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/module.h>
19 #include <linux/sysctl.h>
20 #include <linux/smpboot.h>
21 #include <linux/sched/rt.h>
22
23 #include <asm/irq_regs.h>
24 #include <linux/kvm_para.h>
25 #include <linux/perf_event.h>
26
27 int watchdog_user_enabled = 1;
28 int __read_mostly watchdog_thresh = 10;
29 #ifdef CONFIG_SMP
30 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
31 #else
32 #define sysctl_softlockup_all_cpu_backtrace 0
33 #endif
34
35 static int __read_mostly watchdog_running;
36 static u64 __read_mostly sample_period;
37
38 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
39 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
40 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
41 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
42 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
43 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
44 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
45 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
46 #ifdef CONFIG_HARDLOCKUP_DETECTOR
47 static DEFINE_PER_CPU(bool, hard_watchdog_warn);
48 static DEFINE_PER_CPU(bool, watchdog_nmi_touch);
49 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
50 #endif
51 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
52 static cpumask_t __read_mostly watchdog_cpus;
53 #endif
54 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
55 static DEFINE_PER_CPU(struct perf_event *, watchdog_ev);
56 #endif
57 static unsigned long soft_lockup_nmi_warn;
58
59 /* boot commands */
60 /*
61 * Should we panic when a soft-lockup or hard-lockup occurs:
62 */
63 #ifdef CONFIG_HARDLOCKUP_DETECTOR
64 static int hardlockup_panic =
65 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
66
67 static bool hardlockup_detector_enabled = true;
68 /*
69 * We may not want to enable hard lockup detection by default in all cases,
70 * for example when running the kernel as a guest on a hypervisor. In these
71 * cases this function can be called to disable hard lockup detection. This
72 * function should only be executed once by the boot processor before the
73 * kernel command line parameters are parsed, because otherwise it is not
74 * possible to override this in hardlockup_panic_setup().
75 */
watchdog_enable_hardlockup_detector(bool val)76 void watchdog_enable_hardlockup_detector(bool val)
77 {
78 hardlockup_detector_enabled = val;
79 }
80
watchdog_hardlockup_detector_is_enabled(void)81 bool watchdog_hardlockup_detector_is_enabled(void)
82 {
83 return hardlockup_detector_enabled;
84 }
85
hardlockup_panic_setup(char * str)86 static int __init hardlockup_panic_setup(char *str)
87 {
88 if (!strncmp(str, "panic", 5))
89 hardlockup_panic = 1;
90 else if (!strncmp(str, "nopanic", 7))
91 hardlockup_panic = 0;
92 else if (!strncmp(str, "0", 1))
93 watchdog_user_enabled = 0;
94 else if (!strncmp(str, "1", 1) || !strncmp(str, "2", 1)) {
95 /*
96 * Setting 'nmi_watchdog=1' or 'nmi_watchdog=2' (legacy option)
97 * has the same effect.
98 */
99 watchdog_user_enabled = 1;
100 watchdog_enable_hardlockup_detector(true);
101 }
102 return 1;
103 }
104 __setup("nmi_watchdog=", hardlockup_panic_setup);
105 #endif
106
107 unsigned int __read_mostly softlockup_panic =
108 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
109
softlockup_panic_setup(char * str)110 static int __init softlockup_panic_setup(char *str)
111 {
112 softlockup_panic = simple_strtoul(str, NULL, 0);
113
114 return 1;
115 }
116 __setup("softlockup_panic=", softlockup_panic_setup);
117
nowatchdog_setup(char * str)118 static int __init nowatchdog_setup(char *str)
119 {
120 watchdog_user_enabled = 0;
121 return 1;
122 }
123 __setup("nowatchdog", nowatchdog_setup);
124
125 /* deprecated */
nosoftlockup_setup(char * str)126 static int __init nosoftlockup_setup(char *str)
127 {
128 watchdog_user_enabled = 0;
129 return 1;
130 }
131 __setup("nosoftlockup", nosoftlockup_setup);
132 /* */
133 #ifdef CONFIG_SMP
softlockup_all_cpu_backtrace_setup(char * str)134 static int __init softlockup_all_cpu_backtrace_setup(char *str)
135 {
136 sysctl_softlockup_all_cpu_backtrace =
137 !!simple_strtol(str, NULL, 0);
138 return 1;
139 }
140 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
141 #endif
142
143 /*
144 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
145 * lockups can have false positives under extreme conditions. So we generally
146 * want a higher threshold for soft lockups than for hard lockups. So we couple
147 * the thresholds with a factor: we make the soft threshold twice the amount of
148 * time the hard threshold is.
