Commit | Line | Data |
---|---|---|
58687acb DZ |
1 | /* |
2 | * Detect hard and soft lockups on a system | |
3 | * | |
4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. | |
5 | * | |
6 | * this code detects hard lockups: incidents in where on a CPU | |
7 | * the kernel does not respond to anything except NMI. | |
8 | * | |
9 | * Note: Most of this code is borrowed heavily from softlockup.c, | |
10 | * so thanks to Ingo for the initial implementation. | |
11 | * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks | |
12 | * to those contributors as well. | |
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 | ||
30 | int watchdog_enabled; | |
31 | int __read_mostly softlockup_thresh = 60; | |
32 | ||
33 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
d7c54733 | 34 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); |
58687acb DZ |
35 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); |
36 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
37 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
38 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); | |
39 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); | |
40 | #ifdef CONFIG_PERF_EVENTS_NMI | |
41 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); | |
42 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); | |
43 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); | |
44 | #endif | |
45 | ||
46 | static int __read_mostly did_panic; | |
47 | static int __initdata no_watchdog; | |
48 | ||
49 | ||
50 | /* boot commands */ | |
51 | /* | |
52 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
53 | */ | |
54 | #ifdef CONFIG_PERF_EVENTS_NMI | |
55 | static int hardlockup_panic; | |
56 | ||
57 | static int __init hardlockup_panic_setup(char *str) | |
58 | { | |
59 | if (!strncmp(str, "panic", 5)) | |
60 | hardlockup_panic = 1; | |
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 | ||
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 | ||
77 | static int __init nowatchdog_setup(char *str) | |
78 | { | |
79 | no_watchdog = 1; | |
80 | return 1; | |
81 | } | |
82 | __setup("nowatchdog", nowatchdog_setup); | |
83 | ||
84 | /* deprecated */ | |
85 | static int __init nosoftlockup_setup(char *str) | |
86 | { | |
87 | no_watchdog = 1; | |
88 | return 1; | |
89 | } | |
90 | __setup("nosoftlockup", nosoftlockup_setup); | |
91 | /* */ | |
92 | ||
93 | ||
94 | /* | |
95 | * Returns seconds, approximately. We don't need nanosecond | |
96 | * resolution, and we don't need to waste time with a big divide when | |
97 | * 2^30ns == 1.074s. | |
98 | */ | |
99 | static unsigned long get_timestamp(int this_cpu) | |
100 | { | |
101 | return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */ | |
102 | } | |
103 | ||
104 | static unsigned long get_sample_period(void) | |
105 | { | |
106 | /* | |
107 | * convert softlockup_thresh from seconds to ns | |
108 | * the divide by 5 is to give hrtimer 5 chances to | |
109 | * increment before the hardlockup detector generates | |
110 | * a warning | |
111 | */ | |
112 | return softlockup_thresh / 5 * NSEC_PER_SEC; | |
113 | } | |
114 | ||
115 | /* Commands for resetting the watchdog */ | |
116 | static void __touch_watchdog(void) | |
117 | { | |
118 | int this_cpu = raw_smp_processor_id(); | |
119 | ||
120 | __get_cpu_var(watchdog_touch_ts) = get_timestamp(this_cpu); | |
121 | } | |
122 | ||
332fbdbc | 123 | void touch_softlockup_watchdog(void) |
58687acb DZ |
124 | { |
125 | __get_cpu_var(watchdog_touch_ts) = 0; | |
126 | } | |
58687acb | 127 | |
