changes to the watchdog infrastructure */
proc_watchdog_update(void)729 static void proc_watchdog_update(void)
730 {
731 	/* Remove impossible cpus to keep sysctl output clean. */
732 	cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
733 	__lockup_detector_reconfigure();
734 }
735 
736 /*
737  * common function for watchdog, nmi_watchdog and soft_watchdog parameter
738  *
739  * caller             | table->data points to            | 'which'
740  * -------------------|----------------------------------|-------------------------------
741  * proc_watchdog      | watchdog_user_enabled            | WATCHDOG_HARDLOCKUP_ENABLED |
742  *                    |                                  | WATCHDOG_SOFTOCKUP_ENABLED
743  * -------------------|----------------------------------|-------------------------------
744  * proc_nmi_watchdog  | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED
745  * -------------------|----------------------------------|-------------------------------
746  * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED
747  */
proc_watchdog_common(int which,struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)748 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
749 				void *buffer, size_t *lenp, loff_t *ppos)
750 {
751 	int err, old, *param = table->data;
752 
753 	mutex_lock(&watchdog_mutex);
754 
755 	if (!write) {
756 		/*
757 		 * On read synchronize the userspace interface. This is a
758 		 * racy snapshot.
759 		 */
760 		*param = (watchdog_enabled & which) != 0;
761 		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
762 	} else {
763 		old = READ_ONCE(*param);
764 		err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
765 		if (!err && old != READ_ONCE(*param))
766 			proc_watchdog_update();
767 	}
768 	mutex_unlock(&watchdog_mutex);
769 	return err;
770 }
771 
772 /*
773  * /proc/sys/kernel/watchdog
774  */
proc_watchdog(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)775 int proc_watchdog(struct ctl_table *table, int write,
776 		  void *buffer, size_t *lenp, loff_t *ppos)
777 {
778 	return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED |
779 				    WATCHDOG_SOFTOCKUP_ENABLED,
780 				    table, write, buffer, lenp, ppos);
781 }
782 
783 /*
784  * /proc/sys/kernel/nmi_watchdog
785  */
proc_nmi_watchdog(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)786 int proc_nmi_watchdog(struct ctl_table *table, int write,
787 		      void *buffer, size_t *lenp, loff_t *ppos)
788 {
789 	if (!watchdog_hardlockup_available && write)
790 		return -ENOTSUPP;
791 	return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED,
792 				    table, write, buffer, lenp, ppos);
793 }
794 
795 /*
796  * /proc/sys/kernel/soft_watchdog
797  */
proc_soft_watchdog(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)798 int proc_soft_watchdog(struct ctl_table *table, int write,
799 			void *buffer, size_t *lenp, loff_t *ppos)
800 {
801 	return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED,
802 				    table, write, buffer, lenp, ppos);
803 }
804 
805 /*
806  * /proc/sys/kernel/watchdog_thresh
807  */
proc_watchdog_thresh(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)808 int proc_watchdog_thresh(struct ctl_table *table, int write,
809 			 void *buffer, size_t *lenp, loff_t *ppos)
810 {
811 	int err, old;
812 
813 	mutex_lock(&watchdog_mutex);
814 
815 	old = READ_ONCE(watchdog_thresh);
816 	err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
817 
818 	if (!err && write && old != READ_ONCE(watchdog_thresh))
819 		proc_watchdog_update();
820 
821 	mutex_unlock(&watchdog_mutex);
822 	return err;
823 }
824 
825 /*
826  * The cpumask is the mask of possible cpus that the watchdog can run
827  * on, not the mask of cpus it is actually running on.  This allows the
828  * user to specify a mask that will include cpus that have not yet
829  * been brought online, if desired.
830  */
proc_watchdog_cpumask(struct ctl_table * table,int write,void * buffer,size_t * lenp,loff_t * ppos)831 int proc_watchdog_cpumask(struct ctl_table *table, int write,
832 			  void *buffer, size_t *lenp, loff_t *ppos)
833 {
834 	int err;
835 
836 	mutex_lock(&watchdog_mutex);
837 
838 	err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
839 	if (!err && write)
840 		proc_watchdog_update();
841 
842 	mutex_unlock(&watchdog_mutex);
843 	return err;
844 }
845 
846 static const int sixty = 60;
847 
848 static struct ctl_table watchdog_sysctls[] = {
849 	{
850 		.procname       = "watchdog",
851 		.data		= &watchdog_user_enabled,
852 		.maxlen		= sizeof(int),
853 		.mode		= 0644,
854 		.proc_handler   = proc_watchdog,
855 		.extra1		= SYSCTL_ZERO,
856 		.extra2		= SYSCTL_ONE,
857 	},
858 	{
859 		.procname	= "watchdog_thresh",
860 		.data		= &watchdog_thresh,
861 		.maxlen		= sizeof(int),
862 		.mode		= 0644,
