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 * this code detects hard lockups: incidents in where on a CPU
7 * the kernel does not respond to anything except NMI.
8 *
9 * Note: Most of this code is borrowed heavily from softlockup.c,
10 * so thanks to Ingo for the initial implementation.
11 * Some chunks also taken from arch/x86/kernel/apic/nmi.c, thanks
12 * to those contributors as well.
13 */
14
15#include <linux/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/delay.h>
20#include <linux/freezer.h>
21#include <linux/kthread.h>
22#include <linux/lockdep.h>
23#include <linux/notifier.h>
24#include <linux/module.h>
25#include <linux/sysctl.h>
26
27#include <asm/irq_regs.h>
28#include <linux/perf_event.h>
29
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 5 chances to
121 * increment before the hardlockup detector generates
122 * a warning
123 */
124 return get_softlockup_thresh() * (NSEC_PER_SEC / 5);
125}
126
127/* Commands for resetting the watchdog */
128static void __touch_watchdog(void)
129{
130 int this_cpu = smp_processor_id();
131
132 __this_cpu_write(watchdog_touch_ts, get_timestamp(this_cpu));
133}
134
135void touch_softlockup_watchdog(void)
136{
137 __this_cpu_write(watchdog_touch_ts, 0);
138}
139EXPORT_SYMBOL(touch_softlockup_watchdog);
140
141void touch_all_softlockup_watchdogs(void)
142{
143 int cpu;
144
145 /*
146 * this is done lockless
147 * do we care if a 0 races with a timestamp?
148 * all it means is the softlock check starts one cycle later
149 */
150 for_each_online_cpu(cpu)
151 per_cpu(watchdog_touch_ts, cpu) = 0;
152}
153
154#ifdef CONFIG_HARDLOCKUP_DETECTOR
155void touch_nmi_watchdog(void)
156{
157 if (watchdog_enabled) {
158 unsigned cpu;
159
160 for_each_present_cpu(cpu) {
161 if (per_cpu(watchdog_nmi_touch, cpu) != true)
162 per_cpu(watchdog_nmi_touch, cpu) = true;
163 }
164 }
165 touch_softlockup_watchdog();
166}
167EXPORT_SYMBOL(touch_nmi_watchdog);
168
169#endif
170
171void touch_softlockup_watchdog_sync(void)
172{
173 __raw_get_cpu_var(softlockup_touch_sync) = true;
174 __raw_get_cpu_var(watchdog_touch_ts) = 0;
175}
176
177#ifdef CONFIG_HARDLOCKUP_DETECTOR
178/* watchdog detector functions */
179static int is_hardlockup(void)
180{
181 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
182
183 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
184 return 1;
185
186 __this_cpu_write(hrtimer_interrupts_saved, hrint);
187 return 0;
188}
189#endif
190
191static int is_softlockup(unsigned long touch_ts)
192{
193 unsigned long now = get_timestamp(smp_processor_id());
194
195 /* Warn about unreasonable delays: */
196 if (time_after(now, touch_ts + get_softlockup_thresh()))
197 return now - touch_ts;
198
199 return 0;
200}
201
202#ifdef CONFIG_HARDLOCKUP_DETECTOR
203
204static struct perf_event_attr wd_hw_attr = {
205 .type = PERF_TYPE_HARDWARE,
206 .config = PERF_COUNT_HW_CPU_CYCLES,
207 .size = sizeof(struct perf_event_attr),
208 .pinned = 1,
209 .disabled = 1,
210};
211
212/* Callback function for perf event subsystem */
213static void watchdog_overflow_callback(struct perf_event *event,
214 struct perf_sample_data *data,
215 struct pt_regs *regs)
216{
217 /* Ensure the watchdog never gets throttled */
218 event->hw.interrupts = 0;
219
220 if (__this_cpu_read(watchdog_nmi_touch) == true) {
221 __this_cpu_write(watchdog_nmi_touch, false);
222 return;
223 }
224
225 /* check for a hardlockup
226 * This is done by making sure our timer interrupt