332fbdbc | 128 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
129 | { |
130 | int cpu; | |
131 | ||
132 | /* | |
133 | * this is done lockless | |
134 | * do we care if a 0 races with a timestamp? | |
135 | * all it means is the softlock check starts one cycle later | |
136 | */ | |
137 | for_each_online_cpu(cpu) | |
138 | per_cpu(watchdog_touch_ts, cpu) = 0; | |
139 | } | |
140 | ||
141 | void touch_nmi_watchdog(void) | |
142 | { | |
d7c54733 | 143 | __get_cpu_var(watchdog_nmi_touch) = true; |
332fbdbc | 144 | touch_softlockup_watchdog(); |
58687acb DZ |
145 | } |
146 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
147 | ||
58687acb DZ |
148 | void touch_softlockup_watchdog_sync(void) |
149 | { | |
150 | __raw_get_cpu_var(softlockup_touch_sync) = true; | |
151 | __raw_get_cpu_var(watchdog_touch_ts) = 0; | |
152 | } | |
153 | ||
58687acb DZ |
154 | #ifdef CONFIG_PERF_EVENTS_NMI |
155 | /* watchdog detector functions */ | |
156 | static int is_hardlockup(int cpu) | |
157 | { | |
158 | unsigned long hrint = per_cpu(hrtimer_interrupts, cpu); | |
159 | ||
160 | if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint) | |
161 | return 1; | |
162 | ||
163 | per_cpu(hrtimer_interrupts_saved, cpu) = hrint; | |
164 | return 0; | |
165 | } | |
166 | #endif | |
167 | ||
168 | static int is_softlockup(unsigned long touch_ts, int cpu) | |
169 | { | |
170 | unsigned long now = get_timestamp(cpu); | |
171 | ||
172 | /* Warn about unreasonable delays: */ | |
173 | if (time_after(now, touch_ts + softlockup_thresh)) | |
174 | return now - touch_ts; | |
175 | ||
176 | return 0; | |
177 | } | |
178 | ||
179 | static int | |
180 | watchdog_panic(struct notifier_block *this, unsigned long event, void *ptr) | |
181 | { | |
182 | did_panic = 1; | |
183 | ||
184 | return NOTIFY_DONE; | |
185 | } | |
186 | ||
187 | static struct notifier_block panic_block = { | |
188 | .notifier_call = watchdog_panic, | |
189 | }; | |
190 | ||
191 | #ifdef CONFIG_PERF_EVENTS_NMI | |
192 | static struct perf_event_attr wd_hw_attr = { | |
193 | .type = PERF_TYPE_HARDWARE, | |
194 | .config = PERF_COUNT_HW_CPU_CYCLES, | |
195 | .size = sizeof(struct perf_event_attr), | |
196 | .pinned = 1, | |
197 | .disabled = 1, | |
198 | }; | |
199 | ||
200 | /* Callback function for perf event subsystem */ | |
201 | void watchdog_overflow_callback(struct perf_event *event, int nmi, | |
202 | struct perf_sample_data *data, | |
203 | struct pt_regs *regs) | |
204 | { | |
205 | int this_cpu = smp_processor_id(); | |
58687acb | 206 | |
d7c54733 DZ |
207 | if (__get_cpu_var(watchdog_nmi_touch) == true) { |
208 | __get_cpu_var(watchdog_nmi_touch) = false; | |
58687acb DZ |
209 | return; |
210 | } | |
211 | ||
212 | /* check for a hardlockup | |
213 | * This is done by making sure our timer interrupt | |
214 | * is incrementing. The timer interrupt should have | |
215 | * fired multiple times before we overflow'd. If it hasn't | |
216 | * then this is a good indication the cpu is stuck | |
217 | */ | |
218 | if (is_hardlockup(this_cpu)) { | |
219 | /* only print hardlockups once */ | |
220 | if (__get_cpu_var(hard_watchdog_warn) == true) | |
221 | return; | |
222 | ||
223 | if (hardlockup_panic) | |
224 | panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu); | |
225 | else | |
226 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu); | |
227 | ||
228 | __get_cpu_var(hard_watchdog_warn) = true; | |
229 | return; | |
230 | } | |
231 | ||
232 | __get_cpu_var(hard_watchdog_warn) = false; | |
233 | return; | |
234 | } | |