149 */
get_softlockup_thresh(void)150 static int get_softlockup_thresh(void)
151 {
152 return watchdog_thresh * 2;
153 }
154
155 /*
156 * Returns seconds, approximately. We don't need nanosecond
157 * resolution, and we don't need to waste time with a big divide when
158 * 2^30ns == 1.074s.
159 */
get_timestamp(void)160 static unsigned long get_timestamp(void)
161 {
162 return local_clock() >> 30LL; /* 2^30 ~= 10^9 */
163 }
164
set_sample_period(void)165 static void set_sample_period(void)
166 {
167 /*
168 * convert watchdog_thresh from seconds to ns
169 * the divide by 5 is to give hrtimer several chances (two
170 * or three with the current relation between the soft
171 * and hard thresholds) to increment before the
172 * hardlockup detector generates a warning
173 */
174 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
175 }
176
177 /* Commands for resetting the watchdog */
__touch_watchdog(void)178 static void __touch_watchdog(void)
179 {
180 __this_cpu_write(watchdog_touch_ts, get_timestamp());
181 }
182
touch_softlockup_watchdog(void)183 void touch_softlockup_watchdog(void)
184 {
185 /*
186 * Preemption can be enabled. It doesn't matter which CPU's timestamp
187 * gets zeroed here, so use the raw_ operation.
188 */
189 raw_cpu_write(watchdog_touch_ts, 0);
190 }
191 EXPORT_SYMBOL(touch_softlockup_watchdog);
192
touch_all_softlockup_watchdogs(void)193 void touch_all_softlockup_watchdogs(void)
194 {
195 int cpu;
196
197 /*
198 * this is done lockless
199 * do we care if a 0 races with a timestamp?
200 * all it means is the softlock check starts one cycle later
201 */
202 for_each_online_cpu(cpu)
203 per_cpu(watchdog_touch_ts, cpu) = 0;
204 }
205
206 #ifdef CONFIG_HARDLOCKUP_DETECTOR
touch_nmi_watchdog(void)207 void touch_nmi_watchdog(void)
208 {
209 /*
210 * Using __raw here because some code paths have
211 * preemption enabled. If preemption is enabled
212 * then interrupts should be enabled too, in which
213 * case we shouldn't have to worry about the watchdog
214 * going off.
215 */
216 raw_cpu_write(watchdog_nmi_touch, true);
217 touch_softlockup_watchdog();
218 }
219 EXPORT_SYMBOL(touch_nmi_watchdog);
220
221 #endif
222
touch_softlockup_watchdog_sync(void)223 void touch_softlockup_watchdog_sync(void)
224 {
225 __this_cpu_write(softlockup_touch_sync, true);
226 __this_cpu_write(watchdog_touch_ts, 0);
227 }
228
229 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
230 /* watchdog detector functions */
is_hardlockup(void)231 static int is_hardlockup(void)
232 {
233 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
234
235 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
236 return 1;
237
238 __this_cpu_write(hrtimer_interrupts_saved, hrint);
239 return 0;
240 }
241 #endif
242
243 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
watchdog_next_cpu(unsigned int cpu)244 static unsigned int watchdog_next_cpu(unsigned int cpu)
245 {
246 cpumask_t cpus = watchdog_cpus;
247 unsigned int next_cpu;
248
249 next_cpu = cpumask_next(cpu, &cpus);
250 if (next_cpu >= nr_cpu_ids)
251 next_cpu = cpumask_first(&cpus);
252
253 if (next_cpu == cpu)
254 return nr_cpu_ids;
255
256 return next_cpu;
257 }
258
is_hardlockup_other_cpu(unsigned int cpu)259 static int is_hardlockup_other_cpu(unsigned int cpu)
260 {
261 unsigned long hrint = per_cpu(hrtimer_interrupts, cpu);
262
263 if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
264 return 1;
265
266 per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
267 return 0;
268 }
269
watchdog_check_hardlockup_other_cpu(void)270 static void watchdog_check_hardlockup_other_cpu(void)
271 {
272 unsigned int next_cpu;
273
274 /*
275 * Test for hardlockups every 3 samples. The sample period is
276 * watchdog_thresh * 2 / 5, so 3 samples gets us back to slightly over
277 * watchdog_thresh (over by 20%).