227 * is incrementing. The timer interrupt should have
228 * fired multiple times before we overflow'd. If it hasn't
229 * then this is a good indication the cpu is stuck
230 */
231 if (is_hardlockup()) {
232 int this_cpu = smp_processor_id();
233
234 /* only print hardlockups once */
235 if (__this_cpu_read(hard_watchdog_warn) == true)
236 return;
237
238 if (hardlockup_panic)
239 panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu);
240 else
241 WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu);
242
243 __this_cpu_write(hard_watchdog_warn, true);
244 return;
245 }
246
247 __this_cpu_write(hard_watchdog_warn, false);
248 return;
249}
250static void watchdog_interrupt_count(void)
251{
252 __this_cpu_inc(hrtimer_interrupts);
253}
254#else
255static inline void watchdog_interrupt_count(void) { return; }
256#endif /* CONFIG_HARDLOCKUP_DETECTOR */
257
258/* watchdog kicker functions */
259static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
260{
261 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
262 struct pt_regs *regs = get_irq_regs();
263 int duration;
264
265 /* kick the hardlockup detector */
266 watchdog_interrupt_count();
267
268 /* kick the softlockup detector */
269 wake_up_process(__this_cpu_read(softlockup_watchdog));
270
271 /* .. and repeat */
272 hrtimer_forward_now(hrtimer, ns_to_ktime(get_sample_period()));
273
274 if (touch_ts == 0) {
275 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
276 /*
277 * If the time stamp was touched atomically
278 * make sure the scheduler tick is up to date.
279 */
280 __this_cpu_write(softlockup_touch_sync, false);
281 sched_clock_tick();
282 }
283 __touch_watchdog();
284 return HRTIMER_RESTART;
285 }
286
287 /* check for a softlockup
288 * This is done by making sure a high priority task is
289 * being scheduled. The task touches the watchdog to
290 * indicate it is getting cpu time. If it hasn't then
291 * this is a good indication some task is hogging the cpu
292 */
293 duration = is_softlockup(touch_ts);
294 if (unlikely(duration)) {
295 /* only warn once */
296 if (__this_cpu_read(soft_watchdog_warn) == true)
297 return HRTIMER_RESTART;
298
299 printk(KERN_ERR "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
300 smp_processor_id(), duration,
301 current->comm, task_pid_nr(current));
302 print_modules();
303 print_irqtrace_events(current);
304 if (regs)
305 show_regs(regs);
306 else
307 dump_stack();
308
309 if (softlockup_panic)
310 panic("softlockup: hung tasks");
311 __this_cpu_write(soft_watchdog_warn, true);
312 } else
313 __this_cpu_write(soft_watchdog_warn, false);
314
315 return HRTIMER_RESTART;
316}
317
318
319/*
320 * The watchdog thread - touches the timestamp.
321 */
322static int watchdog(void *unused)
323{
324 static struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 };
325 struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer);
326
327 sched_setscheduler(current, SCHED_FIFO, ¶m);
328
329 /* initialize timestamp */
330 __touch_watchdog();
331
332 /* kick off the timer for the hardlockup detector */
333 /* done here because hrtimer_start can only pin to smp_processor_id() */
334 hrtimer_start(hrtimer, ns_to_ktime(get_sample_period()),
335 HRTIMER_MODE_REL_PINNED);
336
337 set_current_state(TASK_INTERRUPTIBLE);
338 /*
339 * Run briefly once per second to reset the softlockup timestamp.
340 * If this gets delayed for more than 60 seconds then the
341 * debug-printout triggers in watchdog_timer_fn().
