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/delay.h>
19 #include <linux/freezer.h>
20 #include <linux/kthread.h>
21 #include <linux/lockdep.h>
22 #include <linux/notifier.h>
23 #include <linux/module.h>
24 #include <linux/sysctl.h>
25 
26 #include <asm/irq_regs.h>
27 #include <linux/perf_event.h>
28 
29 int watchdog_enabled = 1;
30 int __read_mostly watchdog_thresh = 10;
31 
32 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
33 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
34 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
35 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
36 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
37 #ifdef CONFIG_HARDLOCKUP_DETECTOR
38 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
39 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
40 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
41 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
42 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
43 #endif
44 
45 /* boot commands */
46 /*
47  * Should we panic when a soft-lockup or hard-lockup occurs:
48  */
49 #ifdef CONFIG_HARDLOCKUP_DETECTOR
50 static int hardlockup_panic =
51 			CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
52 
hardlockup_panic_setup(char * str)53 static int __init hardlockup_panic_setup(char *str)
54 {
55 	if (!strncmp(str, "panic", 5))
56 		hardlockup_panic = 1;
57 	else if (!strncmp(str, "nopanic", 7))
58 		hardlockup_panic = 0;
59 	else if (!strncmp(str, "0", 1))
60 		watchdog_enabled = 0;
61 	return 1;
62 }
63 __setup("nmi_watchdog=", hardlockup_panic_setup);
64 #endif
65 
66 unsigned int __read_mostly softlockup_panic =
67 			CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
68 
softlockup_panic_setup(char * str)69 static int __init softlockup_panic_setup(char *str)
70 {
71 	softlockup_panic = simple_strtoul(str, NULL, 0);
72 
73 	return 1;
74 }
75 __setup("softlockup_panic=", softlockup_panic_setup);
76 
nowatchdog_setup(char * str)77 static int __init nowatchdog_setup(char *str)
78 {
79 	watchdog_enabled = 0;
80 	return 1;
81 }
82 __setup("nowatchdog", nowatchdog_setup);
83 
84 /* deprecated */
nosoftlockup_setup(char * str)85 static int __init nosoftlockup_setup(char *str)
86 {
87 	watchdog_enabled = 0;
88 	return 1;
89 }
90 __setup("nosoftlockup", nosoftlockup_setup);
91 /*  */
92 
93 /*
94  * Hard-lockup warnings should be triggered after just a few seconds. Soft-
95  * lockups can have false positives under extreme conditions. So we generally
96  * want a higher threshold for soft lockups than for hard lockups. So we couple
97  * the thresholds with a factor: we make the soft threshold twice the amount of
98  * time the hard threshold is.
99  */
get_softlockup_thresh(void)100 static int get_softlockup_thresh(void)
101 {
102 	return watchdog_thresh * 2;
103 }
104 
105 /*
106  * Returns seconds, approximately.  We don't need nanosecond
107  * resolution, and we don't need to waste time with a big divide when
108  * 2^30ns == 1.074s.
109  */
get_timestamp(int this_cpu)110 static unsigned long get_timestamp(int this_cpu)
111 {
112 	return cpu_clock(this_cpu) >> 30LL;  /* 2^30 ~= 10^9 */
113 }
114 
get_sample_period(void)115 static u64 get_sample_period(void)
116 {
117 	/*
118 	 * convert watchdog_thresh from seconds to ns
119 	 * the divide by 5 is to give hrtimer several chances (two
120 	 * or three with the current relation between the soft
121 	 * and hard thresholds) to increment before the
122 	 * hardlockup detector generates a warning
123 	 */
124 	return get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
125 }
126 
127 /* Commands for resetting the watchdog */
__touch_watchdog(void)128 static void __touch_watchdog(void)
129 {
130 	int this_cpu = smp_processor_id();
131 
132 	__this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
133 }
134 
touch_softlockup_watchdog(void)135 void touch_softlockup_watchdog(void)
136 {
137 	__this_cpu_write(watchdog_touch_ts, 0);
138 }
139 EXPORT_SYMBOL(touch_softlockup_watchdog);
140 
touch_all_softlockup_watchdogs(void)141 void touch_all_softlockup_watchdogs(void)
142 {
143 	int cpu;
144 
145 	/*
146 	 * this is done lockless
147 	 * do we care if a 0 races with a timestamp?
