Loading...
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/kvm_para.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/* boot commands */
47/*
48 * Should we panic when a soft-lockup or hard-lockup occurs:
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/* deprecated */
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 * 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 */
101static int get_softlockup_thresh(void)
102{
103 return watchdog_thresh * 2;
104}
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 /*
119 * convert watchdog_thresh from seconds to ns
120 * the divide by 5 is to give hrtimer several chances (two
121 * or three with the current relation between the soft
122 * and hard thresholds) to increment before the
123 * hardlockup detector generates a warning
124 */
125 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
126}
127
128/* Commands for resetting the watchdog */
129static void __touch_watchdog(void)
130{
131 int this_cpu = smp_processor_id();
132
133 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
134}
135
136void touch_softlockup_watchdog(void)
137{
138 __this_cpu_write(watchdog_touch_ts, 0);
139}
140EXPORT_SYMBOL(touch_softlockup_watchdog);
141
142void touch_all_softlockup_watchdogs(void)
143{
144 int cpu;
145
146 /*
147 * this is done lockless
148 * do we care if a 0 races with a timestamp?
149 * all it means is the softlock check starts one cycle later
150 */
151 for_each_online_cpu(cpu)
152 per_cpu(watchdog_touch_ts, cpu) = 0;
153}
154
155#ifdef CONFIG_HARDLOCKUP_DETECTOR
156void touch_nmi_watchdog(void)
157{
158 if (watchdog_enabled) {
159 unsigned cpu;
160
161 for_each_present_cpu(cpu) {
162 if (per_cpu(watchdog_nmi_touch, cpu) != true)
163 per_cpu(watchdog_nmi_touch, cpu) = true;
164 }
165 }
166 touch_softlockup_watchdog();
167}
168EXPORT_SYMBOL(touch_nmi_watchdog);
169
170#endif
171
172void touch_softlockup_watchdog_sync(void)
173{
174 __raw_get_cpu_var(softlockup_touch_sync) = true;
175 __raw_get_cpu_var(watchdog_touch_ts) = 0;
176}
177
178#ifdef CONFIG_HARDLOCKUP_DETECTOR
179/* watchdog detector functions */
180static int is_hardlockup(void)
181{
182 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
183
184 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
185 return 1;
186
187 __this_cpu_write(hrtimer_interrupts_saved, hrint);
188 return 0;
189}
190#endif
191
192static int is_softlockup(unsigned long touch_ts)
193{
194 unsigned long now = get_timestamp(smp_processor_id());
195
196 /* Warn about unreasonable delays: */
197 if (time_after(now, touch_ts + get_softlockup_thresh()))
198 return now - touch_ts;
199
200 return 0;
201}
202
203#ifdef CONFIG_HARDLOCKUP_DETECTOR
204
205static struct perf_event_attr wd_hw_attr = {
206 .type = PERF_TYPE_HARDWARE,
207 .config = PERF_COUNT_HW_CPU_CYCLES,
208 .size = sizeof(struct perf_event_attr),
209 .pinned = 1,
210 .disabled = 1,
211};
212
213/* Callback function for perf event subsystem */
214static void watchdog_overflow_callback(struct perf_event *event,
215 struct perf_sample_data *data,
216 struct pt_regs *regs)
217{
218 /* Ensure the watchdog never gets throttled */
219 event->hw.interrupts = 0;
220
221 if (__this_cpu_read(watchdog_nmi_touch) == true) {
222 __this_cpu_write(watchdog_nmi_touch, false);
223 return;
224 }
225
226 /* check for a hardlockup
227 * This is done by making sure our timer interrupt
228 * is incrementing. The timer interrupt should have
229 * fired multiple times before we overflow'd. If it hasn't
230 * then this is a good indication the cpu is stuck
231 */
232 if (is_hardlockup()) {
233 int this_cpu = smp_processor_id();
234
235 /* only print hardlockups once */
236 if (__this_cpu_read(hard_watchdog_warn) == true)
237 return;
238
239 if (hardlockup_panic)
240 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
241 else
242 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
243
244 __this_cpu_write(hard_watchdog_warn, true);
245 return;
246 }
247
248 __this_cpu_write(hard_watchdog_warn, false);
249 return;
250}
251static void watchdog_interrupt_count(void)
252{
253 __this_cpu_inc(hrtimer_interrupts);
254}
255#else
256static inline void watchdog_interrupt_count(void) { return; }
257#endif /* CONFIG_HARDLOCKUP_DETECTOR */
258
259/* watchdog kicker functions */
260static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
261{
262 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
263 struct pt_regs *regs = get_irq_regs();
264 int duration;
265
266 /* kick the hardlockup detector */
267 watchdog_interrupt_count();
268
269 /* kick the softlockup detector */
270 wake_up_process(__this_cpu_read(softlockup_watchdog));
271
272 /* .. and repeat */
273 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
274
275 if (touch_ts == 0) {
276 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
277 /*
278 * If the time stamp was touched atomically
279 * make sure the scheduler tick is up to date.