278 */
279 if (__this_cpu_read(hrtimer_interrupts) % 3 != 0)
280 return;
281
282 /* check for a hardlockup on the next cpu */
283 next_cpu = watchdog_next_cpu(smp_processor_id());
284 if (next_cpu >= nr_cpu_ids)
285 return;
286
287 smp_rmb();
288
289 if (per_cpu(watchdog_nmi_touch, next_cpu) == true) {
290 per_cpu(watchdog_nmi_touch, next_cpu) = false;
291 return;
292 }
293
294 if (is_hardlockup_other_cpu(next_cpu)) {
295 /* only warn once */
296 if (per_cpu(hard_watchdog_warn, next_cpu) == true)
297 return;
298
299 if (hardlockup_panic)
300 panic("Watchdog detected hard LOCKUP on cpu %u", next_cpu);
301 else
302 WARN(1, "Watchdog detected hard LOCKUP on cpu %u", next_cpu);
303
304 per_cpu(hard_watchdog_warn, next_cpu) = true;
305 } else {
306 per_cpu(hard_watchdog_warn, next_cpu) = false;
307 }
308 }
309 #else
watchdog_check_hardlockup_other_cpu(void)310 static inline void watchdog_check_hardlockup_other_cpu(void) { return; }
311 #endif
312
is_softlockup(unsigned long touch_ts)313 static int is_softlockup(unsigned long touch_ts)
314 {
315 unsigned long now = get_timestamp();
316
317 /* Warn about unreasonable delays: */
318 if (time_after(now, touch_ts + get_softlockup_thresh()))
319 return now - touch_ts;
320
321 return 0;
322 }
323
324 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
325
326 static struct perf_event_attr wd_hw_attr = {
327 .type = PERF_TYPE_HARDWARE,
328 .config = PERF_COUNT_HW_CPU_CYCLES,
329 .size = sizeof(struct perf_event_attr),
330 .pinned = 1,
331 .disabled = 1,
332 };
333
334 /* Callback function for perf event subsystem */
watchdog_overflow_callback(struct perf_event * event,struct perf_sample_data * data,struct pt_regs * regs)335 static void watchdog_overflow_callback(struct perf_event *event,
336 struct perf_sample_data *data,
337 struct pt_regs *regs)
338 {
339 /* Ensure the watchdog never gets throttled */
340 event->hw.interrupts = 0;
341
342 if (__this_cpu_read(watchdog_nmi_touch) == true) {
343 __this_cpu_write(watchdog_nmi_touch, false);
344 return;
345 }
346
347 /* check for a hardlockup
348 * This is done by making sure our timer interrupt
349 * is incrementing. The timer interrupt should have
350 * fired multiple times before we overflow'd. If it hasn't
351 * then this is a good indication the cpu is stuck
352 */
353 if (is_hardlockup()) {
354 int this_cpu = smp_processor_id();
355
356 /* only print hardlockups once */
357 if (__this_cpu_read(hard_watchdog_warn) == true)
358 return;
359
360 if (hardlockup_panic)
361 panic("Watchdog detected hard LOCKUP on cpu %d",
362 this_cpu);
363 else
364 WARN(1, "Watchdog detected hard LOCKUP on cpu %d",
365 this_cpu);
366
367 __this_cpu_write(hard_watchdog_warn, true);
368 return;
369 }
370
371 __this_cpu_write(hard_watchdog_warn, false);
372 return;
373 }
374 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
375
watchdog_interrupt_count(void)376 static void watchdog_interrupt_count(void)
377 {
378 __this_cpu_inc(hrtimer_interrupts);
379 }
380
381 static int watchdog_nmi_enable(unsigned int cpu);
382 static void watchdog_nmi_disable(unsigned int cpu);
383