148 	 * all it means is the softlock check starts one cycle later
149 	 */
150 	for_each_online_cpu(cpu)
151 		per_cpu(watchdog_touch_ts, cpu) = 0;
152 }
153 
154 #ifdef CONFIG_HARDLOCKUP_DETECTOR
touch_nmi_watchdog(void)155 void touch_nmi_watchdog(void)
156 {
157 	if (watchdog_enabled) {
158 		unsigned cpu;
159 
160 		for_each_present_cpu(cpu) {
161 			if (per_cpu(watchdog_nmi_touch, cpu) != true)
162 				per_cpu(watchdog_nmi_touch, cpu) = true;
163 		}
164 	}
165 	touch_softlockup_watchdog();
166 }
167 EXPORT_SYMBOL(touch_nmi_watchdog);
168 
169 #endif
170 
touch_softlockup_watchdog_sync(void)171 void touch_softlockup_watchdog_sync(void)
172 {
173 	__raw_get_cpu_var(softlockup_touch_sync) = true;
174 	__raw_get_cpu_var(watchdog_touch_ts) = 0;
175 }
176 
177 #ifdef CONFIG_HARDLOCKUP_DETECTOR
178 /* watchdog detector functions */
is_hardlockup(void)179 static int is_hardlockup(void)
180 {
181 	unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
182 
183 	if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
184 		return 1;
185 
186 	__this_cpu_write(hrtimer_interrupts_saved, hrint);
187 	return 0;
188 }
189 #endif
190 
is_softlockup(unsigned long touch_ts)191 static int is_softlockup(unsigned long touch_ts)
192 {
193 	unsigned long now = get_timestamp(smp_processor_id());
194 
195 	/* Warn about unreasonable delays: */
196 	if (time_after(now, touch_ts + get_softlockup_thresh()))
197 		return now - touch_ts;
198 
199 	return 0;
200 }
201 
202 #ifdef CONFIG_HARDLOCKUP_DETECTOR
203 
204 static struct perf_event_attr wd_hw_attr = {
205 	.type		= PERF_TYPE_HARDWARE,
206 	.config		= PERF_COUNT_HW_CPU_CYCLES,
207 	.size		= sizeof(struct perf_event_attr),
208 	.pinned		= 1,
209 	.disabled	= 1,
210 };
211 
212 /* Callback function for perf event subsystem */
watchdog_overflow_callback(struct perf_event * event,struct perf_sample_data * data,struct pt_regs * regs)213 static void watchdog_overflow_callback(struct perf_event *event,
214 		 struct perf_sample_data *data,
215 		 struct pt_regs *regs)
216 {
217 	/* Ensure the watchdog never gets throttled */
218 	event->hw.interrupts = 0;
219 
220 	if (__this_cpu_read(watchdog_nmi_touch) == true) {
221 		__this_cpu_write(watchdog_nmi_touch, false);
222 		return;
223 	}
224 
225 	/* check for a hardlockup
226 	 * This is done by making sure our timer interrupt
227 	 * is incrementing.  The timer interrupt should have
228 	 * fired multiple times before we overflow'd.  If it hasn't
229 	 * then this is a good indication the cpu is stuck
230 	 */
231 	if (is_hardlockup()) {
232 		int this_cpu = smp_processor_id();
233 
234 		/* only print hardlockups once */
235 		if (__this_cpu_read(hard_watchdog_warn) == true)
236 			return;
237 
238 		if (hardlockup_panic)
239 			panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
240 		else
241 			WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
242 
243 		__this_cpu_write(hard_watchdog_warn, true);
244 		return;
245 	}
246 
247 	__this_cpu_write(hard_watchdog_warn, false);
248 	return;
249 }
watchdog_interrupt_count(void)250 static void watchdog_interrupt_count(void)
251 {
252 	__this_cpu_inc(hrtimer_interrupts);
253 }
254 #else
watchdog_interrupt_count(void)255 static inline void watchdog_interrupt_count(void) { return; }
256 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
257 
258 /* watchdog kicker functions */
watchdog_timer_fn(struct hrtimer * hrtimer)259 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
260 {
261 	unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
262 	struct pt_regs *regs = get_irq_regs();
263 	int duration;
264 
265 	/* kick the hardlockup detector */
266 	watchdog_interrupt_count();
267 
268 	/* kick the softlockup detector */
269 	wake_up_process(__this_cpu_read(softlockup_watchdog));
270 
271 	/* .. and repeat */
272 	hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
273 
274 	if (touch_ts == 0) {
275 		if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
276 			/*
277 			 * If the time stamp was touched atomically
278 			 * make sure the scheduler tick is up to date.
