1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Detect hard and soft lockups on a system
4 *
5 * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc.
6 *
7 * Note: Most of this code is borrowed heavily from the original softlockup
8 * detector, so thanks to Ingo for the initial implementation.
9 * Some chunks also taken from the old x86-specific nmi watchdog code, thanks
10 * to those contributors as well.
11 */
12
13 #define pr_fmt(fmt) "watchdog: " fmt
14
15 #include <linux/cpu.h>
16 #include <linux/init.h>
17 #include <linux/irq.h>
18 #include <linux/irqdesc.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/kvm_para.h>
21 #include <linux/math64.h>
22 #include <linux/mm.h>
23 #include <linux/module.h>
24 #include <linux/nmi.h>
25 #include <linux/stop_machine.h>
26 #include <linux/sysctl.h>
27 #include <linux/tick.h>
28
29 #include <linux/sched/clock.h>
30 #include <linux/sched/debug.h>
31 #include <linux/sched/isolation.h>
32
33 #include <asm/irq_regs.h>
34
35 static DEFINE_MUTEX(watchdog_mutex);
36
37 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HARDLOCKUP_DETECTOR_SPARC64)
38 # define WATCHDOG_HARDLOCKUP_DEFAULT 1
39 #else
40 # define WATCHDOG_HARDLOCKUP_DEFAULT 0
41 #endif
42
43 #define NUM_SAMPLE_PERIODS 5
44
45 unsigned long __read_mostly watchdog_enabled;
46 int __read_mostly watchdog_user_enabled = 1;
47 static int __read_mostly watchdog_hardlockup_user_enabled = WATCHDOG_HARDLOCKUP_DEFAULT;
48 static int __read_mostly watchdog_softlockup_user_enabled = 1;
49 int __read_mostly watchdog_thresh = 10;
50 static int __read_mostly watchdog_hardlockup_available;
51
52 struct cpumask watchdog_cpumask __read_mostly;
53 unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
54
55 #ifdef CONFIG_HARDLOCKUP_DETECTOR
56
57 # ifdef CONFIG_SMP
58 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
59 # endif /* CONFIG_SMP */
60
61 /*
62 * Should we panic when a soft-lockup or hard-lockup occurs:
63 */
64 unsigned int __read_mostly hardlockup_panic =
65 IS_ENABLED(CONFIG_BOOTPARAM_HARDLOCKUP_PANIC);
66 /*
67 * We may not want to enable hard lockup detection by default in all cases,
68 * for example when running the kernel as a guest on a hypervisor. In these
69 * cases this function can be called to disable hard lockup detection. This
70 * function should only be executed once by the boot processor before the
71 * kernel command line parameters are parsed, because otherwise it is not
72 * possible to override this in hardlockup_panic_setup().
73 */
74 void __init hardlockup_detector_disable(void)
75 {
76 watchdog_hardlockup_user_enabled = 0;
77 }
78
79 static int __init hardlockup_panic_setup(char *str)
80 {
81 next:
82 if (!strncmp(str, "panic", 5))
83 hardlockup_panic = 1;
84 else if (!strncmp(str, "nopanic", 7))
85 hardlockup_panic = 0;
86 else if (!strncmp(str, "", 1))
87 watchdog_hardlockup_user_enabled = 0;
88 else if (!strncmp(str, "1", 1))
89 watchdog_hardlockup_user_enabled = 1;
90 else if (!strncmp(str, "r", 1))
91 hardlockup_config_perf_event(str + 1);
92 while (*(str++)) {
93 if (*str == ',') {
94 str++;
95 goto next;
96 }
97 }
98 return 1;
99 }
100 __setup("nmi_watchdog=", hardlockup_panic_setup);
101
102 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
103
104 #if defined(CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER)
105
106 static DEFINE_PER_CPU(atomic_t, hrtimer_interrupts);
107 static DEFINE_PER_CPU(int, hrtimer_interrupts_saved);
108 static DEFINE_PER_CPU(bool, watchdog_hardlockup_warned);
109 static DEFINE_PER_CPU(bool, watchdog_hardlockup_touched);
110 static unsigned long hard_lockup_nmi_warn;
111
112 notrace void arch_touch_nmi_watchdog(void)
113 {
114 /*
115 * Using __raw here because some code paths have
116 * preemption enabled. If preemption is enabled
117 * then interrupts should be enabled too, in which
118 * case we shouldn't have to worry about the watchdog
119 * going off.
120 */
121 raw_cpu_write(watchdog_hardlockup_touched, true);
122 }
123 EXPORT_SYMBOL(arch_touch_nmi_watchdog);
124
125 void watchdog_hardlockup_touch_cpu(unsigned int cpu)
126 {
127 per_cpu(watchdog_hardlockup_touched, cpu) = true;
128 }
129
130 static bool is_hardlockup(unsigned int cpu)
131 {
132 int hrint = atomic_read(&per_cpu(hrtimer_interrupts, cpu));
133
134 if (per_cpu(hrtimer_interrupts_saved, cpu) == hrint)
135 return true;
136
137 /*
138 * NOTE: we don't need any fancy atomic_t or READ_ONCE/WRITE_ONCE
139 * for hrtimer_interrupts_saved. hrtimer_interrupts_saved is
140 * written/read by a single CPU.