384 /* watchdog kicker functions */
watchdog_timer_fn(struct hrtimer * hrtimer)385 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
386 {
387 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
388 struct pt_regs *regs = get_irq_regs();
389 int duration;
390 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
391
392 /* kick the hardlockup detector */
393 watchdog_interrupt_count();
394
395 /* test for hardlockups on the next cpu */
396 watchdog_check_hardlockup_other_cpu();
397
398 /* kick the softlockup detector */
399 wake_up_process(__this_cpu_read(softlockup_watchdog));
400
401 /* .. and repeat */
402 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
403
404 if (touch_ts == 0) {
405 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
406 /*
407 * If the time stamp was touched atomically
408 * make sure the scheduler tick is up to date.
409 */
410 __this_cpu_write(softlockup_touch_sync, false);
411 sched_clock_tick();
412 }
413
414 /* Clear the guest paused flag on watchdog reset */
415 kvm_check_and_clear_guest_paused();
416 __touch_watchdog();
417 return HRTIMER_RESTART;
418 }
419
420 /* check for a softlockup
421 * This is done by making sure a high priority task is
422 * being scheduled. The task touches the watchdog to
423 * indicate it is getting cpu time. If it hasn't then
424 * this is a good indication some task is hogging the cpu
425 */
426 duration = is_softlockup(touch_ts);
427 if (unlikely(duration)) {
428 /*
429 * If a virtual machine is stopped by the host it can look to
430 * the watchdog like a soft lockup, check to see if the host
431 * stopped the vm before we issue the warning
432 */
433 if (kvm_check_and_clear_guest_paused())
434 return HRTIMER_RESTART;
435
436 /* only warn once */
437 if (__this_cpu_read(soft_watchdog_warn) == true) {
438 /*
439 * When multiple processes are causing softlockups the
440 * softlockup detector only warns on the first one
441 * because the code relies on a full quiet cycle to
442 * re-arm. The second process prevents the quiet cycle
443 * and never gets reported. Use task pointers to detect
444 * this.
445 */
446 if (__this_cpu_read(softlockup_task_ptr_saved) !=
447 current) {
448 __this_cpu_write(soft_watchdog_warn, false);
449 __touch_watchdog();
450 }
451 return HRTIMER_RESTART;
452 }
453
454 if (softlockup_all_cpu_backtrace) {
455 /* Prevent multiple soft-lockup reports if one cpu is already
456 * engaged in dumping cpu back traces
457 */
458 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
459 /* Someone else will report us. Let's give up */
460 __this_cpu_write(soft_watchdog_warn, true);
461 return HRTIMER_RESTART;
462 }
463 }
464
465 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
466 smp_processor_id(), duration,
467 current->comm, task_pid_nr(current));
468 __this_cpu_write(softlockup_task_ptr_saved, current);
469 print_modules();
470 print_irqtrace_events(current);
471 if (regs)
472 show_regs(regs);
473 else
474 dump_stack();
475
476 if (softlockup_all_cpu_backtrace) {
477 /* Avoid generating two back traces for current
478 * given that one is already made above