280 */
281 __this_cpu_write(softlockup_touch_sync, false);
282 sched_clock_tick();
283 }
284
285 /* Clear the guest paused flag on watchdog reset */
286 kvm_check_and_clear_guest_paused();
287 __touch_watchdog();
288 return HRTIMER_RESTART;
289 }
290
291 /* check for a softlockup
292 * This is done by making sure a high priority task is
293 * being scheduled. The task touches the watchdog to
294 * indicate it is getting cpu time. If it hasn't then
295 * this is a good indication some task is hogging the cpu
296 */
297 duration = is_softlockup(touch_ts);
298 if (unlikely(duration)) {
299 /*
300 * If a virtual machine is stopped by the host it can look to
301 * the watchdog like a soft lockup, check to see if the host
302 * stopped the vm before we issue the warning
303 */
304 if (kvm_check_and_clear_guest_paused())
305 return HRTIMER_RESTART;
306
307 /* only warn once */
308 if (__this_cpu_read(soft_watchdog_warn) == true)
309 return HRTIMER_RESTART;
310
311 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
312 smp_processor_id(), duration,
313 current->comm, task_pid_nr(current));
314 print_modules();
315 print_irqtrace_events(current);
316 if (regs)
317 show_regs(regs);
318 else
319 dump_stack();
320
321 if (softlockup_panic)
322 panic("softlockup: hung tasks");
323 __this_cpu_write(soft_watchdog_warn, true);
324 } else
325 __this_cpu_write(soft_watchdog_warn, false);
326
327 return HRTIMER_RESTART;
328}
329
330
331/*
332 * The watchdog thread - touches the timestamp.
333 */
334static int watchdog(void *unused)
335{
336 struct sched_param param = { .sched_priority = 0 };
337 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
338
339 /* initialize timestamp */
340 __touch_watchdog();
341
342 /* kick off the timer for the hardlockup detector */
343 /* done here because hrtimer_start can only pin to smp_processor_id() */
344 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
345 HRTIMER_MODE_REL_PINNED);
346
347 set_current_state(TASK_INTERRUPTIBLE);
348 /*
349 * Run briefly (kicked by the hrtimer callback function) once every
350 * get_sample_period() seconds (4 seconds by default) to reset the
351 * softlockup timestamp. If this gets delayed for more than
352 * 2*watchdog_thresh seconds then the debug-printout triggers in
353 * watchdog_timer_fn().
354 */
355 while (!kthread_should_stop()) {
356 __touch_watchdog();
357 schedule();
358
359 if (kthread_should_stop())
360 break;
361
362 set_current_state(TASK_INTERRUPTIBLE);
363 }
364 /*
365 * Drop the policy/priority elevation during thread exit to avoid a
366 * scheduling latency spike.
367 */
368 __set_current_state(TASK_RUNNING);
369 sched_setscheduler(current, SCHED_NORMAL, ¶m);
370 return 0;
371}
372
373
374#ifdef CONFIG_HARDLOCKUP_DETECTOR
375/*
376 * People like the simple clean cpu node info on boot.
377 * Reduce the watchdog noise by only printing messages
378 * that are different from what cpu0 displayed.