141 */
142 per_cpu(hrtimer_interrupts_saved, cpu) = hrint;
143
144 return false;
145 }
146
147 static void watchdog_hardlockup_kick(void)
148 {
149 int new_interrupts;
150
151 new_interrupts = atomic_inc_return(this_cpu_ptr(&hrtimer_interrupts));
152 watchdog_buddy_check_hardlockup(new_interrupts);
153 }
154
155 void watchdog_hardlockup_check(unsigned int cpu, struct pt_regs *regs)
156 {
157 if (per_cpu(watchdog_hardlockup_touched, cpu)) {
158 per_cpu(watchdog_hardlockup_touched, cpu) = false;
159 return;
160 }
161
162 /*
163 * Check for a hardlockup by making sure the CPU's timer
164 * interrupt is incrementing. The timer interrupt should have
165 * fired multiple times before we overflow'd. If it hasn't
166 * then this is a good indication the cpu is stuck
167 */
168 if (is_hardlockup(cpu)) {
169 unsigned int this_cpu = smp_processor_id();
170 unsigned long flags;
171
172 /* Only print hardlockups once. */
173 if (per_cpu(watchdog_hardlockup_warned, cpu))
174 return;
175
176 /*
177 * Prevent multiple hard-lockup reports if one cpu is already
178 * engaged in dumping all cpu back traces.
179 */
180 if (sysctl_hardlockup_all_cpu_backtrace) {
181 if (test_and_set_bit_lock(0, &hard_lockup_nmi_warn))
182 return;
183 }
184
185 /*
186 * NOTE: we call printk_cpu_sync_get_irqsave() after printing
187 * the lockup message. While it would be nice to serialize
188 * that printout, we really want to make sure that if some
189 * other CPU somehow locked up while holding the lock associated
190 * with printk_cpu_sync_get_irqsave() that we can still at least
191 * get the message about the lockup out.
192 */
193 pr_emerg("Watchdog detected hard LOCKUP on cpu %d\n", cpu);
194 printk_cpu_sync_get_irqsave(flags);
195
196 print_modules();
197 print_irqtrace_events(current);
198 if (cpu == this_cpu) {
199 if (regs)
200 show_regs(regs);
201 else
202 dump_stack();
203 printk_cpu_sync_put_irqrestore(flags);
204 } else {
205 printk_cpu_sync_put_irqrestore(flags);
206 trigger_single_cpu_backtrace(cpu);
207 }
208
209 if (sysctl_hardlockup_all_cpu_backtrace) {
210 trigger_allbutcpu_cpu_backtrace(cpu);
211 if (!hardlockup_panic)
212 clear_bit_unlock(0, &hard_lockup_nmi_warn);
213 }
214
215 if (hardlockup_panic)
216 nmi_panic(regs, "Hard LOCKUP");
217
218 per_cpu(watchdog_hardlockup_warned, cpu) = true;
219 } else {
220 per_cpu(watchdog_hardlockup_warned, cpu) = false;
221 }
222 }
223
224 #else /* CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
225
226 static inline void watchdog_hardlockup_kick(void) { }
227
228 #endif /* !CONFIG_HARDLOCKUP_DETECTOR_COUNTS_HRTIMER */
229
230 /*
231 * These functions can be overridden based on the configured hardlockdup detector.
232 *
233 * watchdog_hardlockup_enable/disable can be implemented to start and stop when
234 * softlockup watchdog start and stop. The detector must select the
235 * SOFTLOCKUP_DETECTOR Kconfig.
236 */
237 void __weak watchdog_hardlockup_enable(unsigned int cpu) { }
238
239 void __weak watchdog_hardlockup_disable(unsigned int cpu) { }
240
241 /*
242 * Watchdog-detector specific API.
243 *
244 * Return 0 when hardlockup watchdog is available, negative value otherwise.
245 * Note that the negative value means that a delayed probe might
246 * succeed later.
247 */
248 int __weak __init watchdog_hardlockup_probe(void)
249 {
250 return -ENODEV;
251 }
252
253 /**
254 * watchdog_hardlockup_stop - Stop the watchdog for reconfiguration
255 *
256 * The reconfiguration steps are:
257 * watchdog_hardlockup_stop();
258 * update_variables();
259 * watchdog_hardlockup_start();
260 */
261 void __weak watchdog_hardlockup_stop(void) { }
262
263 /**
264 * watchdog_hardlockup_start - Start the watchdog after reconfiguration
265 *
266 * Counterpart to watchdog_hardlockup_stop().
267 *
268 * The following variables have been updated in update_variables() and
269 * contain the currently valid configuration:
270 * - watchdog_enabled
271 * - watchdog_thresh
272 * - watchdog_cpumask
273 */
274 void __weak watchdog_hardlockup_start(void) { }
275
276 /**
277 * lockup_detector_update_enable - Update the sysctl enable bit
278 *
279 * Caller needs to make sure that the hard watchdogs are off, so this
280 * can't race with watchdog_hardlockup_disable().
281 */
282 static void lockup_detector_update_enable(void)
283 {
284 watchdog_enabled = 0;
285 if (!watchdog_user_enabled)
286 return;
287 if (watchdog_hardlockup_available && watchdog_hardlockup_user_enabled)
288 watchdog_enabled |= WATCHDOG_HARDLOCKUP_ENABLED;
289 if (watchdog_softlockup_user_enabled)
290 watchdog_enabled |= WATCHDOG_SOFTOCKUP_ENABLED;
291 }
292
293 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
294
295 /*
296 * Delay the soflockup report when running a known slow code.
297 * It does _not_ affect the timestamp of the last successdul reschedule.
