]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blame - kernel/watchdog.c
Merge commit 'v3.2-rc2' into fbdev-next
[mirror_ubuntu-zesty-kernel.git] / kernel / watchdog.c
CommitLineData
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
4135038a 30int watchdog_enabled = 1;
4eec42f3 31int __read_mostly watchdog_thresh = 10;
58687acb
DZ
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);
58687acb 37static DEFINE_PER_CPU(bool, soft_watchdog_warn);
23637d47 38#ifdef CONFIG_HARDLOCKUP_DETECTOR
cafcd80d
DZ
39static DEFINE_PER_CPU(bool, hard_watchdog_warn);
40static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
58687acb
DZ
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
58687acb
DZ
46/* boot commands */
47/*
48 * Should we panic when a soft-lockup or hard-lockup occurs:
49 */
23637d47 50#ifdef CONFIG_HARDLOCKUP_DETECTOR
fef2c9bc
DZ
51static int hardlockup_panic =
52 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
58687acb
DZ
53
54static int __init hardlockup_panic_setup(char *str)
55{
56 if (!strncmp(str, "panic", 5))
57 hardlockup_panic = 1;
fef2c9bc
DZ
58 else if (!strncmp(str, "nopanic", 7))
59 hardlockup_panic = 0;
5dc30558 60 else if (!strncmp(str, "0", 1))
4135038a 61 watchdog_enabled = 0;
58687acb
DZ
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{
4135038a 80 watchdog_enabled = 0;
58687acb
DZ
81 return 1;
82}
83__setup("nowatchdog", nowatchdog_setup);
84
85/* deprecated */
86static int __init nosoftlockup_setup(char *str)
87{
4135038a 88 watchdog_enabled = 0;
58687acb
DZ
89 return 1;
90}
91__setup("nosoftlockup", nosoftlockup_setup);
92/* */
93
4eec42f3
MSB
94/*
95 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
96 * lockups can have false positives under extreme conditions. So we generally
97 * want a higher threshold for soft lockups than for hard lockups. So we couple
98 * the thresholds with a factor: we make the soft threshold twice the amount of
99 * time the hard threshold is.
100 */
6e9101ae 101static int get_softlockup_thresh(void)
4eec42f3
MSB
102{
103 return watchdog_thresh * 2;
104}
58687acb
DZ
105
106/*
107 * Returns seconds, approximately. We don't need nanosecond
108 * resolution, and we don't need to waste time with a big divide when
109 * 2^30ns == 1.074s.
110 */
111static unsigned long get_timestamp(int this_cpu)
112{
113 return cpu_clock(this_cpu) >> 30LL; /* 2^30 ~= 10^9 */
114}
115
116static unsigned long get_sample_period(void)
117{
118 /*
586692a5 119 * convert watchdog_thresh from seconds to ns
58687acb
DZ
120 * the divide by 5 is to give hrtimer 5 chances to
121 * increment before the hardlockup detector generates
122 * a warning
123 */
4eec42f3 124 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
58687acb
DZ
125}
126
127/* Commands for resetting the watchdog */
128static void __touch_watchdog(void)
129{
26e09c6e 130 int this_cpu = smp_processor_id();
58687acb 131
909ea964 132 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
58687acb
DZ
133}
134
332fbdbc 135void touch_softlockup_watchdog(void)
58687acb 136{
909ea964 137 __this_cpu_write(watchdog_touch_ts, 0);
58687acb 138}
0167c781 139EXPORT_SYMBOL(touch_softlockup_watchdog);
58687acb 140
332fbdbc 141void touch_all_softlockup_watchdogs(void)
58687acb
DZ
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
cafcd80d 154#ifdef CONFIG_HARDLOCKUP_DETECTOR
58687acb
DZ
155void touch_nmi_watchdog(void)
156{
68d3f1d8
DZ
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 }
332fbdbc 165 touch_softlockup_watchdog();
58687acb
DZ
166}
167EXPORT_SYMBOL(touch_nmi_watchdog);
168
cafcd80d
DZ
169#endif
170
58687acb
DZ
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
23637d47 177#ifdef CONFIG_HARDLOCKUP_DETECTOR
58687acb 178/* watchdog detector functions */
26e09c6e 179static int is_hardlockup(void)
58687acb 180{
909ea964 181 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
58687acb 182
909ea964 183 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
58687acb
DZ