342 */
343 while (!kthread_should_stop()) {
344 __touch_watchdog();
345 schedule();
346
347 if (kthread_should_stop())
348 break;
349
350 set_current_state(TASK_INTERRUPTIBLE);
351 }
352 __set_current_state(TASK_RUNNING);
353
354 return 0;
355}
356
357
358#ifdef CONFIG_HARDLOCKUP_DETECTOR
359static int watchdog_nmi_enable(int cpu)
360{
361 struct perf_event_attr *wd_attr;
362 struct perf_event *event = per_cpu(watchdog_ev, cpu);
363
364 /* is it already setup and enabled? */
365 if (event && event->state > PERF_EVENT_STATE_OFF)
366 goto out;
367
368 /* it is setup but not enabled */
369 if (event != NULL)
370 goto out_enable;
371
372 wd_attr = &wd_hw_attr;
373 wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh);
374
375 /* Try to register using hardware perf events */
376 event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL);
377 if (!IS_ERR(event)) {
378 printk(KERN_INFO "NMI watchdog enabled, takes one hw-pmu counter.\n");
379 goto out_save;
380 }
381
382
383 /* vary the KERN level based on the returned errno */
384 if (PTR_ERR(event) == -EOPNOTSUPP)
385 printk(KERN_INFO "NMI watchdog disabled (cpu%i): not supported (no LAPIC?)\n", cpu);
386 else if (PTR_ERR(event) == -ENOENT)
387 printk(KERN_WARNING "NMI watchdog disabled (cpu%i): hardware events not enabled\n", cpu);
388 else
389 printk(KERN_ERR "NMI watchdog disabled (cpu%i): unable to create perf event: %ld\n", cpu, PTR_ERR(event));
390 return PTR_ERR(event);
391
392 /* success path */
393out_save:
394 per_cpu(watchdog_ev, cpu) = event;
395out_enable:
396 perf_event_enable(per_cpu(watchdog_ev, cpu));
397out:
398 return 0;
399}
400
401static void watchdog_nmi_disable(int cpu)
402{
403 struct perf_event *event = per_cpu(watchdog_ev, cpu);
404
405 if (event) {
406 perf_event_disable(event);
407 per_cpu(watchdog_ev, cpu) = NULL;
408
409 /* should be in cleanup, but blocks oprofile */
410 perf_event_release_kernel(event);
411 }
412 return;
413}
414#else
415static int watchdog_nmi_enable(int cpu) { return 0; }
416static void watchdog_nmi_disable(int cpu) { return; }
417#endif /* CONFIG_HARDLOCKUP_DETECTOR */
418
419/* prepare/enable/disable routines */
420static void watchdog_prepare_cpu(int cpu)
421{
422 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
423
424 WARN_ON(per_cpu(softlockup_watchdog, cpu));
425 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
426 hrtimer->function = watchdog_timer_fn;
427}
428
429static int watchdog_enable(int cpu)
430{
431 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
432 int err = 0;
433
434 /* enable the perf event */
435 err = watchdog_nmi_enable(cpu);
436
437 /* Regardless of err above, fall through and start softlockup */
438
439 /* create the watchdog thread */
440 if (!p) {
441 p = kthread_create(watchdog, (void *)(unsigned long)cpu, "watchdog/%d", cpu);
442 if (IS_ERR(p)) {
443 printk(KERN_ERR "softlockup watchdog for %i failed\n", cpu);
444 if (!err) {
445 /* if hardlockup hasn't already set this */
446 err = PTR_ERR(p);
447 /* and disable the perf event */
448 watchdog_nmi_disable(cpu);
449 }
450 goto out;
451 }
452 kthread_bind(p, cpu);
453 per_cpu(watchdog_touch_ts, cpu) = 0;
454 per_cpu(softlockup_watchdog, cpu) = p;
455 wake_up_process(p);
456 }
457
458out:
459 return err;
460}
461
462static void watchdog_disable(int cpu)
463{
464 struct task_struct *p = per_cpu(softlockup_watchdog, cpu);
465 struct hrtimer *hrtimer = &per_cpu(watchdog_hrtimer, cpu);
466
467 /*
468 * cancel the timer first to stop incrementing the stats
469 * and waking up the kthread
470 */
471 hrtimer_cancel(hrtimer);
472
473 /* disable the perf event */
474 watchdog_nmi_disable(cpu);
475
476 /* stop the watchdog thread */
477 if (p) {
478 per_cpu(softlockup_watchdog, cpu) = NULL;
479 kthread_stop(p);
480 }
481}
482
483static void watchdog_enable_all_cpus(void)
484{
485 int cpu;
486
487 watchdog_enabled = 0;
488
489 for_each_online_cpu(cpu)
490 if (!watchdog_enable(cpu))
491 /* if any cpu succeeds, watchdog is considered
492 enabled for the system */
493 watchdog_enabled = 1;
494
495 if (!watchdog_enabled)
496 printk(KERN_ERR "watchdog: failed to be enabled on some cpus\n");