298 */
299 #define SOFTLOCKUP_DELAY_REPORT ULONG_MAX
300
301 #ifdef CONFIG_SMP
302 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
303 #endif
304
305 static struct cpumask watchdog_allowed_mask __read_mostly;
306
307 /* Global variables, exported for sysctl */
308 unsigned int __read_mostly softlockup_panic =
309 IS_ENABLED(CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC);
310
311 static bool softlockup_initialized __read_mostly;
312 static u64 __read_mostly sample_period;
313
314 /* Timestamp taken after the last successful reschedule. */
315 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
316 /* Timestamp of the last softlockup report. */
317 static DEFINE_PER_CPU(unsigned long, watchdog_report_ts);
318 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
319 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
320 static unsigned long soft_lockup_nmi_warn;
321
322 static int __init softlockup_panic_setup(char *str)
323 {
324 softlockup_panic = simple_strtoul(str, NULL, 0);
325 return 1;
326 }
327 __setup("softlockup_panic=", softlockup_panic_setup);
328
329 static int __init nowatchdog_setup(char *str)
330 {
331 watchdog_user_enabled = 0;
332 return 1;
333 }
334 __setup("nowatchdog", nowatchdog_setup);
335
336 static int __init nosoftlockup_setup(char *str)
337 {
338 watchdog_softlockup_user_enabled = 0;
339 return 1;
340 }
341 __setup("nosoftlockup", nosoftlockup_setup);
342
343 static int __init watchdog_thresh_setup(char *str)
344 {
345 get_option(&str, &watchdog_thresh);
346 return 1;
347 }
348 __setup("watchdog_thresh=", watchdog_thresh_setup);
349
350 static void __lockup_detector_cleanup(void);
351
352 #ifdef CONFIG_SOFTLOCKUP_DETECTOR_INTR_STORM
353 enum stats_per_group {
354 STATS_SYSTEM,
355 STATS_SOFTIRQ,
356 STATS_HARDIRQ,
357 STATS_IDLE,
358 NUM_STATS_PER_GROUP,
359 };
360
361 static const enum cpu_usage_stat tracked_stats[NUM_STATS_PER_GROUP] = {
362 CPUTIME_SYSTEM,
363 CPUTIME_SOFTIRQ,
364 CPUTIME_IRQ,
365 CPUTIME_IDLE,
366 };
367
368 static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
369 static DEFINE_PER_CPU(u8, cpustat_util[NUM_SAMPLE_PERIODS][NUM_STATS_PER_GROUP]);
370 static DEFINE_PER_CPU(u8, cpustat_tail);
371
372 /*
373 * We don't need nanosecond resolution. A granularity of 16ms is
374 * sufficient for our precision, allowing us to use u16 to store
375 * cpustats, which will roll over roughly every ~1000 seconds.
376 * 2^24 ~= 16 * 10^6
377 */
378 static u16 get_16bit_precision(u64 data_ns)
379 {
380 return data_ns >> 24LL; /* 2^24ns ~= 16.8ms */
381 }
382
383 static void update_cpustat(void)
384 {
385 int i;
386 u8 util;
387 u16 old_stat, new_stat;
388 struct kernel_cpustat kcpustat;
389 u64 *cpustat = kcpustat.cpustat;
390 u8 tail = __this_cpu_read(cpustat_tail);
391 u16 sample_period_16 = get_16bit_precision(sample_period);
392
393 kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
394
395 for (i = 0; i < NUM_STATS_PER_GROUP; i++) {
396 old_stat = __this_cpu_read(cpustat_old[i]);
397 new_stat = get_16bit_precision(cpustat[tracked_stats[i]]);
398 util = DIV_ROUND_UP(100 * (new_stat - old_stat), sample_period_16);
399 __this_cpu_write(cpustat_util[tail][i], util);
400 __this_cpu_write(cpustat_old[i], new_stat);
401 }
402
403 __this_cpu_write(cpustat_tail, (tail + 1) % NUM_SAMPLE_PERIODS);
404 }
405
406 static void print_cpustat(void)
407 {
408 int i, group;
409 u8 tail = __this_cpu_read(cpustat_tail);
410 u64 sample_period_second = sample_period;
411
412 do_div(sample_period_second, NSEC_PER_SEC);
413
414 /*
415 * Outputting the "watchdog" prefix on every line is redundant and not
416 * concise, and the original alarm information is sufficient for
417 * positioning in logs, hence here printk() is used instead of pr_crit().
418 */
419 printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
420 smp_processor_id(), sample_period_second);
421
422 for (i = 0; i < NUM_SAMPLE_PERIODS; i++) {
423 group = (tail + i) % NUM_SAMPLE_PERIODS;
424 printk(KERN_CRIT "\t#%d: %3u%% system,\t%3u%% softirq,\t"
425 "%3u%% hardirq,\t%3u%% idle\n", i + 1,
426 __this_cpu_read(cpustat_util[group][STATS_SYSTEM]),
427 __this_cpu_read(cpustat_util[group][STATS_SOFTIRQ]),
428 __this_cpu_read(cpustat_util[group][STATS_HARDIRQ]),
429 __this_cpu_read(cpustat_util[group][STATS_IDLE]));
430 }
431 }
432
433 #define HARDIRQ_PERCENT_THRESH 50
434 #define NUM_HARDIRQ_REPORT 5
435 struct irq_counts {
436 int irq;
437 u32 counts;
438 };
439
440 static DEFINE_PER_CPU(bool, snapshot_taken);
441
442 /* Tabulate the most frequent interrupts. */
443 static void tabulate_irq_count(struct irq_counts *irq_counts, int irq, u32 counts, int rank)
444 {
445 int i;
446 struct irq_counts new_count = {irq, counts};
447
448 for (i = 0; i < rank; i++) {
449 if (counts > irq_counts[i].counts)
450 swap(new_count, irq_counts[i]);
451 }
452 }
453
454 /*
455 * If the hardirq time exceeds HARDIRQ_PERCENT_THRESH% of the sample_period,
456 * then the cause of softlockup might be interrupt storm. In this case, it
457 * would be useful to start interrupt counting.