379 */
380static unsigned long cpu0_err;
381
382static int watchdog_nmi_enable(int cpu)
383{
384 struct perf_event_attr *wd_attr;
385 struct perf_event *event = per_cpu(watchdog_ev, cpu);
386
387 /* is it already setup and enabled? */
388 if (event && event->state > PERF_EVENT_STATE_OFF)
389 goto out;
390
391 /* it is setup but not enabled */
392 if (event != NULL)
393 goto out_enable;
394
395 wd_attr = &wd_hw_attr;
396 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
397
398 /* Try to register using hardware perf events */
399 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
400
401 /* save cpu0 error for future comparision */
402 if (cpu == 0 && IS_ERR(event))
403 cpu0_err = PTR_ERR(event);
404
405 if (!IS_ERR(event)) {
406 /* only print for cpu0 or different than cpu0 */
407 if (cpu == 0 || cpu0_err)
408 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
409 goto out_save;
410 }
411
412 /* skip displaying the same error again */
413 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
414 return PTR_ERR(event);
415
416 /* vary the KERN level based on the returned errno */
417 if (PTR_ERR(event) == -EOPNOTSUPP)
418 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
419 else if (PTR_ERR(event) == -ENOENT)
420 pr_warning("disabled (cpu%i): hardware events not enabled\n",
421 cpu);
422 else
423 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
424 cpu, PTR_ERR(event));
425 return PTR_ERR(event);
426
427 /* success path */
428out_save:
429 per_cpu(watchdog_ev, cpu) = event;
430out_enable:
431 perf_event_enable(per_cpu(watchdog_ev, cpu));
432out:
433 return 0;
434}
435
436static void watchdog_nmi_disable(int cpu)
437{
438 struct perf_event *event = per_cpu(watchdog_ev, cpu);
439
440 if (event) {
441 perf_event_disable(event);
442 per_cpu(watchdog_ev, cpu) = NULL;
443
444 /* should be in cleanup, but blocks oprofile */
445 perf_event_release_kernel(event);
446 }
447 return;
448}
449#else
450static int watchdog_nmi_enable(int cpu) { return 0; }
451static void watchdog_nmi_disable(int cpu) { return; }
452#endif /* CONFIG_HARDLOCKUP_DETECTOR */
453
454/* prepare/enable/disable routines */
455static void watchdog_prepare_cpu(int cpu)
456{
457 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
458
459 WARN_ON(per_cpu(softlockup_watchdog, cpu));
460 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
461 hrtimer->function = watchdog_timer_fn;
462}
463
464static int watchdog_enable(int cpu)
465{
466 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
467 int err = 0;
468
469 /* enable the perf event */
470 err = watchdog_nmi_enable(cpu);
471
472 /* Regardless of err above, fall through and start softlockup */
473
474 /* create the watchdog thread */
475 if (!p) {
476 struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
477 p = kthread_create_on_node(watchdog, NULL, cpu_to_node(cpu), "watchdog/%d", cpu);
478 if (IS_ERR(p)) {
479 pr_err("softlockup watchdog for %i failed\n", cpu);
480 if (!err) {
481 /* if hardlockup hasn't already set this */
482 err = PTR_ERR(p);
483 /* and disable the perf event */
484 watchdog_nmi_disable(cpu);
485 }
486 goto out;
487 }
488 sched_setscheduler(p, SCHED_FIFO, ¶m);
489 kthread_bind(p, cpu);
490 per_cpu(watchdog_touch_ts, cpu) = 0;
491 per_cpu(softlockup_watchdog, cpu) = p;
492 wake_up_process(p);
493 }
494
495out:
496 return err;
497}
498
499static void watchdog_disable(int cpu)
500{
501 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
502 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
503
504 /*
505 * cancel the timer first to stop incrementing the stats
506 * and waking up the kthread
507 */
508 hrtimer_cancel(hrtimer);
509
510 /* disable the perf event */
511 watchdog_nmi_disable(cpu);
512
513 /* stop the watchdog thread */
514 if (p) {
515 per_cpu(softlockup_watchdog, cpu) = NULL;
516 kthread_stop(p);
517 }
518}
519
520/* sysctl functions */
521#ifdef CONFIG_SYSCTL
522static void watchdog_enable_all_cpus(void)
523{
524 int cpu;
525
526 watchdog_enabled = 0;
527
528 for_each_online_cpu(cpu)
529 if (!watchdog_enable(cpu))
530 /* if any cpu succeeds, watchdog is considered
531 enabled for the system */
532 watchdog_enabled = 1;
533
534 if (!watchdog_enabled)
535 pr_err("failed to be enabled on some cpus\n");
536
537}
538
539static void watchdog_disable_all_cpus(void)
540{
541 int cpu;
542
543 for_each_online_cpu(cpu)
544 watchdog_disable(cpu);
545
546 /* if all watchdogs are disabled, then they are disabled for the system */