235 | static void watchdog_interrupt_count(void) | |
236 | { | |
237 | __get_cpu_var(hrtimer_interrupts)++; | |
238 | } | |
239 | #else | |
240 | static inline void watchdog_interrupt_count(void) { return; } | |
241 | #endif /* CONFIG_PERF_EVENTS_NMI */ | |
242 | ||
243 | /* watchdog kicker functions */ | |
244 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
245 | { | |
246 | int this_cpu = smp_processor_id(); | |
247 | unsigned long touch_ts = __get_cpu_var(watchdog_touch_ts); | |
248 | struct pt_regs *regs = get_irq_regs(); | |
249 | int duration; | |
250 | ||
251 | /* kick the hardlockup detector */ | |
252 | watchdog_interrupt_count(); | |
253 | ||
254 | /* kick the softlockup detector */ | |
255 | wake_up_process(__get_cpu_var(softlockup_watchdog)); | |
256 | ||
257 | /* .. and repeat */ | |
258 | hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period())); | |
259 | ||
260 | if (touch_ts == 0) { | |
261 | if (unlikely(per_cpu(softlockup_touch_sync, this_cpu))) { | |
262 | /* | |
263 | * If the time stamp was touched atomically | |
264 | * make sure the scheduler tick is up to date. | |
265 | */ | |
266 | per_cpu(softlockup_touch_sync, this_cpu) = false; | |
267 | sched_clock_tick(); | |
268 | } | |
269 | __touch_watchdog(); | |
270 | return HRTIMER_RESTART; | |
271 | } | |
272 | ||
273 | /* check for a softlockup | |
274 | * This is done by making sure a high priority task is | |
275 | * being scheduled. The task touches the watchdog to | |
276 | * indicate it is getting cpu time. If it hasn't then | |
277 | * this is a good indication some task is hogging the cpu | |
278 | */ | |
279 | duration = is_softlockup(touch_ts, this_cpu); | |
280 | if (unlikely(duration)) { | |
281 | /* only warn once */ | |
282 | if (__get_cpu_var(soft_watchdog_warn) == true) | |
283 | return HRTIMER_RESTART; | |
284 | ||
285 | printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", | |
286 | this_cpu, duration, | |
287 | current->comm, task_pid_nr(current)); | |
288 | print_modules(); | |
289 | print_irqtrace_events(current); | |
290 | if (regs) | |
291 | show_regs(regs); | |
292 | else | |
293 | dump_stack(); | |
294 | ||
295 | if (softlockup_panic) | |
296 | panic("softlockup: hung tasks"); | |
297 | __get_cpu_var(soft_watchdog_warn) = true; | |
298 | } else | |
299 | __get_cpu_var(soft_watchdog_warn) = false; | |
300 | ||
301 | return HRTIMER_RESTART; | |
302 | } | |
303 | ||
304 | ||
305 | /* | |
306 | * The watchdog thread - touches the timestamp. | |
307 | */ | |
308 | static int watchdog(void *__bind_cpu) | |
309 | { | |
310 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; | |
311 | struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, (unsigned long)__bind_cpu); | |
312 | ||
313 | sched_setscheduler(current, SCHED_FIFO, ¶m); | |
314 | ||
315 | /* initialize timestamp */ | |
316 | __touch_watchdog(); | |
317 | ||
318 | /* kick off the timer for the hardlockup detector */ | |
319 | /* done here because hrtimer_start can only pin to smp_processor_id() */ | |
320 | hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()), | |
321 | HRTIMER_MODE_REL_PINNED); | |
322 | ||
323 | set_current_state(TASK_INTERRUPTIBLE); | |
324 | /* | |
325 | * Run briefly once per second to reset the softlockup timestamp. | |
326 | * If this gets delayed for more than 60 seconds then the | |
327 | * debug-printout triggers in softlockup_tick(). | |
328 | */ | |
329 | while (!kthread_should_stop()) { | |
330 | __touch_watchdog(); | |
331 | schedule(); | |
332 | ||