479 */
480 trigger_allbutself_cpu_backtrace();
481
482 clear_bit(0, &soft_lockup_nmi_warn);
483 /* Barrier to sync with other cpus */
484 smp_mb__after_atomic();
485 }
486
487 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
488 if (softlockup_panic)
489 panic("softlockup: hung tasks");
490 __this_cpu_write(soft_watchdog_warn, true);
491 } else
492 __this_cpu_write(soft_watchdog_warn, false);
493
494 return HRTIMER_RESTART;
495 }
496
watchdog_set_prio(unsigned int policy,unsigned int prio)497 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
498 {
499 struct sched_param param = { .sched_priority = prio };
500
501 sched_setscheduler(current, policy, ¶m);
502 }
503
watchdog_enable(unsigned int cpu)504 static void watchdog_enable(unsigned int cpu)
505 {
506 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
507
508 /* kick off the timer for the hardlockup detector */
509 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
510 hrtimer->function = watchdog_timer_fn;
511
512 /* Enable the perf event */
513 watchdog_nmi_enable(cpu);
514
515 /* done here because hrtimer_start can only pin to smp_processor_id() */
516 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
517 HRTIMER_MODE_REL_PINNED);
518
519 /* initialize timestamp */
520 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
521 __touch_watchdog();
522 }
523
watchdog_disable(unsigned int cpu)524 static void watchdog_disable(unsigned int cpu)
525 {
526 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
527
528 watchdog_set_prio(SCHED_NORMAL, 0);
529 hrtimer_cancel(hrtimer);
530 /* disable the perf event */
531 watchdog_nmi_disable(cpu);
532 }
533
watchdog_cleanup(unsigned int cpu,bool online)534 static void watchdog_cleanup(unsigned int cpu, bool online)
535 {
536 watchdog_disable(cpu);
537 }
538
watchdog_should_run(unsigned int cpu)539 static int watchdog_should_run(unsigned int cpu)
540 {
541 return __this_cpu_read(hrtimer_interrupts) !=
542 __this_cpu_read(soft_lockup_hrtimer_cnt);
543 }
544
545 /*
546 * The watchdog thread function - touches the timestamp.
547 *
548 * It only runs once every sample_period seconds (4 seconds by
549 * default) to reset the softlockup timestamp. If this gets delayed
550 * for more than 2*watchdog_thresh seconds then the debug-printout
551 * triggers in watchdog_timer_fn().
552 */
watchdog(unsigned int cpu)553 static void watchdog(unsigned int cpu)
554 {
555 __this_cpu_write(soft_lockup_hrtimer_cnt,
556 __this_cpu_read(hrtimer_interrupts));
557 __touch_watchdog();
558 }
559
560 #ifdef CONFIG_HARDLOCKUP_DETECTOR_NMI
561 /*
562 * People like the simple clean cpu node info on boot.
563 * Reduce the watchdog noise by only printing messages
564 * that are different from what cpu0 displayed.
565 */
566 static unsigned long cpu0_err;
567
watchdog_nmi_enable(unsigned int cpu)568 static int watchdog_nmi_enable(unsigned int cpu)
569 {
570 struct perf_event_attr *wd_attr;
571 struct perf_event *event = per_cpu(watchdog_ev, cpu);
572
573 /*
574 * Some kernels need to default hard lockup detection to
575 * 'disabled', for example a guest on a hypervisor.