458 */
459 static bool need_counting_irqs(void)
460 {
461 u8 util;
462 int tail = __this_cpu_read(cpustat_tail);
463
464 tail = (tail + NUM_HARDIRQ_REPORT - 1) % NUM_HARDIRQ_REPORT;
465 util = __this_cpu_read(cpustat_util[tail][STATS_HARDIRQ]);
466 return util > HARDIRQ_PERCENT_THRESH;
467 }
468
469 static void start_counting_irqs(void)
470 {
471 if (!__this_cpu_read(snapshot_taken)) {
472 kstat_snapshot_irqs();
473 __this_cpu_write(snapshot_taken, true);
474 }
475 }
476
477 static void stop_counting_irqs(void)
478 {
479 __this_cpu_write(snapshot_taken, false);
480 }
481
482 static void print_irq_counts(void)
483 {
484 unsigned int i, count;
485 struct irq_counts irq_counts_sorted[NUM_HARDIRQ_REPORT] = {
486 {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}, {-1, 0}
487 };
488
489 if (__this_cpu_read(snapshot_taken)) {
490 for_each_active_irq(i) {
491 count = kstat_get_irq_since_snapshot(i);
492 tabulate_irq_count(irq_counts_sorted, i, count, NUM_HARDIRQ_REPORT);
493 }
494
495 /*
496 * Outputting the "watchdog" prefix on every line is redundant and not
497 * concise, and the original alarm information is sufficient for
498 * positioning in logs, hence here printk() is used instead of pr_crit().
499 */
500 printk(KERN_CRIT "CPU#%d Detect HardIRQ Time exceeds %d%%. Most frequent HardIRQs:\n",
501 smp_processor_id(), HARDIRQ_PERCENT_THRESH);
502
503 for (i = 0; i < NUM_HARDIRQ_REPORT; i++) {
504 if (irq_counts_sorted[i].irq == -1)
505 break;
506
507 printk(KERN_CRIT "\t#%u: %-10u\tirq#%d\n",
508 i + 1, irq_counts_sorted[i].counts,
509 irq_counts_sorted[i].irq);
510 }
511
512 /*
513 * If the hardirq time is less than HARDIRQ_PERCENT_THRESH% in the last
514 * sample_period, then we suspect the interrupt storm might be subsiding.
515 */
516 if (!need_counting_irqs())
517 stop_counting_irqs();
518 }
519 }
520
521 static void report_cpu_status(void)
522 {
523 print_cpustat();
524 print_irq_counts();
525 }
526 #else
527 static inline void update_cpustat(void) { }
528 static inline void report_cpu_status(void) { }
529 static inline bool need_counting_irqs(void) { return false; }
530 static inline void start_counting_irqs(void) { }
531 static inline void stop_counting_irqs(void) { }
532 #endif
533
534 /*
535 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
536 * lockups can have false positives under extreme conditions. So we generally
537 * want a higher threshold for soft lockups than for hard lockups. So we couple
538 * the thresholds with a factor: we make the soft threshold twice the amount of
539 * time the hard threshold is.
540 */
541 static int get_softlockup_thresh(void)
542 {
543 return watchdog_thresh * 2;
544 }
545
546 /*
547 * Returns seconds, approximately. We don't need nanosecond
548 * resolution, and we don't need to waste time with a big divide when
549 * 2^30ns == 1.074s.
550 */
551 static unsigned long get_timestamp(void)
552 {
553 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
554 }
555
556 static void set_sample_period(void)
557 {
558 /*
559 * convert watchdog_thresh from seconds to ns
560 * the divide by 5 is to give hrtimer several chances (two
561 * or three with the current relation between the soft
562 * and hard thresholds) to increment before the
563 * hardlockup detector generates a warning
564 */
565 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / NUM_SAMPLE_PERIODS);
566 watchdog_update_hrtimer_threshold(sample_period);
567 }
568
569 static void update_report_ts(void)
570 {
571 __this_cpu_write(watchdog_report_ts, get_timestamp());
572 }
573
574 /* Commands for resetting the watchdog */
575 static void update_touch_ts(void)
576 {
577 __this_cpu_write(watchdog_touch_ts, get_timestamp());
578 update_report_ts();
579 }
580
581 /**
582 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
583 *
584 * Call when the scheduler may have stalled for legitimate reasons
585 * preventing the watchdog task from executing - e.g. the scheduler
586 * entering idle state. This should only be used for scheduler events.
587 * Use touch_softlockup_watchdog() for everything else.
588 */
589 notrace void touch_softlockup_watchdog_sched(void)
590 {
591 /*
592 * Preemption can be enabled. It doesn't matter which CPU's watchdog
593 * report period gets restarted here, so use the raw_ operation.
594 */
595 raw_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
596 }
597
598 notrace void touch_softlockup_watchdog(void)
599 {
600 touch_softlockup_watchdog_sched();
601 wq_watchdog_touch(raw_smp_processor_id());
602 }
603 EXPORT_SYMBOL(touch_softlockup_watchdog);
604
605 void touch_all_softlockup_watchdogs(void)
606 {
607 int cpu;
608
609 /*
610 * watchdog_mutex cannpt be taken here, as this might be called
611 * from (soft)interrupt context, so the access to
612 * watchdog_allowed_cpumask might race with a concurrent update.
