1
2
3
4
5
6
7
8
9
10
11
12
13
14
15#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/delay.h>
20#include <linux/freezer.h>
21#include <linux/kthread.h>
22#include <linux/lockdep.h>
23#include <linux/notifier.h>
24#include <linux/module.h>
25#include <linux/sysctl.h>
26
27#include <asm/irq_regs.h>
28#include <linux/perf_event.h>
29
30int watchdog_enabled = 1;
31int __read_mostly watchdog_thresh = 10;
32
33static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
34static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
35static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
36static DEFINE_PER_CPU(bool, softlockup_touch_sync);
37static DEFINE_PER_CPU(bool, soft_watchdog_warn);
38#ifdef CONFIG_HARDLOCKUP_DETECTOR
39static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
41static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
42static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
43static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
44#endif
45
46
47
48
49
50#ifdef CONFIG_HARDLOCKUP_DETECTOR
51static int hardlockup_panic =
52 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
53
54static int __init hardlockup_panic_setup(char *str)
55{
56 if (!strncmp(str, "panic", 5))
57 hardlockup_panic = 1;
58 else if (!strncmp(str, "nopanic", 7))
59 hardlockup_panic = 0;
60 else if (!strncmp(str, "0", 1))
61 watchdog_enabled = 0;
62 return 1;
63}
64__setup("nmi_watchdog=", hardlockup_panic_setup);
65#endif
66
67unsigned int __read_mostly softlockup_panic =
68 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
69
70static int __init softlockup_panic_setup(char *str)
71{
72 softlockup_panic = simple_strtoul(str, NULL, 0);
73
74 return 1;
75}
76__setup("softlockup_panic=", softlockup_panic_setup);
77
78static int __init nowatchdog_setup(char *str)
79{
80 watchdog_enabled = 0;
81 return 1;
82}
83__setup("nowatchdog", nowatchdog_setup);
84
85
86static int __init nosoftlockup_setup(char *str)
87{
88 watchdog_enabled = 0;
89 return 1;
90}
91__setup("nosoftlockup", nosoftlockup_setup);
92
93
94
95
96
97
98
99
100
101static int get_softlockup_thresh(void)
102{
103 return watchdog_thresh * 2;
104}
105
106
107
108
109
110
111static unsigned long get_timestamp(int this_cpu)
112{
113 return cpu_clock(this_cpu) >> 30LL;
114}
115
116static unsigned long get_sample_period(void)
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 = MAX_RT_PRIO-1 };
325 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
326
327 sched_setscheduler(current, SCHED_FIFO, ¶m);
328
329
330 __touch_watchdog();
331
332
333
334 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
335 HRTIMER_MODE_REL_PINNED);
336
337 set_current_state(TASK_INTERRUPTIBLE);
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 __set_current_state(TASK_RUNNING);
353 param.sched_priority = 0;
354 sched_setscheduler(current, SCHED_NORMAL, ¶m);
355 return 0;
356}
357
358
359#ifdef CONFIG_HARDLOCKUP_DETECTOR
360static int watchdog_nmi_enable(int cpu)
361{
362 struct perf_event_attr *wd_attr;
363 struct perf_event *event = per_cpu(watchdog_ev, cpu);
364
365
366 if (event && event->state > PERF_EVENT_STATE_OFF)
367 goto out;
368
369
370 if (event != NULL)
371 goto out_enable;
372
373 wd_attr = &wd_hw_attr;
374 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
375
376
377 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
378 if (!IS_ERR(event)) {
379 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
380 goto out_save;
381 }
382
383
384
385 if (PTR_ERR(event) == -EOPNOTSUPP)
386 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
387 else if (PTR_ERR(event) == -ENOENT)
388 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
389 else
390 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
391 return PTR_ERR(event);
392
393
394out_save:
395 per_cpu(watchdog_ev, cpu) = event;
396out_enable:
397 perf_event_enable(per_cpu(watchdog_ev, cpu));
398out:
399 return 0;
400}
401
402static void watchdog_nmi_disable(int cpu)
403{
404 struct perf_event *event = per_cpu(watchdog_ev, cpu);
405
406 if (event) {
407 perf_event_disable(event);
408 per_cpu(watchdog_ev, cpu) = NULL;
409
410
411 perf_event_release_kernel(event);
412 }
413 return;
414}
415#else
416static int watchdog_nmi_enable(int cpu) { return 0; }
417static void watchdog_nmi_disable(int cpu) { return; }
418#endif
419
420
421static void watchdog_prepare_cpu(int cpu)
422{
423 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
424
425 WARN_ON(per_cpu(softlockup_watchdog, cpu));
426 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
427 hrtimer->function = watchdog_timer_fn;
428}
429
430static int watchdog_enable(int cpu)
431{
432 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
433 int err = 0;
434
435
436 err = watchdog_nmi_enable(cpu);
437
438
439
440
441 if (!p) {
442 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
443 if (IS_ERR(p)) {
444 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
445 if (!err) {
446
447 err = PTR_ERR(p);
448
449 watchdog_nmi_disable(cpu);
450 }
451 goto out;
452 }
453 kthread_bind(p, cpu);
454 per_cpu(watchdog_touch_ts, cpu) = 0;
455 per_cpu(softlockup_watchdog, cpu) = p;
456 wake_up_process(p);
457 }
458
459out:
460 return err;
461}
462
463static void watchdog_disable(int cpu)
464{
465 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
466 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
467
468
469
470
471
472 hrtimer_cancel(hrtimer);
473
474
475 watchdog_nmi_disable(cpu);
476
477
478 if (p) {
479 per_cpu(softlockup_watchdog, cpu) = NULL;
480 kthread_stop(p);
481 }
482}
483
484
485#ifdef CONFIG_SYSCTL
486static void watchdog_enable_all_cpus(void)
487{
488 int cpu;
489
490 watchdog_enabled = 0;
491
492 for_each_online_cpu(cpu)
493 if (!watchdog_enable(cpu))
494
495
496 watchdog_enabled = 1;
497
498 if (!watchdog_enabled)
499 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
500
501}
502
503static void watchdog_disable_all_cpus(void)
504{
505 int cpu;
506
507 for_each_online_cpu(cpu)
508 watchdog_disable(cpu);
509
510
511 watchdog_enabled = 0;
512}
513
514
515
516
517
518
519int proc_dowatchdog(struct ctl_table *table, int write,
520 void __user *buffer, size_t *lenp, loff_t *ppos)
521{
522 int ret;
523
524 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
525 if (ret || !write)
526 goto out;
527
528 if (watchdog_enabled && watchdog_thresh)
529 watchdog_enable_all_cpus();
530 else
531 watchdog_disable_all_cpus();
532
533out:
534 return ret;
535}
536#endif
537
538
539
540
541
542static int __cpuinit
543cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
544{
545 int hotcpu = (unsigned long)hcpu;
546
547 switch (action) {
548 case CPU_UP_PREPARE:
549 case CPU_UP_PREPARE_FROZEN:
550 watchdog_prepare_cpu(hotcpu);
551 break;
552 case CPU_ONLINE:
553 case CPU_ONLINE_FROZEN:
554 if (watchdog_enabled)
555 watchdog_enable(hotcpu);
556 break;
557#ifdef CONFIG_HOTPLUG_CPU
558 case CPU_UP_CANCELED:
559 case CPU_UP_CANCELED_FROZEN:
560 watchdog_disable(hotcpu);
561 break;
562 case CPU_DEAD:
563 case CPU_DEAD_FROZEN:
564 watchdog_disable(hotcpu);
565 break;
566#endif
567 }
568
569
570
571
572
573
574 return NOTIFY_OK;
575}
576
577static struct notifier_block __cpuinitdata cpu_nfb = {
578 .notifier_call = cpu_callback
579};
580
581void __init lockup_detector_init(void)
582{
583 void *cpu = (void *)(long)smp_processor_id();
584 int err;
585
586 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
587 WARN_ON(notifier_to_errno(err));
588
589 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
590 register_cpu_notifier(&cpu_nfb);
591
592 return;
593}
594