547 watchdog_enabled = 0;
548}
549
550
551/*
552 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
553 */
554
555int proc_dowatchdog(struct ctl_table *table, int write,
556 void __user *buffer, size_t *lenp, loff_t *ppos)
557{
558 int ret;
559
560 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
561 if (ret || !write)
562 goto out;
563
564 if (watchdog_enabled && watchdog_thresh)
565 watchdog_enable_all_cpus();
566 else
567 watchdog_disable_all_cpus();
568
569out:
570 return ret;
571}
572#endif /* CONFIG_SYSCTL */
573
574
575/*
576 * Create/destroy watchdog threads as CPUs come and go:
577 */
578static int __cpuinit
579cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
580{
581 int hotcpu = (unsigned long)hcpu;
582
583 switch (action) {
584 case CPU_UP_PREPARE:
585 case CPU_UP_PREPARE_FROZEN:
586 watchdog_prepare_cpu(hotcpu);
587 break;
588 case CPU_ONLINE:
589 case CPU_ONLINE_FROZEN:
590 if (watchdog_enabled)
591 watchdog_enable(hotcpu);
592 break;
593#ifdef CONFIG_HOTPLUG_CPU
594 case CPU_UP_CANCELED:
595 case CPU_UP_CANCELED_FROZEN:
596 watchdog_disable(hotcpu);
597 break;
598 case CPU_DEAD:
599 case CPU_DEAD_FROZEN:
600 watchdog_disable(hotcpu);
601 break;
602#endif /* CONFIG_HOTPLUG_CPU */
603 }
604
605 /*
606 * hardlockup and softlockup are not important enough
607 * to block cpu bring up. Just always succeed and
608 * rely on printk output to flag problems.
609 */
610 return NOTIFY_OK;
611}
612
613static struct notifier_block __cpuinitdata cpu_nfb = {
614 .notifier_call = cpu_callback
615};
616
617void __init lockup_detector_init(void)
618{
619 void *cpu = (void *)(long)smp_processor_id();
620 int err;
621
622 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
623 WARN_ON(notifier_to_errno(err));
624
625 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
626 register_cpu_notifier(&cpu_nfb);
627
628 return;
629}
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#include <linux/smpboot.h>
26#include <linux/sched/rt.h>
27
28#include <asm/irq_regs.h>
29#include <linux/kvm_para.h>
30#include <linux/perf_event.h>
31
32int watchdog_user_enabled = 1;
33int __read_mostly watchdog_thresh = 10;
34static int __read_mostly watchdog_running;
35static u64 __read_mostly sample_period;
36
37static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
38static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
39static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
40static DEFINE_PER_CPU(bool, softlockup_touch_sync);
41static DEFINE_PER_CPU(bool, soft_watchdog_warn);
42static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
43static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
44#ifdef CONFIG_HARDLOCKUP_DETECTOR
45static DEFINE_PER_CPU(bool, hard_watchdog_warn);
46static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
47static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
48static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
49#endif
50
51/* boot commands */
52/*
53 * Should we panic when a soft-lockup or hard-lockup occurs:
54 */
55#ifdef CONFIG_HARDLOCKUP_DETECTOR
56static int hardlockup_panic =
57 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
58
59static int __init hardlockup_panic_setup(char *str)
60{
61 if (!strncmp(str, "panic", 5))
62 hardlockup_panic = 1;
63 else if (!strncmp(str, "nopanic", 7))
64 hardlockup_panic = 0;
65 else if (!strncmp(str, "0", 1))
66 watchdog_user_enabled = 0;
67 return 1;
68}
69__setup("nmi_watchdog=", hardlockup_panic_setup);
70#endif
71
72unsigned int __read_mostly softlockup_panic =
73 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
74
75static int __init softlockup_panic_setup(char *str)
76{
77 softlockup_panic = simple_strtoul(str, NULL, 0);
78
79 return 1;
80}
81__setup("softlockup_panic=", softlockup_panic_setup);
82
83static int __init nowatchdog_setup(char *str)
84{
85 watchdog_user_enabled = 0;
86 return 1;
87}
88__setup("nowatchdog", nowatchdog_setup);
89
90/* deprecated */
91static int __init nosoftlockup_setup(char *str)
92{
93 watchdog_user_enabled = 0;
94 return 1;
95}
96__setup("nosoftlockup", nosoftlockup_setup);
97/* */
98
99/*
100 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
101 * lockups can have false positives under extreme conditions. So we generally
102 * want a higher threshold for soft lockups than for hard lockups. So we couple
103 * the thresholds with a factor: we make the soft threshold twice the amount of
104 * time the hard threshold is.