333 | if (kthread_should_stop()) | |
334 | break; | |
335 | ||
336 | set_current_state(TASK_INTERRUPTIBLE); | |
337 | } | |
338 | __set_current_state(TASK_RUNNING); | |
339 | ||
340 | return 0; | |
341 | } | |
342 | ||
343 | ||
344 | #ifdef CONFIG_PERF_EVENTS_NMI | |
345 | static int watchdog_nmi_enable(int cpu) | |
346 | { | |
347 | struct perf_event_attr *wd_attr; | |
348 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
349 | ||
350 | /* is it already setup and enabled? */ | |
351 | if (event && event->state > PERF_EVENT_STATE_OFF) | |
352 | goto out; | |
353 | ||
354 | /* it is setup but not enabled */ | |
355 | if (event != NULL) | |
356 | goto out_enable; | |
357 | ||
358 | /* Try to register using hardware perf events */ | |
359 | wd_attr = &wd_hw_attr; | |
360 | wd_attr->sample_period = hw_nmi_get_sample_period(); | |
361 | event = perf_event_create_kernel_counter(wd_attr, cpu, -1, watchdog_overflow_callback); | |
362 | if (!IS_ERR(event)) { | |
363 | printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n"); | |
364 | goto out_save; | |
365 | } | |
366 | ||
367 | printk(KERN_ERR "NMI watchdog failed to create perf event on cpu%i: %p\n", cpu, event); | |
368 | return -1; | |
369 | ||
370 | /* success path */ | |
371 | out_save: | |
372 | per_cpu(watchdog_ev, cpu) = event; | |
373 | out_enable: | |
374 | perf_event_enable(per_cpu(watchdog_ev, cpu)); | |
375 | out: | |
376 | return 0; | |
377 | } | |
378 | ||
379 | static void watchdog_nmi_disable(int cpu) | |
380 | { | |
381 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
382 | ||
383 | if (event) { | |
384 | perf_event_disable(event); | |
385 | per_cpu(watchdog_ev, cpu) = NULL; | |
386 | ||
387 | /* should be in cleanup, but blocks oprofile */ | |
388 | perf_event_release_kernel(event); | |
389 | } | |
390 | return; | |
391 | } | |
392 | #else | |
393 | static int watchdog_nmi_enable(int cpu) { return 0; } | |
394 | static void watchdog_nmi_disable(int cpu) { return; } | |
395 | #endif /* CONFIG_PERF_EVENTS_NMI */ | |
396 | ||
397 | /* prepare/enable/disable routines */ | |
398 | static int watchdog_prepare_cpu(int cpu) | |
399 | { | |
400 | struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu); | |
401 | ||
402 | WARN_ON(per_cpu(softlockup_watchdog, cpu)); | |
403 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
404 | hrtimer->function = watchdog_timer_fn; | |
405 | ||
406 | return 0; | |
407 | } | |
408 | ||
409 | static int watchdog_enable(int cpu) | |
410 | { | |
411 | struct task_struct *p = per_cpu(softlockup_watchdog, cpu); | |
412 | ||
413 | /* enable the perf event */ | |
414 | if (watchdog_nmi_enable(cpu) != 0) | |
415 | return -1; | |
416 | ||
417 | /* create the watchdog thread */ | |
418 | if (!p) { | |
419 | p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu); | |
420 | if (IS_ERR(p)) { | |
421 | printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu); | |
422 | return -1; | |
423 | } | |
424 | kthread_bind(p, cpu); | |
425 | per_cpu(watchdog_touch_ts, cpu) = 0; | |
426 | per_cpu(softlockup_watchdog, cpu) = p; | |
427 | wake_up_process(p); | |
428 | } | |
429 | ||
430 | return 0; | |
431 | } | |
432 | ||
433 | static void watchdog_disable(int cpu) | |
434 | { | |
435 | struct task_struct *p = per_cpu(softlockup_watchdog, cpu); | |
436 | struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu); | |
437 | ||
438 | /* | |
439 | * cancel the timer first to stop incrementing the stats | |
440 | * and waking up the kthread | |
441 | */ | |
442 | hrtimer_cancel(hrtimer); | |
443 | ||