613 *
614 * The watchdog time stamp can race against a concurrent real
615 * update as well, the only side effect might be a cycle delay for
616 * the softlockup check.
617 */
618 for_each_cpu(cpu, &watchdog_allowed_mask) {
619 per_cpu(watchdog_report_ts, cpu) = SOFTLOCKUP_DELAY_REPORT;
620 wq_watchdog_touch(cpu);
621 }
622 }
623
624 void touch_softlockup_watchdog_sync(void)
625 {
626 __this_cpu_write(softlockup_touch_sync, true);
627 __this_cpu_write(watchdog_report_ts, SOFTLOCKUP_DELAY_REPORT);
628 }
629
630 static int is_softlockup(unsigned long touch_ts,
631 unsigned long period_ts,
632 unsigned long now)
633 {
634 if ((watchdog_enabled & WATCHDOG_SOFTOCKUP_ENABLED) && watchdog_thresh) {
635 /*
636 * If period_ts has not been updated during a sample_period, then
637 * in the subsequent few sample_periods, period_ts might also not
638 * be updated, which could indicate a potential softlockup. In
639 * this case, if we suspect the cause of the potential softlockup
640 * might be interrupt storm, then we need to count the interrupts
641 * to find which interrupt is storming.
642 */
643 if (time_after_eq(now, period_ts + get_softlockup_thresh() / NUM_SAMPLE_PERIODS) &&
644 need_counting_irqs())
645 start_counting_irqs();
646
647 /* Warn about unreasonable delays. */
648 if (time_after(now, period_ts + get_softlockup_thresh()))
649 return now - touch_ts;
650 }
651 return 0;
652 }
653
654 /* watchdog detector functions */
655 static DEFINE_PER_CPU(struct completion, softlockup_completion);
656 static DEFINE_PER_CPU(struct cpu_stop_work, softlockup_stop_work);
657
658 /*
659 * The watchdog feed function - touches the timestamp.
660 *
661 * It only runs once every sample_period seconds (4 seconds by
662 * default) to reset the softlockup timestamp. If this gets delayed
663 * for more than 2*watchdog_thresh seconds then the debug-printout
664 * triggers in watchdog_timer_fn().
665 */
666 static int softlockup_fn(void *data)
667 {
668 update_touch_ts();
669 stop_counting_irqs();
670 complete(this_cpu_ptr(&softlockup_completion));
671
672 return 0;
673 }
674
675 /* watchdog kicker functions */
676 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
677 {
678 unsigned long touch_ts, period_ts, now;
679 struct pt_regs *regs = get_irq_regs();
680 int duration;
681 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
682 unsigned long flags;
683
684 if (!watchdog_enabled)
685 return HRTIMER_NORESTART;
686
687 watchdog_hardlockup_kick();
688
689 /* kick the softlockup detector */
690 if (completion_done(this_cpu_ptr(&softlockup_completion))) {
691 reinit_completion(this_cpu_ptr(&softlockup_completion));
692 stop_one_cpu_nowait(smp_processor_id(),
693 softlockup_fn, NULL,
694 this_cpu_ptr(&softlockup_stop_work));
695 }
696
697 /* .. and repeat */
698 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
699
700 /*
701 * Read the current timestamp first. It might become invalid anytime
702 * when a virtual machine is stopped by the host or when the watchog
703 * is touched from NMI.
704 */
705 now = get_timestamp();
706 /*
707 * If a virtual machine is stopped by the host it can look to
708 * the watchdog like a soft lockup. This function touches the watchdog.
709 */
710 kvm_check_and_clear_guest_paused();
711 /*
712 * The stored timestamp is comparable with @now only when not touched.
713 * It might get touched anytime from NMI. Make sure that is_softlockup()
714 * uses the same (valid) value.
715 */
716 period_ts = READ_ONCE(*this_cpu_ptr(&watchdog_report_ts));
717
718 update_cpustat();
719
720 /* Reset the interval when touched by known problematic code. */
721 if (period_ts == SOFTLOCKUP_DELAY_REPORT) {
722 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
723 /*
724 * If the time stamp was touched atomically
725 * make sure the scheduler tick is up to date.
726 */
727 __this_cpu_write(softlockup_touch_sync, false);
728 sched_clock_tick();
729 }
730
731 update_report_ts();
732 return HRTIMER_RESTART;
733 }
734
735 /* Check for a softlockup. */
736 touch_ts = __this_cpu_read(watchdog_touch_ts);
737 duration = is_softlockup(touch_ts, period_ts, now);
738 if (unlikely(duration)) {
739 /*
740 * Prevent multiple soft-lockup reports if one cpu is already
741 * engaged in dumping all cpu back traces.