184 return 1;
185
909ea964 186 __this_cpu_write(hrtimer_interrupts_saved, hrint);
58687acb
DZ
187 return 0;
188}
189#endif
190
26e09c6e 191static int is_softlockup(unsigned long touch_ts)
58687acb 192{
26e09c6e 193 unsigned long now = get_timestamp(smp_processor_id());
58687acb
DZ
194
195 /* Warn about unreasonable delays: */
4eec42f3 196 if (time_after(now, touch_ts + get_softlockup_thresh()))
58687acb
DZ
197 return now - touch_ts;
198
199 return 0;
200}
201
23637d47 202#ifdef CONFIG_HARDLOCKUP_DETECTOR
1880c4ae 203
58687acb
DZ
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/* Callback function for perf event subsystem */
a8b0ca17 213static void watchdog_overflow_callback(struct perf_event *event,
58687acb
DZ
214 struct perf_sample_data *data,
215 struct pt_regs *regs)
216{
c6db67cd
PZ
217 /* Ensure the watchdog never gets throttled */
218 event->hw.interrupts = 0;
219
909ea964
CL
220 if (__this_cpu_read(watchdog_nmi_touch) == true) {
221 __this_cpu_write(watchdog_nmi_touch, false);
58687acb
DZ
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 */
26e09c6e
DZ
231 if (is_hardlockup()) {
232 int this_cpu = smp_processor_id();
233
58687acb 234 /* only print hardlockups once */
909ea964 235 if (__this_cpu_read(hard_watchdog_warn) == true)
58687acb
DZ
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
909ea964 243 __this_cpu_write(hard_watchdog_warn, true);
58687acb
DZ
244 return;
245 }
246
909ea964 247 __this_cpu_write(hard_watchdog_warn, false);
58687acb
DZ
248 return;
249}
250static void watchdog_interrupt_count(void)
251{
909ea964 252 __this_cpu_inc(hrtimer_interrupts);
58687acb
DZ
253}
254#else
255static inline void watchdog_interrupt_count(void) { return; }
23637d47 256#endif /* CONFIG_HARDLOCKUP_DETECTOR */
58687acb
DZ
257
258/* watchdog kicker functions */
259static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
260{
909ea964 261 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
58687acb
DZ
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 */
909ea964 269 wake_up_process(__this_cpu_read(softlockup_watchdog));
58687acb
DZ
270
271 /* .. and repeat */
272 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
273
274 if (touch_ts == 0) {
909ea964 275 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
58687acb
DZ
276 /*
277 * If the time stamp was touched atomically
278 * make sure the scheduler tick is up to date.
279 */
909ea964 280 __this_cpu_write(softlockup_touch_sync, false);
58687acb
DZ
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 */
26e09c6e 293 duration = is_softlockup(touch_ts);
58687acb
DZ
294 if (unlikely(duration)) {
295 /* only warn once */
909ea964 296 if (__this_cpu_read(soft_watchdog_warn) == true)
58687acb
DZ
297 return HRTIMER_RESTART;
298
299 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
26e09c6e 300 smp_processor_id(), duration,
58687acb
DZ
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");
909ea964 311 __this_cpu_write(soft_watchdog_warn, true);
58687acb 312 } else
909ea964 313 __this_cpu_write(soft_watchdog_warn, false);
58687acb
DZ
314
315 return HRTIMER_RESTART;
316}
317
318
319/*
320 * The watchdog thread - touches the timestamp.
321 */
26e09c6e 322static int watchdog(void *unused)
58687acb 323{
cba9bd22 324 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
26e09c6e 325 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
58687acb
DZ
326
327 sched_setscheduler(current, SCHED_FIFO, &param);
328
329 /* initialize timestamp */
330 __touch_watchdog();
331
332 /* kick off the timer for the hardlockup detector */
333 /* done here because hrtimer_start can only pin to smp_processor_id() */
334 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
335 HRTIMER_MODE_REL_PINNED);
336
337 set_current_state(TASK_INTERRUPTIBLE);
338 /*
339 * Run briefly once per second to reset the softlockup timestamp.
340 * If this gets delayed for more than 60 seconds then the
26e09c6e 341 * debug-printout triggers in watchdog_timer_fn().