444 | /* disable the perf event */ | |
445 | watchdog_nmi_disable(cpu); | |
446 | ||
447 | /* stop the watchdog thread */ | |
448 | if (p) { | |
449 | per_cpu(softlockup_watchdog, cpu) = NULL; | |
450 | kthread_stop(p); | |
451 | } | |
452 | ||
453 | /* if any cpu succeeds, watchdog is considered enabled for the system */ | |
454 | watchdog_enabled = 1; | |
455 | } | |
456 | ||
457 | static void watchdog_enable_all_cpus(void) | |
458 | { | |
459 | int cpu; | |
460 | int result; | |
461 | ||
462 | for_each_online_cpu(cpu) | |
463 | result += watchdog_enable(cpu); | |
464 | ||
465 | if (result) | |
466 | printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n"); | |
467 | ||
468 | } | |
469 | ||
470 | static void watchdog_disable_all_cpus(void) | |
471 | { | |
472 | int cpu; | |
473 | ||
474 | for_each_online_cpu(cpu) | |
475 | watchdog_disable(cpu); | |
476 | ||
477 | /* if all watchdogs are disabled, then they are disabled for the system */ | |
478 | watchdog_enabled = 0; | |
479 | } | |
480 | ||
481 | ||
482 | /* sysctl functions */ | |
483 | #ifdef CONFIG_SYSCTL | |
484 | /* | |
485 | * proc handler for /proc/sys/kernel/nmi_watchdog | |
486 | */ | |
487 | ||
488 | int proc_dowatchdog_enabled(struct ctl_table *table, int write, | |
489 | void __user *buffer, size_t *length, loff_t *ppos) | |
490 | { | |
491 | proc_dointvec(table, write, buffer, length, ppos); | |
492 | ||
493 | if (watchdog_enabled) | |
494 | watchdog_enable_all_cpus(); | |
495 | else | |
496 | watchdog_disable_all_cpus(); | |
497 | return 0; | |
498 | } | |
499 | ||
500 | int proc_dowatchdog_thresh(struct ctl_table *table, int write, | |
501 | void __user *buffer, | |
502 | size_t *lenp, loff_t *ppos) | |
503 | { | |
504 | return proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
505 | } | |
58687acb DZ |
506 | #endif /* CONFIG_SYSCTL */ |
507 | ||
508 | ||
509 | /* | |
510 | * Create/destroy watchdog threads as CPUs come and go: | |
511 | */ | |
512 | static int __cpuinit | |
513 | cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu) | |
514 | { | |
515 | int hotcpu = (unsigned long)hcpu; | |
516 | ||
517 | switch (action) { | |
518 | case CPU_UP_PREPARE: | |
519 | case CPU_UP_PREPARE_FROZEN: | |
520 | if (watchdog_prepare_cpu(hotcpu)) | |
521 | return NOTIFY_BAD; | |
522 | break; | |
523 | case CPU_ONLINE: | |
524 | case CPU_ONLINE_FROZEN: | |
525 | if (watchdog_enable(hotcpu)) | |
526 | return NOTIFY_BAD; | |
527 | break; | |
528 | #ifdef CONFIG_HOTPLUG_CPU | |
529 | case CPU_UP_CANCELED: | |
530 | case CPU_UP_CANCELED_FROZEN: | |
531 | watchdog_disable(hotcpu); | |
532 | break; | |
533 | case CPU_DEAD: | |
534 | case CPU_DEAD_FROZEN: | |
535 | watchdog_disable(hotcpu); | |
536 | break; | |
537 | #endif /* CONFIG_HOTPLUG_CPU */ | |
538 | } | |
539 | return NOTIFY_OK; | |
540 | } | |
541 | ||
542 | static struct notifier_block __cpuinitdata cpu_nfb = { | |
543 | .notifier_call = cpu_callback | |
544 | }; | |
545 | ||
546 | static int __init spawn_watchdog_task(void) | |
547 | { | |
548 | void *cpu = (void *)(long)smp_processor_id(); | |
549 | int err; | |
550 | ||
551 | if (no_watchdog) | |
552 | return 0; | |
553 | ||
554 | err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu); | |
555 | WARN_ON(err == NOTIFY_BAD); | |
556 | ||
557 | cpu_callback(&cpu_nfb, CPU_ONLINE, cpu); | |
558 | register_cpu_notifier(&cpu_nfb); | |
559 | ||
560 | atomic_notifier_chain_register(&panic_notifier_list, &panic_block); | |
561 | ||
562 | return 0; | |
563 | } | |
564 | early_initcall(spawn_watchdog_task); |