742 */
743 if (softlockup_all_cpu_backtrace) {
744 if (test_and_set_bit_lock(0, &soft_lockup_nmi_warn))
745 return HRTIMER_RESTART;
746 }
747
748 /* Start period for the next softlockup warning. */
749 update_report_ts();
750
751 printk_cpu_sync_get_irqsave(flags);
752 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
753 smp_processor_id(), duration,
754 current->comm, task_pid_nr(current));
755 report_cpu_status();
756 print_modules();
757 print_irqtrace_events(current);
758 if (regs)
759 show_regs(regs);
760 else
761 dump_stack();
762 printk_cpu_sync_put_irqrestore(flags);
763
764 if (softlockup_all_cpu_backtrace) {
765 trigger_allbutcpu_cpu_backtrace(smp_processor_id());
766 if (!softlockup_panic)
767 clear_bit_unlock(0, &soft_lockup_nmi_warn);
768 }
769
770 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
771 if (softlockup_panic)
772 panic("softlockup: hung tasks");
773 }
774
775 return HRTIMER_RESTART;
776 }
777
778 static void watchdog_enable(unsigned int cpu)
779 {
780 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
781 struct completion *done = this_cpu_ptr(&softlockup_completion);
782
783 WARN_ON_ONCE(cpu != smp_processor_id());
784
785 init_completion(done);
786 complete(done);
787
788 /*
789 * Start the timer first to prevent the hardlockup watchdog triggering
790 * before the timer has a chance to fire.
791 */
792 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_HARD);
793 hrtimer->function = watchdog_timer_fn;
794 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
795 HRTIMER_MODE_REL_PINNED_HARD);
796
797 /* Initialize timestamp */
798 update_touch_ts();
799 /* Enable the hardlockup detector */
800 if (watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED)
801 watchdog_hardlockup_enable(cpu);
802 }
803
804 static void watchdog_disable(unsigned int cpu)
805 {
806 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
807
808 WARN_ON_ONCE(cpu != smp_processor_id());
809
810 /*
811 * Disable the hardlockup detector first. That prevents that a large
812 * delay between disabling the timer and disabling the hardlockup
813 * detector causes a false positive.
814 */
815 watchdog_hardlockup_disable(cpu);
816 hrtimer_cancel(hrtimer);
817 wait_for_completion(this_cpu_ptr(&softlockup_completion));
818 }
819
820 static int softlockup_stop_fn(void *data)
821 {
822 watchdog_disable(smp_processor_id());
823 return 0;
824 }
825
826 static void softlockup_stop_all(void)
827 {
828 int cpu;
829
830 if (!softlockup_initialized)
831 return;
832
833 for_each_cpu(cpu, &watchdog_allowed_mask)
834 smp_call_on_cpu(cpu, softlockup_stop_fn, NULL, false);
835
836 cpumask_clear(&watchdog_allowed_mask);
837 }
838
839 static int softlockup_start_fn(void *data)
840 {
841 watchdog_enable(smp_processor_id());
842 return 0;
843 }
844
845 static void softlockup_start_all(void)
846 {
847 int cpu;
848
849 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
850 for_each_cpu(cpu, &watchdog_allowed_mask)
851 smp_call_on_cpu(cpu, softlockup_start_fn, NULL, false);
852 }
853
854 int lockup_detector_online_cpu(unsigned int cpu)
855 {
856 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
857 watchdog_enable(cpu);
858 return 0;
859 }
860
861 int lockup_detector_offline_cpu(unsigned int cpu)
862 {
863 if (cpumask_test_cpu(cpu, &watchdog_allowed_mask))
864 watchdog_disable(cpu);
865 return 0;
866 }
867
868 static void __lockup_detector_reconfigure(void)
869 {
870 cpus_read_lock();
871 watchdog_hardlockup_stop();
872
873 softlockup_stop_all();
874 set_sample_period();
875 lockup_detector_update_enable();
876 if (watchdog_enabled && watchdog_thresh)
877 softlockup_start_all();
878
879 watchdog_hardlockup_start();
880 cpus_read_unlock();
881 /*
882 * Must be called outside the cpus locked section to prevent
883 * recursive locking in the perf code.
884 */
885 __lockup_detector_cleanup();
886 }
887
888 void lockup_detector_reconfigure(void)
889 {
890 mutex_lock(&watchdog_mutex);
891 __lockup_detector_reconfigure();
892 mutex_unlock(&watchdog_mutex);
893 }
894
895 /*
896 * Create the watchdog infrastructure and configure the detector(s).
897 */
898 static __init void lockup_detector_setup(void)
899 {
900 /*
901 * If sysctl is off and watchdog got disabled on the command line,
902 * nothing to do here.
903 */
904 lockup_detector_update_enable();
905
906 if (!IS_ENABLED(CONFIG_SYSCTL) &&
907 !(watchdog_enabled && watchdog_thresh))
908 return;
909
910 mutex_lock(&watchdog_mutex);
911 __lockup_detector_reconfigure();
912 softlockup_initialized = true;
913 mutex_unlock(&watchdog_mutex);
914 }
915
916 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
917 static void __lockup_detector_reconfigure(void)
918 {
919 cpus_read_lock();
920 watchdog_hardlockup_stop();
921 lockup_detector_update_enable();
922 watchdog_hardlockup_start();
923 cpus_read_unlock();
924 }
925 void lockup_detector_reconfigure(void)
926 {
927 __lockup_detector_reconfigure();
928 }
929 static inline void lockup_detector_setup(void)
930 {
931 __lockup_detector_reconfigure();
932 }
933 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
934
935 static void __lockup_detector_cleanup(void)
936 {
937 lockdep_assert_held(&watchdog_mutex);
938 hardlockup_detector_perf_cleanup();
939 }
940
941 /**
942 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
943 *
944 * Caller must not hold the cpu hotplug rwsem.
945 */
946 void lockup_detector_cleanup(void)
947 {
948 mutex_lock(&watchdog_mutex);
949 __lockup_detector_cleanup();
950 mutex_unlock(&watchdog_mutex);
951 }
952
953 /**
954 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
955 *
956 * Special interface for parisc. It prevents lockup detector warnings from
957 * the default pm_poweroff() function which busy loops forever.