279 			 */
280 			__this_cpu_write(softlockup_touch_sync, false);
281 			sched_clock_tick();
282 		}
283 		__touch_watchdog();
284 		return HRTIMER_RESTART;
285 	}
286 
287 	/* check for a softlockup
288 	 * This is done by making sure a high priority task is
289 	 * being scheduled.  The task touches the watchdog to
290 	 * indicate it is getting cpu time.  If it hasn't then
291 	 * this is a good indication some task is hogging the cpu
292 	 */
293 	duration = is_softlockup(touch_ts);
294 	if (unlikely(duration)) {
295 		/* only warn once */
296 		if (__this_cpu_read(soft_watchdog_warn) == true)
297 			return HRTIMER_RESTART;
298 
299 		printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
300 			smp_processor_id(), duration,
301 			current->comm, task_pid_nr(current));
302 		print_modules();
303 		print_irqtrace_events(current);
304 		if (regs)
305 			show_regs(regs);
306 		else
307 			dump_stack();
308 
309 		if (softlockup_panic)
310 			panic("softlockup: hung tasks");
311 		__this_cpu_write(soft_watchdog_warn, true);
312 	} else
313 		__this_cpu_write(soft_watchdog_warn, false);
314 
315 	return HRTIMER_RESTART;
316 }
317 
318 
319 /*
320  * The watchdog thread - touches the timestamp.
321  */
watchdog(void * unused)322 static int watchdog(void *unused)
323 {
324 	struct sched_param param = { .sched_priority = 0 };
325 	struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
326 
327 	/* initialize timestamp */
328 	__touch_watchdog();
329 
330 	/* kick off the timer for the hardlockup detector */
331 	/* done here because hrtimer_start can only pin to smp_processor_id() */
332 	hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
333 		      HRTIMER_MODE_REL_PINNED);
334 
335 	set_current_state(TASK_INTERRUPTIBLE);
336 	/*
337 	 * Run briefly (kicked by the hrtimer callback function) once every
338 	 * get_sample_period() seconds (4 seconds by default) to reset the
339 	 * softlockup timestamp. If this gets delayed for more than
340 	 * 2*watchdog_thresh seconds then the debug-printout triggers in
341 	 * watchdog_timer_fn().
342 	 */
343 	while (!kthread_should_stop()) {
344 		__touch_watchdog();
345 		schedule();
346 
347 		if (kthread_should_stop())
348 			break;
349 
350 		set_current_state(TASK_INTERRUPTIBLE);
351 	}
352 	/*
353 	 * Drop the policy/priority elevation during thread exit to avoid a
354 	 * scheduling latency spike.