105 */
106static int get_softlockup_thresh(void)
107{
108 return watchdog_thresh * 2;
109}
110
111/*
112 * Returns seconds, approximately. We don't need nanosecond
113 * resolution, and we don't need to waste time with a big divide when
114 * 2^30ns == 1.074s.
115 */
116static unsigned long get_timestamp(void)
117{
118 return local_clock() >> 30LL; /* 2^30 ~= 10^9 */
119}
120
121static void set_sample_period(void)
122{
123 /*
124 * convert watchdog_thresh from seconds to ns
125 * the divide by 5 is to give hrtimer several chances (two
126 * or three with the current relation between the soft
127 * and hard thresholds) to increment before the
128 * hardlockup detector generates a warning
129 */
130 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
131}
132
133/* Commands for resetting the watchdog */
134static void __touch_watchdog(void)
135{
136 __this_cpu_write(watchdog_touch_ts, get_timestamp());
137}
138
139void touch_softlockup_watchdog(void)
140{
141 /*
142 * Preemption can be enabled. It doesn't matter which CPU's timestamp
143 * gets zeroed here, so use the raw_ operation.
144 */
145 raw_cpu_write(watchdog_touch_ts, 0);
146}
147EXPORT_SYMBOL(touch_softlockup_watchdog);
148
149void touch_all_softlockup_watchdogs(void)
150{
151 int cpu;
152
153 /*
154 * this is done lockless
155 * do we care if a 0 races with a timestamp?
156 * all it means is the softlock check starts one cycle later
157 */
158 for_each_online_cpu(cpu)
159 per_cpu(watchdog_touch_ts, cpu) = 0;
160}
161
162#ifdef CONFIG_HARDLOCKUP_DETECTOR
163void touch_nmi_watchdog(void)
164{
165 /*
166 * Using __raw here because some code paths have
167 * preemption enabled. If preemption is enabled
168 * then interrupts should be enabled too, in which
169 * case we shouldn't have to worry about the watchdog
170 * going off.