576 */
577 if (!watchdog_hardlockup_detector_is_enabled()) {
578 event = ERR_PTR(-ENOENT);
579 goto handle_err;
580 }
581
582 /* is it already setup and enabled? */
583 if (event && event->state > PERF_EVENT_STATE_OFF)
584 goto out;
585
586 /* it is setup but not enabled */
587 if (event != NULL)
588 goto out_enable;
589
590 wd_attr = &wd_hw_attr;
591 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
592
593 /* Try to register using hardware perf events */
594 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
595
596 handle_err:
597 /* save cpu0 error for future comparision */
598 if (cpu == 0 && IS_ERR(event))
599 cpu0_err = PTR_ERR(event);
600
601 if (!IS_ERR(event)) {
602 /* only print for cpu0 or different than cpu0 */
603 if (cpu == 0 || cpu0_err)
604 pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n");
605 goto out_save;
606 }
607
608 /* skip displaying the same error again */
609 if (cpu > 0 && (PTR_ERR(event) == cpu0_err))
610 return PTR_ERR(event);
611
612 /* vary the KERN level based on the returned errno */
613 if (PTR_ERR(event) == -EOPNOTSUPP)
614 pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
615 else if (PTR_ERR(event) == -ENOENT)
616 pr_warn("disabled (cpu%i): hardware events not enabled\n",
617 cpu);
618 else
619 pr_err("disabled (cpu%i): unable to create perf event: %ld\n",
620 cpu, PTR_ERR(event));
621 return PTR_ERR(event);
622
623 /* success path */
624 out_save:
625 per_cpu(watchdog_ev, cpu) = event;
626 out_enable:
627 perf_event_enable(per_cpu(watchdog_ev, cpu));
628 out:
629 return 0;
630 }
631
watchdog_nmi_disable(unsigned int cpu)632 static void watchdog_nmi_disable(unsigned int cpu)
633 {
634 struct perf_event *event = per_cpu(watchdog_ev, cpu);
635
636 if (event) {
637 perf_event_disable(event);
638 per_cpu(watchdog_ev, cpu) = NULL;
639
640 /* should be in cleanup, but blocks oprofile */
641 perf_event_release_kernel(event);
642 }
643 if (cpu == 0) {
644 /* watchdog_nmi_enable() expects this to be zero initially. */
645 cpu0_err = 0;
646 }
647 }
648 #else
649 #ifdef CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU
watchdog_nmi_enable(unsigned int cpu)650 static int watchdog_nmi_enable(unsigned int cpu)
651 {
652 /*
653 * The new cpu will be marked online before the first hrtimer interrupt
654 * runs on it. If another cpu tests for a hardlockup on the new cpu
655 * before it has run its first hrtimer, it will get a false positive.
656 * Touch the watchdog on the new cpu to delay the first check for at
657 * least 3 sampling periods to guarantee one hrtimer has run on the new
658 * cpu.
659 */
660 per_cpu(watchdog_nmi_touch, cpu) = true;
661 smp_wmb();
662 cpumask_set_cpu(cpu, &watchdog_cpus);
663 return 0;
664 }
665
watchdog_nmi_disable(unsigned int cpu)666 static void watchdog_nmi_disable(unsigned int cpu)
667 {
668 unsigned int next_cpu = watchdog_next_cpu(cpu);
669
670 /*
671 * Offlining this cpu will cause the cpu before this one to start
672 * checking the one after this one. If this cpu just finished checking
673 * the next cpu and updating hrtimer_interrupts_saved, and then the
674 * previous cpu checks it within one sample period, it will trigger a
675 * false positive. Touch the watchdog on the next cpu to prevent it.
676 */
677 if (next_cpu < nr_cpu_ids)
678 per_cpu(watchdog_nmi_touch, next_cpu) = true;
679 smp_wmb();
680 cpumask_clear_cpu(cpu, &watchdog_cpus);
681 }
682 #else
watchdog_nmi_enable(unsigned int cpu)683 static int watchdog_nmi_enable(unsigned int cpu) { return 0; }
watchdog_nmi_disable(unsigned int cpu)684 static void watchdog_nmi_disable(unsigned int cpu) { return; }
685 #endif /* CONFIG_HARDLOCKUP_DETECTOR_OTHER_CPU */
686 #endif /* CONFIG_HARDLOCKUP_DETECTOR_NMI */
687
688 static struct smp_hotplug_thread watchdog_threads = {
689 .store = &softlockup_watchdog,
690 .thread_should_run = watchdog_should_run,
691 .thread_fn = watchdog,
692 .thread_comm = "watchdog/%u",
693 .setup = watchdog_enable,
694 .cleanup = watchdog_cleanup,
695 .park = watchdog_disable,
696 .unpark = watchdog_enable,
697 };
698
restart_watchdog_hrtimer(void * info)699 static void restart_watchdog_hrtimer(void *info)
700 {
701 struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer);
702 int ret;
703
704 /*
705 * No need to cancel and restart hrtimer if it is currently executing
706 * because it will reprogram itself with the new period now.