958 */
959 void lockup_detector_soft_poweroff(void)
960 {
961 watchdog_enabled = 0;
962 }
963
964 #ifdef CONFIG_SYSCTL
965
966 /* Propagate any changes to the watchdog infrastructure */
967 static void proc_watchdog_update(void)
968 {
969 /* Remove impossible cpus to keep sysctl output clean. */
970 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
971 __lockup_detector_reconfigure();
972 }
973
974 /*
975 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
976 *
977 * caller | table->data points to | 'which'
978 * -------------------|----------------------------------|-------------------------------
979 * proc_watchdog | watchdog_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED |
980 * | | WATCHDOG_SOFTOCKUP_ENABLED
981 * -------------------|----------------------------------|-------------------------------
982 * proc_nmi_watchdog | watchdog_hardlockup_user_enabled | WATCHDOG_HARDLOCKUP_ENABLED
983 * -------------------|----------------------------------|-------------------------------
984 * proc_soft_watchdog | watchdog_softlockup_user_enabled | WATCHDOG_SOFTOCKUP_ENABLED
985 */
986 static int proc_watchdog_common(int which, const struct ctl_table *table, int write,
987 void *buffer, size_t *lenp, loff_t *ppos)
988 {
989 int err, old, *param = table->data;
990
991 mutex_lock(&watchdog_mutex);
992
993 if (!write) {
994 /*
995 * On read synchronize the userspace interface. This is a
996 * racy snapshot.
997 */
998 *param = (watchdog_enabled & which) != 0;
999 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1000 } else {
1001 old = READ_ONCE(*param);
1002 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1003 if (!err && old != READ_ONCE(*param))
1004 proc_watchdog_update();
1005 }
1006 mutex_unlock(&watchdog_mutex);
1007 return err;
1008 }
1009
1010 /*
1011 * /proc/sys/kernel/watchdog
1012 */
1013 static int proc_watchdog(const struct ctl_table *table, int write,
1014 void *buffer, size_t *lenp, loff_t *ppos)
1015 {
1016 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED |
1017 WATCHDOG_SOFTOCKUP_ENABLED,
1018 table, write, buffer, lenp, ppos);
1019 }
1020
1021 /*
1022 * /proc/sys/kernel/nmi_watchdog
1023 */
1024 static int proc_nmi_watchdog(const struct ctl_table *table, int write,
1025 void *buffer, size_t *lenp, loff_t *ppos)
1026 {
1027 if (!watchdog_hardlockup_available && write)
1028 return -ENOTSUPP;
1029 return proc_watchdog_common(WATCHDOG_HARDLOCKUP_ENABLED,
1030 table, write, buffer, lenp, ppos);
1031 }
1032
1033 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
1034 /*
1035 * /proc/sys/kernel/soft_watchdog
1036 */
1037 static int proc_soft_watchdog(const struct ctl_table *table, int write,
1038 void *buffer, size_t *lenp, loff_t *ppos)
1039 {
1040 return proc_watchdog_common(WATCHDOG_SOFTOCKUP_ENABLED,
1041 table, write, buffer, lenp, ppos);
1042 }
1043 #endif
1044
1045 /*
1046 * /proc/sys/kernel/watchdog_thresh
1047 */
1048 static int proc_watchdog_thresh(const struct ctl_table *table, int write,
1049 void *buffer, size_t *lenp, loff_t *ppos)
1050 {
1051 int err, old;
1052
1053 mutex_lock(&watchdog_mutex);
1054
1055 old = READ_ONCE(watchdog_thresh);
1056 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1057
1058 if (!err && write && old != READ_ONCE(watchdog_thresh))
1059 proc_watchdog_update();
1060
1061 mutex_unlock(&watchdog_mutex);
1062 return err;
1063 }
1064
1065 /*
1066 * The cpumask is the mask of possible cpus that the watchdog can run
1067 * on, not the mask of cpus it is actually running on. This allows the
1068 * user to specify a mask that will include cpus that have not yet
1069 * been brought online, if desired.