171 */
172 __raw_get_cpu_var(watchdog_nmi_touch) = true;
173 touch_softlockup_watchdog();
174}
175EXPORT_SYMBOL(touch_nmi_watchdog);
176
177#endif
178
179void touch_softlockup_watchdog_sync(void)
180{
181 __raw_get_cpu_var(softlockup_touch_sync) = true;
182 __raw_get_cpu_var(watchdog_touch_ts) = 0;
183}
184
185#ifdef CONFIG_HARDLOCKUP_DETECTOR
186/* watchdog detector functions */
187static int is_hardlockup(void)
188{
189 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
190
191 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
192 return 1;
193
194 __this_cpu_write(hrtimer_interrupts_saved, hrint);
195 return 0;
196}
197#endif
198
199static int is_softlockup(unsigned long touch_ts)
200{
201 unsigned long now = get_timestamp();
202
203 /* Warn about unreasonable delays: */
204 if (time_after(now, touch_ts + get_softlockup_thresh()))
205 return now - touch_ts;
206
207 return 0;
208}
209
210#ifdef CONFIG_HARDLOCKUP_DETECTOR
211
212static struct perf_event_attr wd_hw_attr = {
213 .type = PERF_TYPE_HARDWARE,
214 .config = PERF_COUNT_HW_CPU_CYCLES,
215 .size = sizeof(struct perf_event_attr),
216 .pinned = 1,
217 .disabled = 1,
218};
219
220/* Callback function for perf event subsystem */
221static void watchdog_overflow_callback(struct perf_event *event,
222 struct perf_sample_data *data,
223 struct pt_regs *regs)
224{
225 /* Ensure the watchdog never gets throttled */
226 event->hw.interrupts = 0;
227
228 if (__this_cpu_read(watchdog_nmi_touch) == true) {
229 __this_cpu_write(watchdog_nmi_touch, false);
230 return;
231 }
232
233 /* check for a hardlockup
234 * This is done by making sure our timer interrupt
235 * is incrementing. The timer interrupt should have
236 * fired multiple times before we overflow'd. If it hasn't
237 * then this is a good indication the cpu is stuck
238 */
239 if (is_hardlockup()) {
240 int this_cpu = smp_processor_id();
241
242 /* only print hardlockups once */
243 if (__this_cpu_read(hard_watchdog_warn) == true)
244 return;
245
246 if (hardlockup_panic)
247 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
248 else
249 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
250
251 __this_cpu_write(hard_watchdog_warn, true);
252 return;
253 }
254
255 __this_cpu_write(hard_watchdog_warn, false);
256 return;
257}
258#endif /* CONFIG_HARDLOCKUP_DETECTOR */
259
260static void watchdog_interrupt_count(void)
261{
262 __this_cpu_inc(hrtimer_interrupts);
263}
264
265static int watchdog_nmi_enable(unsigned int cpu);
266static void watchdog_nmi_disable(unsigned int cpu);
267
268/* watchdog kicker functions */
269static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
270{
271 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
272 struct pt_regs *regs = get_irq_regs();
273 int duration;
274
275 /* kick the hardlockup detector */
276 watchdog_interrupt_count();
277
278 /* kick the softlockup detector */
279 wake_up_process(__this_cpu_read(softlockup_watchdog));
280
281 /* .. and repeat */
282 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
283
284 if (touch_ts == 0) {
285 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
286 /*
287 * If the time stamp was touched atomically
288 * make sure the scheduler tick is up to date.
289 */
290 __this_cpu_write(softlockup_touch_sync, false);
291 sched_clock_tick();
292 }
293
294 /* Clear the guest paused flag on watchdog reset */
295 kvm_check_and_clear_guest_paused();
296 __touch_watchdog();
297 return HRTIMER_RESTART;
298 }
299
300 /* check for a softlockup
301 * This is done by making sure a high priority task is
302 * being scheduled. The task touches the watchdog to
303 * indicate it is getting cpu time. If it hasn't then
304 * this is a good indication some task is hogging the cpu
305 */
306 duration = is_softlockup(touch_ts);
307 if (unlikely(duration)) {
308 /*
309 * If a virtual machine is stopped by the host it can look to
310 * the watchdog like a soft lockup, check to see if the host
311 * stopped the vm before we issue the warning
312 */
313 if (kvm_check_and_clear_guest_paused())
314 return HRTIMER_RESTART;
315
316 /* only warn once */
317 if (__this_cpu_read(soft_watchdog_warn) == true)
318 return HRTIMER_RESTART;
319
320 printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
321 smp_processor_id(), duration,
322 current->comm, task_pid_nr(current));
323 print_modules();
324 print_irqtrace_events(current);
325 if (regs)
326 show_regs(regs);
327 else
328 dump_stack();
329
330 if (softlockup_panic)
331 panic("softlockup: hung tasks");
332 __this_cpu_write(soft_watchdog_warn, true);
333 } else
334 __this_cpu_write(soft_watchdog_warn, false);
335
336 return HRTIMER_RESTART;
337}
338
339static void watchdog_set_prio(unsigned int policy, unsigned int prio)