707 * We should never see it unqueued here because we are running per-cpu
708 * with interrupts disabled.
709 */
710 ret = hrtimer_try_to_cancel(hrtimer);
711 if (ret == 1)
712 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
713 HRTIMER_MODE_REL_PINNED);
714 }
715
update_timers(int cpu)716 static void update_timers(int cpu)
717 {
718 /*
719 * Make sure that perf event counter will adopt to a new
720 * sampling period. Updating the sampling period directly would
721 * be much nicer but we do not have an API for that now so
722 * let's use a big hammer.
723 * Hrtimer will adopt the new period on the next tick but this
724 * might be late already so we have to restart the timer as well.
725 */
726 watchdog_nmi_disable(cpu);
727 smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1);
728 watchdog_nmi_enable(cpu);
729 }
730
update_timers_all_cpus(void)731 static void update_timers_all_cpus(void)
732 {
733 int cpu;
734
735 get_online_cpus();
736 for_each_online_cpu(cpu)
737 update_timers(cpu);
738 put_online_cpus();
739 }
740
watchdog_enable_all_cpus(bool sample_period_changed)741 static int watchdog_enable_all_cpus(bool sample_period_changed)
742 {
743 int err = 0;
744
745 if (!watchdog_running) {
746 err = smpboot_register_percpu_thread(&watchdog_threads);
747 if (err)
748 pr_err("Failed to create watchdog threads, disabled\n");
749 else
750 watchdog_running = 1;
751 } else if (sample_period_changed) {
752 update_timers_all_cpus();
753 }
754
755 return err;
756 }
757
758 /* prepare/enable/disable routines */
759 /* sysctl functions */
760 #ifdef CONFIG_SYSCTL
watchdog_disable_all_cpus(void)761 static void watchdog_disable_all_cpus(void)
762 {
763 if (watchdog_running) {
764 watchdog_running = 0;
765 smpboot_unregister_percpu_thread(&watchdog_threads);
766 }
767 }
768
769 /*
770 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
771 */
772
proc_dowatchdog(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)773 int proc_dowatchdog(struct ctl_table *table, int write,
774 void __user *buffer, size_t *lenp, loff_t *ppos)
775 {
776 int err, old_thresh, old_enabled;
777 bool old_hardlockup;
778 static DEFINE_MUTEX(watchdog_proc_mutex);
779
780 mutex_lock(&watchdog_proc_mutex);
781 old_thresh = ACCESS_ONCE(watchdog_thresh);
782 old_enabled = ACCESS_ONCE(watchdog_user_enabled);
783 old_hardlockup = watchdog_hardlockup_detector_is_enabled();
784
785 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
786 if (err || !write)
787 goto out;
788
789 set_sample_period();
790 /*
791 * Watchdog threads shouldn't be enabled if they are
792 * disabled. The 'watchdog_running' variable check in
793 * watchdog_*_all_cpus() function takes care of this.
794 */
795 if (watchdog_user_enabled && watchdog_thresh) {
796 /*
797 * Prevent a change in watchdog_thresh accidentally overriding
798 * the enablement of the hardlockup detector.
799 */
800 if (watchdog_user_enabled != old_enabled)
801 watchdog_enable_hardlockup_detector(true);
802 err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh);
803 } else
804 watchdog_disable_all_cpus();
805
806 /* Restore old values on failure */
807 if (err) {
808 watchdog_thresh = old_thresh;
809 watchdog_user_enabled = old_enabled;
810 watchdog_enable_hardlockup_detector(old_hardlockup);
811 }
812 out:
813 mutex_unlock(&watchdog_proc_mutex);
814 return err;
815 }
816 #endif /* CONFIG_SYSCTL */
817
lockup_detector_init(void)818 void __init lockup_detector_init(void)
819 {
820 set_sample_period();
821
822 if (watchdog_user_enabled)
823 watchdog_enable_all_cpus(false);
824 }
825