497
498}
499
500static void watchdog_disable_all_cpus(void)
501{
502 int cpu;
503
504 for_each_online_cpu(cpu)
505 watchdog_disable(cpu);
506
507 /* if all watchdogs are disabled, then they are disabled for the system */
508 watchdog_enabled = 0;
509}
510
511
512/* sysctl functions */
513#ifdef CONFIG_SYSCTL
514/*
515 * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh
516 */
517
518int proc_dowatchdog(struct ctl_table *table, int write,
519 void __user *buffer, size_t *lenp, loff_t *ppos)
520{
521 int ret;
522
523 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
524 if (ret || !write)
525 goto out;
526
527 if (watchdog_enabled && watchdog_thresh)
528 watchdog_enable_all_cpus();
529 else
530 watchdog_disable_all_cpus();
531
532out:
533 return ret;
534}
535#endif /* CONFIG_SYSCTL */
536
537
538/*
539 * Create/destroy watchdog threads as CPUs come and go:
540 */
541static int __cpuinit
542cpu_callback(struct notifier_block *nfb, unsigned long action, void *hcpu)
543{
544 int hotcpu = (unsigned long)hcpu;
545
546 switch (action) {
547 case CPU_UP_PREPARE:
548 case CPU_UP_PREPARE_FROZEN:
549 watchdog_prepare_cpu(hotcpu);
550 break;
551 case CPU_ONLINE:
552 case CPU_ONLINE_FROZEN:
553 if (watchdog_enabled)
554 watchdog_enable(hotcpu);
555 break;
556#ifdef CONFIG_HOTPLUG_CPU
557 case CPU_UP_CANCELED:
558 case CPU_UP_CANCELED_FROZEN:
559 watchdog_disable(hotcpu);
560 break;
561 case CPU_DEAD:
562 case CPU_DEAD_FROZEN:
563 watchdog_disable(hotcpu);
564 break;
565#endif /* CONFIG_HOTPLUG_CPU */
566 }
567
568 /*
569 * hardlockup and softlockup are not important enough
570 * to block cpu bring up. Just always succeed and
571 * rely on printk output to flag problems.
572 */
573 return NOTIFY_OK;
574}
575
576static struct notifier_block __cpuinitdata cpu_nfb = {
577 .notifier_call = cpu_callback
578};
579
580void __init lockup_detector_init(void)
581{
582 void *cpu = (void *)(long)smp_processor_id();
583 int err;
584
585 err = cpu_callback(&cpu_nfb, CPU_UP_PREPARE, cpu);
586 WARN_ON(notifier_to_errno(err));
587
588 cpu_callback(&cpu_nfb, CPU_ONLINE, cpu);
589 register_cpu_notifier(&cpu_nfb);
590
591 return;
592}
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/mm.h>
16#include <linux/cpu.h>
17#include <linux/nmi.h>
18#include <linux/init.h>
19#include <linux/module.h>
20#include <linux/sysctl.h>
21#include <linux/smpboot.h>
22#include <linux/sched/rt.h>
23#include <uapi/linux/sched/types.h>
24#include <linux/tick.h>
25#include <linux/workqueue.h>
26#include <linux/sched/clock.h>
27#include <linux/sched/debug.h>
28#include <linux/sched/isolation.h>
29
30#include <asm/irq_regs.h>
31#include <linux/kvm_para.h>
32#include <linux/kthread.h>
33
34static DEFINE_MUTEX(watchdog_mutex);
35
36#if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
37# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
38# define NMI_WATCHDOG_DEFAULT 1
39#else
40# define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
41# define NMI_WATCHDOG_DEFAULT 0
42#endif
43
44unsigned long __read_mostly watchdog_enabled;
45int __read_mostly watchdog_user_enabled = 1;
46int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
47int __read_mostly soft_watchdog_user_enabled = 1;
48int __read_mostly watchdog_thresh = 10;
49int __read_mostly nmi_watchdog_available;
50
51struct cpumask watchdog_allowed_mask __read_mostly;
52
53struct cpumask watchdog_cpumask __read_mostly;
54unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask);
55
56#ifdef CONFIG_HARDLOCKUP_DETECTOR
57/*
58 * Should we panic when a soft-lockup or hard-lockup occurs:
59 */
60unsigned int __read_mostly hardlockup_panic =
61 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
62/*
63 * We may not want to enable hard lockup detection by default in all cases,
64 * for example when running the kernel as a guest on a hypervisor. In these
65 * cases this function can be called to disable hard lockup detection. This
66 * function should only be executed once by the boot processor before the
67 * kernel command line parameters are parsed, because otherwise it is not
68 * possible to override this in hardlockup_panic_setup().