863 		.proc_handler	= proc_watchdog_thresh,
864 		.extra1		= SYSCTL_ZERO,
865 		.extra2		= (void *)&sixty,
866 	},
867 	{
868 		.procname	= "watchdog_cpumask",
869 		.data		= &watchdog_cpumask_bits,
870 		.maxlen		= NR_CPUS,
871 		.mode		= 0644,
872 		.proc_handler	= proc_watchdog_cpumask,
873 	},
874 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
875 	{
876 		.procname       = "soft_watchdog",
877 		.data		= &watchdog_softlockup_user_enabled,
878 		.maxlen		= sizeof(int),
879 		.mode		= 0644,
880 		.proc_handler   = proc_soft_watchdog,
881 		.extra1		= SYSCTL_ZERO,
882 		.extra2		= SYSCTL_ONE,
883 	},
884 	{
885 		.procname	= "softlockup_panic",
886 		.data		= &softlockup_panic,
887 		.maxlen		= sizeof(int),
888 		.mode		= 0644,
889 		.proc_handler	= proc_dointvec_minmax,
890 		.extra1		= SYSCTL_ZERO,
891 		.extra2		= SYSCTL_ONE,
892 	},
893 #ifdef CONFIG_SMP
894 	{
895 		.procname	= "softlockup_all_cpu_backtrace",
896 		.data		= &sysctl_softlockup_all_cpu_backtrace,
897 		.maxlen		= sizeof(int),
898 		.mode		= 0644,
899 		.proc_handler	= proc_dointvec_minmax,
900 		.extra1		= SYSCTL_ZERO,
901 		.extra2		= SYSCTL_ONE,
902 	},
903 #endif /* CONFIG_SMP */
904 #endif
905 #ifdef CONFIG_HARDLOCKUP_DETECTOR
906 	{
907 		.procname	= "hardlockup_panic",
908 		.data		= &hardlockup_panic,
909 		.maxlen		= sizeof(int),
910 		.mode		= 0644,
911 		.proc_handler	= proc_dointvec_minmax,
912 		.extra1		= SYSCTL_ZERO,
913 		.extra2		= SYSCTL_ONE,
914 	},
915 #ifdef CONFIG_SMP
916 	{
917 		.procname	= "hardlockup_all_cpu_backtrace",
918 		.data		= &sysctl_hardlockup_all_cpu_backtrace,
919 		.maxlen		= sizeof(int),
920 		.mode		= 0644,
921 		.proc_handler	= proc_dointvec_minmax,
922 		.extra1		= SYSCTL_ZERO,
923 		.extra2		= SYSCTL_ONE,
924 	},
925 #endif /* CONFIG_SMP */
926 #endif
927 	{}
928 };
929 
930 static struct ctl_table watchdog_hardlockup_sysctl[] = {
931 	{
932 		.procname       = "nmi_watchdog",
933 		.data		= &watchdog_hardlockup_user_enabled,
934 		.maxlen		= sizeof(int),
935 		.mode		= 0444,
936 		.proc_handler   = proc_nmi_watchdog,
937 		.extra1		= SYSCTL_ZERO,
938 		.extra2		= SYSCTL_ONE,
939 	},
940 	{}
941 };
942 
watchdog_sysctl_init(void)943 static void __init watchdog_sysctl_init(void)
944 {
945 	register_sysctl_init("kernel", watchdog_sysctls);
946 
947 	if (watchdog_hardlockup_available)
948 		watchdog_hardlockup_sysctl[0].mode = 0644;
949 	register_sysctl_init("kernel", watchdog_hardlockup_sysctl);
950 }
951 
952 #else
953 #define watchdog_sysctl_init() do { } while (0)
954 #endif /* CONFIG_SYSCTL */
955 
956 static void __init lockup_detector_delay_init(struct work_struct *work);
957 static bool allow_lockup_detector_init_retry __initdata;
958 
959 static struct work_struct detector_work __initdata =
960 		__WORK_INITIALIZER(detector_work, lockup_detector_delay_init);
961 
lockup_detector_delay_init(struct work_struct * work)962 static void __init lockup_detector_delay_init(struct work_struct *work)
963 {
964 	int ret;
965 
966 	ret = watchdog_hardlockup_probe();
967 	if (ret) {
968 		pr_info("Delayed init of the lockup detector failed: %d\n", ret);
969 		pr_info("Hard watchdog permanently disabled\n");
970 		return;
971 	}
972 
973 	allow_lockup_detector_init_retry = false;
974 
975 	watchdog_hardlockup_available = true;
976 	lockup_detector_setup();
977 }
978 
979 /*
980  * lockup_detector_retry_init - retry init lockup detector if possible.
981  *
982  * Retry hardlockup detector init. It is useful when it requires some
983  * functionality that has to be initialized later on a particular
984  * platform.
985  */
lockup_detector_retry_init(void)986 void __init lockup_detector_retry_init(void)
987 {
988 	/* Must be called before late init calls */
989 	if (!allow_lockup_detector_init_retry)
990 		return;
991 
992 	schedule_work(&detector_work);
993 }
994 
995 /*
996  * Ensure that optional delayed hardlockup init is proceed before
997  * the init code and memory is freed.
998  */
lockup_detector_check(void)999 static int __init lockup_detector_check(void)
1000 {
1001 	/* Prevent any later retry. */
1002 	allow_lockup_detector_init_retry = false;
1003 
1004 	/* Make sure no work is pending. */
1005 	flush_work(&detector_work);
1006 
1007 	watchdog_sysctl_init();
1008 
1009 	return 0;
1010 
1011 }
1012 late_initcall_sync(lockup_detector_check);
1013 
lockup_detector_init(void)1014 void __init lockup_detector_init(void)
1015 {
1016 	if (tick_nohz_full_enabled())
1017 		pr_info("Disabling watchdog on nohz_full cores by default\n");
1018 
1019 	cpumask_copy(&watchdog_cpumask,
1020 		     housekeeping_cpumask(HK_TYPE_TIMER));
1021 
1022 	if (!watchdog_hardlockup_probe())
1023 		watchdog_hardlockup_available = true;
1024 	else
1025 		allow_lockup_detector_init_retry = true;
1026 
1027 	lockup_detector_setup();
1028 }
1029