355 	 */
356 	__set_current_state(TASK_RUNNING);
357 	sched_setscheduler(current, SCHED_NORMAL, &param);
358 	return 0;
359 }
360 
361 
362 #ifdef CONFIG_HARDLOCKUP_DETECTOR
watchdog_nmi_enable(int cpu)363 static int watchdog_nmi_enable(int cpu)
364 {
365 	struct perf_event_attr *wd_attr;
366 	struct perf_event *event = per_cpu(watchdog_ev, cpu);
367 
368 	/* is it already setup and enabled? */
369 	if (event && event->state > PERF_EVENT_STATE_OFF)
370 		goto out;
371 
372 	/* it is setup but not enabled */
373 	if (event != NULL)
374 		goto out_enable;
375 
376 	wd_attr = &wd_hw_attr;
377 	wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
378 
379 	/* Try to register using hardware perf events */
380 	event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
381 	if (!IS_ERR(event)) {
382 		pr_info("enabled, takes one hw-pmu counter.\n");
383 		goto out_save;
384 	}
385 
386 
387 	/* vary the KERN level based on the returned errno */
388 	if (PTR_ERR(event) == -EOPNOTSUPP)
389 		pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
390 	else if (PTR_ERR(event) == -ENOENT)
391 		pr_warning("disabled (cpu%i): hardware events not enabled\n",
392 			 cpu);
393 	else
394 		pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
395 			cpu, PTR_ERR(event));
396 	return PTR_ERR(event);
397 
398 	/* success path */
399 out_save:
400 	per_cpu(watchdog_ev, cpu) = event;
401 out_enable:
402 	perf_event_enable(per_cpu(watchdog_ev, cpu));
403 out:
404 	return 0;
405 }
406 
watchdog_nmi_disable(int cpu)407 static void watchdog_nmi_disable(int cpu)
408 {
409 	struct perf_event *event = per_cpu(watchdog_ev, cpu);
410 
411 	if (event) {
412 		perf_event_disable(event);
413 		per_cpu(watchdog_ev, cpu) = NULL;
414 
415 		/* should be in cleanup, but blocks oprofile */
416 		perf_event_release_kernel(event);
417 	}
418 	return;
419 }
420 #else
watchdog_nmi_enable(int cpu)421 static int watchdog_nmi_enable(int cpu) { return 0; }
watchdog_nmi_disable(int cpu)422 static void watchdog_nmi_disable(int cpu) { return; }
423 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
424 
425 /* prepare/enable/disable routines */
watchdog_prepare_cpu(int cpu)426 static void watchdog_prepare_cpu(int cpu)
427 {
428 	struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
429 
430 	WARN_ON(per_cpu(softlockup_watchdog, cpu));
431 	hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
432 	hrtimer->function = watchdog_timer_fn;
433 }
434 
watchdog_enable(int cpu)435 static int watchdog_enable(int cpu)
436 {
437 	struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
438 	int err = 0;
439 
440 	/* enable the perf event */
441 	err = watchdog_nmi_enable(cpu);
442 
443 	/* Regardless of err above, fall through and start softlockup */
444 
445 	/* create the watchdog thread */
446 	if (!p) {
447 		struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
448 		p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
449 		if (IS_ERR(p)) {
450 			pr_err("softlockup watchdog for %i failed\n", cpu);
451 			if (!err) {
452 				/* if hardlockup hasn't already set this */
453 				err = PTR_ERR(p);
454 				/* and disable the perf event */
455 				watchdog_nmi_disable(cpu);
456 			}
457 			goto out;
458 		}
459 		sched_setscheduler(p, SCHED_FIFO, &param);
460 		kthread_bind(p, cpu);
461 		per_cpu(watchdog_touch_ts, cpu) = 0;
462 		per_cpu(softlockup_watchdog, cpu) = p;
463 		wake_up_process(p);
464 	}
465 
466 out:
467 	return err;