58687acb
DZ
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);
cba9bd22
TG
353 param.sched_priority = 0;
354 sched_setscheduler(current, SCHED_NORMAL, &param);
58687acb
DZ
355 return 0;
356}
357
358
23637d47 359#ifdef CONFIG_HARDLOCKUP_DETECTOR
58687acb
DZ
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 /* is it already setup and enabled? */
366 if (event && event->state > PERF_EVENT_STATE_OFF)
367 goto out;
368
369 /* it is setup but not enabled */
370 if (event != NULL)
371 goto out_enable;
372
58687acb 373 wd_attr = &wd_hw_attr;
4eec42f3 374 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
1880c4ae
CG
375
376 /* Try to register using hardware perf events */
4dc0da86 377 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
58687acb
DZ
378 if (!IS_ERR(event)) {
379 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
380 goto out_save;
381 }
382
5651f7f4
DZ
383
384 /* vary the KERN level based on the returned errno */
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));
eac24335 391 return PTR_ERR(event);
58687acb
DZ
392
393 /* success path */
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 /* should be in cleanup, but blocks oprofile */
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; }
23637d47 418#endif /* CONFIG_HARDLOCKUP_DETECTOR */
58687acb
DZ
419
420/* prepare/enable/disable routines */
6e9101ae 421static void watchdog_prepare_cpu(int cpu)
58687acb
DZ
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;
58687acb
DZ
428}
429
430static int watchdog_enable(int cpu)
431{
432 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
f99a9933 433 int err = 0;
58687acb
DZ
434
435 /* enable the perf event */
eac24335 436 err = watchdog_nmi_enable(cpu);
f99a9933
DZ
437
438 /* Regardless of err above, fall through and start softlockup */
58687acb
DZ
439
440 /* create the watchdog thread */
441 if (!p) {
18e5a45d 442 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
58687acb
DZ
443 if (IS_ERR(p)) {
444 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
1409f141 445 if (!err) {
f99a9933
DZ
446 /* if hardlockup hasn't already set this */
447 err = PTR_ERR(p);
1409f141
HD
448 /* and disable the perf event */
449 watchdog_nmi_disable(cpu);
450 }
f99a9933 451 goto out;
58687acb
DZ
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
f99a9933
DZ
459out:
460 return err;
58687acb
DZ
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 * cancel the timer first to stop incrementing the stats
470 * and waking up the kthread
471 */
472 hrtimer_cancel(hrtimer);
473
474 /* disable the perf event */
475 watchdog_nmi_disable(cpu);
476
477 /* stop the watchdog thread */
478 if (p) {
479 per_cpu(softlockup_watchdog, cpu) = NULL;
480 kthread_stop(p);
481 }
58687acb
DZ
482}
483
4ff81951
VA
484/* sysctl functions */
485#ifdef CONFIG_SYSCTL
58687acb
DZ
486static void watchdog_enable_all_cpus(void)
487{
488 int cpu;
39735766
MS
489
490 watchdog_enabled = 0;
58687acb
DZ
491
492 for_each_online_cpu(cpu)
39735766
MS
493 if (!watchdog_enable(cpu))
494 /* if any cpu succeeds, watchdog is considered
495 enabled for the system */
496 watchdog_enabled = 1;
58687acb 497
39735766 498 if (!watchdog_enabled)
58687acb
DZ
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 /* if all watchdogs are disabled, then they are disabled for the system */
511 watchdog_enabled = 0;
512}
513
514
58687acb 515/*
586692a5 516 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
58687acb
DZ
517 */
518
586692a5
MSB
519int proc_dowatchdog(struct ctl_table *table, int write,
520 void __user *buffer, size_t *lenp, loff_t *ppos)
58687acb 521{
e04ab2bc 522 int ret;
58687acb 523
586692a5 524 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
e04ab2bc
MSB
525 if (ret || !write)
526 goto out;
527
586692a5 528 if (watchdog_enabled && watchdog_thresh)
e04ab2bc
MSB
529 watchdog_enable_all_cpus();
530 else
531 watchdog_disable_all_cpus();
532
533out:
534 return ret;
58687acb 535}
58687acb
DZ
536#endif /* CONFIG_SYSCTL */
537
538
539/*
540 * Create/destroy watchdog threads as CPUs come and go:
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:
6e9101ae 550 watchdog_prepare_cpu(hotcpu);
58687acb
DZ
551 break;
552 case CPU_ONLINE:
553 case CPU_ONLINE_FROZEN:
4135038a 554 if (watchdog_enabled)
6e9101ae 555 watchdog_enable(hotcpu);
58687acb
DZ
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 /* CONFIG_HOTPLUG_CPU */
567 }
f99a9933
DZ
568
569 /*
570 * hardlockup and softlockup are not important enough
571 * to block cpu bring up. Just always succeed and
572 * rely on printk output to flag problems.
573 */
574 return NOTIFY_OK;
58687acb
DZ
575}
576
577static struct notifier_block __cpuinitdata cpu_nfb = {
578 .notifier_call = cpu_callback
579};
580
004417a6 581void __init lockup_detector_init(void)
58687acb
DZ
582{
583 void *cpu = (void *)(long)smp_processor_id();
584 int err;
585
58687acb 586 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
eac24335 587 WARN_ON(notifier_to_errno(err));
58687acb
DZ
588
589 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
590 register_cpu_notifier(&cpu_nfb);
591
004417a6 592 return;
58687acb 593}