340{
341 struct sched_param param = { .sched_priority = prio };
342
343 sched_setscheduler(current, policy, ¶m);
344}
345
346static void watchdog_enable(unsigned int cpu)
347{
348 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
349
350 /* kick off the timer for the hardlockup detector */
351 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
352 hrtimer->function = watchdog_timer_fn;
353
354 /* Enable the perf event */
355 watchdog_nmi_enable(cpu);
356
357 /* done here because hrtimer_start can only pin to smp_processor_id() */
358 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
359 HRTIMER_MODE_REL_PINNED);
360
361 /* initialize timestamp */
362 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
363 __touch_watchdog();
364}
365
366static void watchdog_disable(unsigned int cpu)
367{
368 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
369
370 watchdog_set_prio(SCHED_NORMAL, 0);
371 hrtimer_cancel(hrtimer);
372 /* disable the perf event */
373 watchdog_nmi_disable(cpu);
374}
375
376static void watchdog_cleanup(unsigned int cpu, bool online)
377{
378 watchdog_disable(cpu);
379}
380
381static int watchdog_should_run(unsigned int cpu)
382{
383 return __this_cpu_read(hrtimer_interrupts) !=
384 __this_cpu_read(soft_lockup_hrtimer_cnt);
385}
386
387/*
388 * The watchdog thread function - touches the timestamp.
389 *
390 * It only runs once every sample_period seconds (4 seconds by
391 * default) to reset the softlockup timestamp. If this gets delayed
392 * for more than 2*watchdog_thresh seconds then the debug-printout
393 * triggers in watchdog_timer_fn().
394 */
395static void watchdog(unsigned int cpu)
396{
397 __this_cpu_write(soft_lockup_hrtimer_cnt,
398 __this_cpu_read(hrtimer_interrupts));
399 __touch_watchdog();
400}
401
402#ifdef CONFIG_HARDLOCKUP_DETECTOR
403/*
404 * People like the simple clean cpu node info on boot.
405 * Reduce the watchdog noise by only printing messages
406 * that are different from what cpu0 displayed.
407 */
408static unsigned long cpu0_err;
409
410static int watchdog_nmi_enable(unsigned int cpu)
411{
412 struct perf_event_attr *wd_attr;
413 struct perf_event *event = per_cpu(watchdog_ev, cpu);
414
415 /* is it already setup and enabled? */
416 if (event && event->state > PERF_EVENT_STATE_OFF)
417 goto out;
418
419 /* it is setup but not enabled */
420 if (event != NULL)
421 goto out_enable;
422
423 wd_attr = &wd_hw_attr;
424 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
425
426 /* Try to register using hardware perf events */
427 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
428
429 /* save cpu0 error for future comparision */
430 if (cpu == 0 && IS_ERR(event))
431 cpu0_err = PTR_ERR(event);
432
433 if (!IS_ERR(event)) {
434 /* only print for cpu0 or different than cpu0 */
435 if (cpu == 0 || cpu0_err)
436 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
437 goto out_save;
438 }
439
440 /* skip displaying the same error again */
441 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
442 return PTR_ERR(event);
443
444 /* vary the KERN level based on the returned errno */
445 if (PTR_ERR(event) == -EOPNOTSUPP)
446 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
447 else if (PTR_ERR(event) == -ENOENT)
448 pr_warning("disabled (cpu%i): hardware events not enabled\n",
449 cpu);
450 else
451 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
452 cpu, PTR_ERR(event));
453 return PTR_ERR(event);
454
455 /* success path */
456out_save:
457 per_cpu(watchdog_ev, cpu) = event;
458out_enable:
459 perf_event_enable(per_cpu(watchdog_ev, cpu));
460out:
461 return 0;
462}
463
464static void watchdog_nmi_disable(unsigned int cpu)
465{
466 struct perf_event *event = per_cpu(watchdog_ev, cpu);
467
468 if (event) {
469 perf_event_disable(event);
470 per_cpu(watchdog_ev, cpu) = NULL;
471
472 /* should be in cleanup, but blocks oprofile */
473 perf_event_release_kernel(event);
474 }
475 return;
476}
477#else
478static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
479static void watchdog_nmi_disable(unsigned int cpu) { return; }
480#endif /* CONFIG_HARDLOCKUP_DETECTOR */
481
482static struct smp_hotplug_thread watchdog_threads = {
483 .store = &softlockup_watchdog,
484 .thread_should_run = watchdog_should_run,
485 .thread_fn = watchdog,
486 .thread_comm = "watchdog/%u",
487 .setup = watchdog_enable,
488 .cleanup = watchdog_cleanup,
489 .park = watchdog_disable,
490 .unpark = watchdog_enable,
491};
492
493static void restart_watchdog_hrtimer(void *info)
494{
495 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
496 int ret;
497
498 /*
499 * No need to cancel and restart hrtimer if it is currently executing
500 * because it will reprogram itself with the new period now.