468 }
469 
watchdog_disable(int cpu)470 static void watchdog_disable(int cpu)
471 {
472 	struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
473 	struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
474 
475 	/*
476 	 * cancel the timer first to stop incrementing the stats
477 	 * and waking up the kthread
478 	 */
479 	hrtimer_cancel(hrtimer);
480 
481 	/* disable the perf event */
482 	watchdog_nmi_disable(cpu);
483 
484 	/* stop the watchdog thread */
485 	if (p) {
486 		per_cpu(softlockup_watchdog, cpu) = NULL;
487 		kthread_stop(p);
488 	}
489 }
490 
491 /* sysctl functions */
492 #ifdef CONFIG_SYSCTL
watchdog_enable_all_cpus(void)493 static void watchdog_enable_all_cpus(void)
494 {
495 	int cpu;
496 
497 	watchdog_enabled = 0;
498 
499 	for_each_online_cpu(cpu)
500 		if (!watchdog_enable(cpu))
501 			/* if any cpu succeeds, watchdog is considered
502 			   enabled for the system */
503 			watchdog_enabled = 1;
504 
505 	if (!watchdog_enabled)
506 		pr_err("failed to be enabled on some cpus\n");
507 
508 }
509 
watchdog_disable_all_cpus(void)510 static void watchdog_disable_all_cpus(void)
511 {
512 	int cpu;
513 
514 	for_each_online_cpu(cpu)
515 		watchdog_disable(cpu);
516 
517 	/* if all watchdogs are disabled, then they are disabled for the system */
518 	watchdog_enabled = 0;
519 }
520 
521 
522 /*
523  * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
524  */
525 
proc_dowatchdog(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)526 int proc_dowatchdog(struct ctl_table *table, int write,
527 		    void __user *buffer, size_t *lenp, loff_t *ppos)
528 {
529 	int ret;
530 
531 	ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
532 	if (ret || !write)
533 		goto out;
534 
535 	if (watchdog_enabled && watchdog_thresh)
536 		watchdog_enable_all_cpus();
537 	else
538 		watchdog_disable_all_cpus();
539 
540 out:
541 	return ret;
542 }
543 #endif /* CONFIG_SYSCTL */
544 
545 
546 /*
547  * Create/destroy watchdog threads as CPUs come and go:
548  */
549 static int __cpuinit
cpu_callback(struct notifier_block * nfb,unsigned long action,void * hcpu)550 cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
551 {
552 	int hotcpu = (unsigned long)hcpu;
553 
554 	switch (action) {
555 	case CPU_UP_PREPARE:
556 	case CPU_UP_PREPARE_FROZEN:
557 		watchdog_prepare_cpu(hotcpu);
558 		break;
559 	case CPU_ONLINE:
560 	case CPU_ONLINE_FROZEN:
561 		if (watchdog_enabled)
562 			watchdog_enable(hotcpu);
563 		break;
564 #ifdef CONFIG_HOTPLUG_CPU
565 	case CPU_UP_CANCELED:
566 	case CPU_UP_CANCELED_FROZEN:
567 		watchdog_disable(hotcpu);
568 		break;
569 	case CPU_DEAD:
570 	case CPU_DEAD_FROZEN:
571 		watchdog_disable(hotcpu);
572 		break;
573 #endif /* CONFIG_HOTPLUG_CPU */
574 	}
575 
576 	/*
577 	 * hardlockup and softlockup are not important enough
578 	 * to block cpu bring up.  Just always succeed and
579 	 * rely on printk output to flag problems.
580 	 */
581 	return NOTIFY_OK;
582 }
583 
584 static struct notifier_block __cpuinitdata cpu_nfb = {
585 	.notifier_call = cpu_callback
586 };
587 
lockup_detector_init(void)588 void __init lockup_detector_init(void)
589 {
590 	void *cpu = (void *)(long)smp_processor_id();
591 	int err;
592 
593 	err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
594 	WARN_ON(notifier_to_errno(err));
595 
596 	cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
597 	register_cpu_notifier(&cpu_nfb);
598 
599 	return;
600 }
601