1
2
3
4
5
6
7
8
9
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
29int watchdog_enabled = 1;
30int __read_mostly watchdog_thresh = 10;
31
32static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
33static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
34static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
35static DEFINE_PER_CPU(bool, softlockup_touch_sync);
36static DEFINE_PER_CPU(bool, soft_watchdog_warn);
37#ifdef CONFIG_HARDLOCKUP_DETECTOR
38static DEFINE_PER_CPU(bool, hard_watchdog_warn);
39static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
40static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
41static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
42static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
43#endif
44
45
46
47
48
49#ifdef CONFIG_HARDLOCKUP_DETECTOR
50static int hardlockup_panic =
51 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
52
53static 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
66unsigned int __read_mostly softlockup_panic =
67 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
68
69static 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
77static int __init nowatchdog_setup(char *str)
78{
79 watchdog_enabled = 0;
80 return 1;
81}
82__setup("nowatchdog", nowatchdog_setup);
83
84
85static int __init nosoftlockup_setup(char *str)
86{
87 watchdog_enabled = 0;
88 return 1;
89}
90__setup("nosoftlockup", nosoftlockup_setup);
91
92
93
94
95
96
97
98
99
100static int get_softlockup_thresh(void)
101{
102 return watchdog_thresh * 2;
103}
104
105
106
107
108
109
110static unsigned long get_timestamp(int this_cpu)
111{
112 return cpu_clock(this_cpu) >> 30LL;
113}
114
115static unsigned long get_sample_period(void)
116{
117
118
119
120
121
122
123
124 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
125}
126
127
128static 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
135void touch_softlockup_watchdog(void)
136{
137 __this_cpu_write(watchdog_touch_ts, 0);
138}
139EXPORT_SYMBOL(touch_softlockup_watchdog);
140
141void touch_all_softlockup_watchdogs(void)
142{
143 int cpu;
144
145
146
147
148
149
150 for_each_online_cpu(cpu)
151 per_cpu(watchdog_touch_ts, cpu) = 0;
152}
153
154#ifdef CONFIG_HARDLOCKUP_DETECTOR
155void 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}
167EXPORT_SYMBOL(touch_nmi_watchdog);
168
169#endif
170
171void 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
179static 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
191static int is_softlockup(unsigned long touch_ts)
192{
193 unsigned long now = get_timestamp(smp_processor_id());
194
195
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
204static 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
213static void watchdog_overflow_callback(struct perf_event *event,
214 struct perf_sample_data *data,
215 struct pt_regs *regs)
216{
217
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
226
227
228
229
230
231 if (is_hardlockup()) {
232 int this_cpu = smp_processor_id();
233
234
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}
250static void watchdog_interrupt_count(void)
251{
252 __this_cpu_inc(hrtimer_interrupts);
253}
254#else
255static inline void watchdog_interrupt_count(void) { return; }
256#endif
257
258
259static 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
266 watchdog_interrupt_count();
267
268
269 wake_up_process(__this_cpu_read(softlockup_watchdog));
270
271
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
278
279
280 __this_cpu_write(softlockup_touch_sync, false);
281 sched_clock_tick();
282 }
283 __touch_watchdog();
284 return HRTIMER_RESTART;
285 }
286
287
288
289
290
291
292
293 duration = is_softlockup(touch_ts);
294 if (unlikely(duration)) {
295
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
321
322static 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
328 __touch_watchdog();
329
330
331
332 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
333 HRTIMER_MODE_REL_PINNED);
334
335 set_current_state(TASK_INTERRUPTIBLE);
336
337
338
339
340
341
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
354
355
356 __set_current_state(TASK_RUNNING);
357 sched_setscheduler(current, SCHED_NORMAL, ¶m);
358 return 0;
359}
360
361
362#ifdef CONFIG_HARDLOCKUP_DETECTOR
363static 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
369 if (event && event->state > PERF_EVENT_STATE_OFF)
370 goto out;
371
372
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
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
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
399out_save:
400 per_cpu(watchdog_ev, cpu) = event;
401out_enable:
402 perf_event_enable(per_cpu(watchdog_ev, cpu));
403out:
404 return 0;
405}
406
407static 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
416 perf_event_release_kernel(event);
417 }
418 return;
419}
420#else
421static int watchdog_nmi_enable(int cpu) { return 0; }
422static void watchdog_nmi_disable(int cpu) { return; }
423#endif
424
425
426static 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
435static int watchdog_enable(int cpu)
436{
437 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
438 int err = 0;
439
440
441 err = watchdog_nmi_enable(cpu);
442
443
444
445
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
453 err = PTR_ERR(p);
454
455 watchdog_nmi_disable(cpu);
456 }
457 goto out;
458 }
459 sched_setscheduler(p, SCHED_FIFO, ¶m);
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
466out:
467 return err;
468}
469
470static 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
477
478
479 hrtimer_cancel(hrtimer);
480
481
482 watchdog_nmi_disable(cpu);
483
484
485 if (p) {
486 per_cpu(softlockup_watchdog, cpu) = NULL;
487 kthread_stop(p);
488 }
489}
490
491
492#ifdef CONFIG_SYSCTL
493static 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
502
503 watchdog_enabled = 1;
504
505 if (!watchdog_enabled)
506 pr_err("failed to be enabled on some cpus\n");
507
508}
509
510static void watchdog_disable_all_cpus(void)
511{
512 int cpu;
513
514 for_each_online_cpu(cpu)
515 watchdog_disable(cpu);
516
517
518 watchdog_enabled = 0;
519}
520
521
522
523
524
525
526int 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
540out:
541 return ret;
542}
543#endif
544
545
546
547
548
549static int __cpuinit
550cpu_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
574 }
575
576
577
578
579
580
581 return NOTIFY_OK;
582}
583
584static struct notifier_block __cpuinitdata cpu_nfb = {
585 .notifier_call = cpu_callback
586};
587
588void __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