501 * We should never see it unqueued here because we are running per-cpu
502 * with interrupts disabled.
503 */
504 ret = hrtimer_try_to_cancel(hrtimer);
505 if (ret == 1)
506 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
507 HRTIMER_MODE_REL_PINNED);
508}
509
510static void update_timers(int cpu)
511{
512 /*
513 * Make sure that perf event counter will adopt to a new
514 * sampling period. Updating the sampling period directly would
515 * be much nicer but we do not have an API for that now so
516 * let's use a big hammer.
517 * Hrtimer will adopt the new period on the next tick but this
518 * might be late already so we have to restart the timer as well.
519 */
520 watchdog_nmi_disable(cpu);
521 smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
522 watchdog_nmi_enable(cpu);
523}
524
525static void update_timers_all_cpus(void)
526{
527 int cpu;
528
529 get_online_cpus();
530 preempt_disable();
531 for_each_online_cpu(cpu)
532 update_timers(cpu);
533 preempt_enable();
534 put_online_cpus();
535}
536
537static int watchdog_enable_all_cpus(bool sample_period_changed)
538{
539 int err = 0;
540
541 if (!watchdog_running) {
542 err = smpboot_register_percpu_thread(&watchdog_threads);
543 if (err)
544 pr_err("Failed to create watchdog threads, disabled\n");
545 else
546 watchdog_running = 1;
547 } else if (sample_period_changed) {
548 update_timers_all_cpus();
549 }
550
551 return err;
552}
553
554/* prepare/enable/disable routines */
555/* sysctl functions */
556#ifdef CONFIG_SYSCTL
557static void watchdog_disable_all_cpus(void)
558{
559 if (watchdog_running) {
560 watchdog_running = 0;
561 smpboot_unregister_percpu_thread(&watchdog_threads);
562 }
563}
564
565/*
566 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
567 */
568
569int proc_dowatchdog(struct ctl_table *table, int write,
570 void __user *buffer, size_t *lenp, loff_t *ppos)
571{
572 int err, old_thresh, old_enabled;
573 static DEFINE_MUTEX(watchdog_proc_mutex);
574
575 mutex_lock(&watchdog_proc_mutex);
576 old_thresh = ACCESS_ONCE(watchdog_thresh);
577 old_enabled = ACCESS_ONCE(watchdog_user_enabled);
578
579 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
580 if (err || !write)
581 goto out;
582
583 set_sample_period();
584 /*
585 * Watchdog threads shouldn't be enabled if they are
586 * disabled. The 'watchdog_running' variable check in
587 * watchdog_*_all_cpus() function takes care of this.
588 */
589 if (watchdog_user_enabled && watchdog_thresh)
590 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
591 else
592 watchdog_disable_all_cpus();
593
594 /* Restore old values on failure */
595 if (err) {
596 watchdog_thresh = old_thresh;
597 watchdog_user_enabled = old_enabled;
598 }
599out:
600 mutex_unlock(&watchdog_proc_mutex);
601 return err;
602}
603#endif /* CONFIG_SYSCTL */
604
605void __init lockup_detector_init(void)
606{
607 set_sample_period();
608
609 if (watchdog_user_enabled)
610 watchdog_enable_all_cpus(false);
611}