1070 */
1071 static int proc_watchdog_cpumask(const struct ctl_table *table, int write,
1072 void *buffer, size_t *lenp, loff_t *ppos)
1073 {
1074 int err;
1075
1076 mutex_lock(&watchdog_mutex);
1077
1078 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
1079 if (!err && write)
1080 proc_watchdog_update();
1081
1082 mutex_unlock(&watchdog_mutex);
1083 return err;
1084 }
1085
1086 static const int sixty = 60;
1087
1088 static struct ctl_table watchdog_sysctls[] = {
1089 {
1090 .procname = "watchdog",
1091 .data = &watchdog_user_enabled,
1092 .maxlen = sizeof(int),
1093 .mode = 0644,
1094 .proc_handler = proc_watchdog,
1095 .extra1 = SYSCTL_ZERO,
1096 .extra2 = SYSCTL_ONE,
1097 },
1098 {
1099 .procname = "watchdog_thresh",
1100 .data = &watchdog_thresh,
1101 .maxlen = sizeof(int),
1102 .mode = 0644,
1103 .proc_handler = proc_watchdog_thresh,
1104 .extra1 = SYSCTL_ZERO,
1105 .extra2 = (void *)&sixty,
1106 },
1107 {
1108 .procname = "watchdog_cpumask",
1109 .data = &watchdog_cpumask_bits,
1110 .maxlen = NR_CPUS,
1111 .mode = 0644,
1112 .proc_handler = proc_watchdog_cpumask,
1113 },
1114 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
1115 {
1116 .procname = "soft_watchdog",
1117 .data = &watchdog_softlockup_user_enabled,
1118 .maxlen = sizeof(int),
1119 .mode = 0644,
1120 .proc_handler = proc_soft_watchdog,
1121 .extra1 = SYSCTL_ZERO,
1122 .extra2 = SYSCTL_ONE,
1123 },
1124 {
1125 .procname = "softlockup_panic",
1126 .data = &softlockup_panic,
1127 .maxlen = sizeof(int),
1128 .mode = 0644,
1129 .proc_handler = proc_dointvec_minmax,
1130 .extra1 = SYSCTL_ZERO,
1131 .extra2 = SYSCTL_ONE,
1132 },
1133 #ifdef CONFIG_SMP
1134 {
1135 .procname = "softlockup_all_cpu_backtrace",
1136 .data = &sysctl_softlockup_all_cpu_backtrace,
1137 .maxlen = sizeof(int),
1138 .mode = 0644,
1139 .proc_handler = proc_dointvec_minmax,
1140 .extra1 = SYSCTL_ZERO,
1141 .extra2 = SYSCTL_ONE,
1142 },
1143 #endif /* CONFIG_SMP */
1144 #endif
1145 #ifdef CONFIG_HARDLOCKUP_DETECTOR
1146 {
1147 .procname = "hardlockup_panic",
1148 .data = &hardlockup_panic,
1149 .maxlen = sizeof(int),
1150 .mode = 0644,
1151 .proc_handler = proc_dointvec_minmax,
1152 .extra1 = SYSCTL_ZERO,
1153 .extra2 = SYSCTL_ONE,
1154 },
1155 #ifdef CONFIG_SMP
1156 {
1157 .procname = "hardlockup_all_cpu_backtrace",
1158 .data = &sysctl_hardlockup_all_cpu_backtrace,
1159 .maxlen = sizeof(int),
1160 .mode = 0644,
1161 .proc_handler = proc_dointvec_minmax,
1162 .extra1 = SYSCTL_ZERO,
1163 .extra2 = SYSCTL_ONE,
1164 },
1165 #endif /* CONFIG_SMP */
1166 #endif
1167 };
1168
1169 static struct ctl_table watchdog_hardlockup_sysctl[] = {
1170 {
1171 .procname = "nmi_watchdog",
1172 .data = &watchdog_hardlockup_user_enabled,
1173 .maxlen = sizeof(int),
1174 .mode = 0444,
1175 .proc_handler = proc_nmi_watchdog,
1176 .extra1 = SYSCTL_ZERO,
1177 .extra2 = SYSCTL_ONE,
1178 },
1179 };
1180
1181 static void __init watchdog_sysctl_init(void)
1182 {
1183 register_sysctl_init("kernel", watchdog_sysctls);
1184
1185 if (watchdog_hardlockup_available)
1186 watchdog_hardlockup_sysctl[0].mode = 0644;
1187 register_sysctl_init("kernel", watchdog_hardlockup_sysctl);
1188 }
1189
1190 #else
1191 #define watchdog_sysctl_init() do { } while (0)
1192 #endif /* CONFIG_SYSCTL */
1193
1194 static void __init lockup_detector_delay_init(struct work_struct *work);
1195 static bool allow_lockup_detector_init_retry __initdata;
1196
1197 static struct work_struct detector_work __initdata =
1198 __WORK_INITIALIZER(detector_work, lockup_detector_delay_init);
1199
1200 static void __init lockup_detector_delay_init(struct work_struct *work)
1201 {
1202 int ret;
1203
1204 ret = watchdog_hardlockup_probe();
1205 if (ret) {
1206 pr_info("Delayed init of the lockup detector failed: %d\n", ret);
1207 pr_info("Hard watchdog permanently disabled\n");
1208 return;
1209 }
1210
1211 allow_lockup_detector_init_retry = false;
1212
1213 watchdog_hardlockup_available = true;
1214 lockup_detector_setup();
1215 }
1216
1217 /*
1218 * lockup_detector_retry_init - retry init lockup detector if possible.
1219 *
1220 * Retry hardlockup detector init. It is useful when it requires some
1221 * functionality that has to be initialized later on a particular
1222 * platform.
1223 */
1224 void __init lockup_detector_retry_init(void)
1225 {
1226 /* Must be called before late init calls */
1227 if (!allow_lockup_detector_init_retry)
1228 return;
1229
1230 schedule_work(&detector_work);
1231 }
1232
1233 /*
1234 * Ensure that optional delayed hardlockup init is proceed before
1235 * the init code and memory is freed.
1236 */
1237 static int __init lockup_detector_check(void)
1238 {
1239 /* Prevent any later retry. */
1240 allow_lockup_detector_init_retry = false;
1241
1242 /* Make sure no work is pending. */
1243 flush_work(&detector_work);
1244
1245 watchdog_sysctl_init();
1246
1247 return 0;
1248
1249 }
1250 late_initcall_sync(lockup_detector_check);
1251
1252 void __init lockup_detector_init(void)
1253 {
1254 if (tick_nohz_full_enabled())
1255 pr_info("Disabling watchdog on nohz_full cores by default\n");
1256
1257 cpumask_copy(&watchdog_cpumask,
1258 housekeeping_cpumask(HK_TYPE_TIMER));
1259
1260 if (!watchdog_hardlockup_probe())
1261 watchdog_hardlockup_available = true;
1262 else
1263 allow_lockup_detector_init_retry = true;
1264
1265 lockup_detector_setup();
1266 }
1267
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.