69 */
70void __init hardlockup_detector_disable(void)
71{
72 nmi_watchdog_user_enabled = 0;
73}
74
75static int __init hardlockup_panic_setup(char *str)
76{
77 if (!strncmp(str, "panic", 5))
78 hardlockup_panic = 1;
79 else if (!strncmp(str, "nopanic", 7))
80 hardlockup_panic = 0;
81 else if (!strncmp(str, "0", 1))
82 nmi_watchdog_user_enabled = 0;
83 else if (!strncmp(str, "1", 1))
84 nmi_watchdog_user_enabled = 1;
85 return 1;
86}
87__setup("nmi_watchdog=", hardlockup_panic_setup);
88
89# ifdef CONFIG_SMP
90int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
91
92static int __init hardlockup_all_cpu_backtrace_setup(char *str)
93{
94 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
95 return 1;
96}
97__setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
98# endif /* CONFIG_SMP */
99#endif /* CONFIG_HARDLOCKUP_DETECTOR */
100
101/*
102 * These functions can be overridden if an architecture implements its
103 * own hardlockup detector.
104 *
105 * watchdog_nmi_enable/disable can be implemented to start and stop when
106 * softlockup watchdog threads start and stop. The arch must select the
107 * SOFTLOCKUP_DETECTOR Kconfig.
108 */
109int __weak watchdog_nmi_enable(unsigned int cpu)
110{
111 hardlockup_detector_perf_enable();
112 return 0;
113}
114
115void __weak watchdog_nmi_disable(unsigned int cpu)
116{
117 hardlockup_detector_perf_disable();
118}
119
120/* Return 0, if a NMI watchdog is available. Error code otherwise */
121int __weak __init watchdog_nmi_probe(void)
122{
123 return hardlockup_detector_perf_init();
124}
125
126/**
127 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
128 *
129 * The reconfiguration steps are:
130 * watchdog_nmi_stop();
131 * update_variables();
132 * watchdog_nmi_start();
133 */
134void __weak watchdog_nmi_stop(void) { }
135
136/**
137 * watchdog_nmi_start - Start the watchdog after reconfiguration
138 *
139 * Counterpart to watchdog_nmi_stop().
140 *
141 * The following variables have been updated in update_variables() and
142 * contain the currently valid configuration:
143 * - watchdog_enabled
144 * - watchdog_thresh
145 * - watchdog_cpumask
146 */
147void __weak watchdog_nmi_start(void) { }
148
149/**
150 * lockup_detector_update_enable - Update the sysctl enable bit
151 *
152 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
153 * can't race with watchdog_nmi_disable().
154 */
155static void lockup_detector_update_enable(void)
156{
157 watchdog_enabled = 0;
158 if (!watchdog_user_enabled)
159 return;
160 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
161 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
162 if (soft_watchdog_user_enabled)
163 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
164}
165
166#ifdef CONFIG_SOFTLOCKUP_DETECTOR
167
168/* Global variables, exported for sysctl */
169unsigned int __read_mostly softlockup_panic =
170 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
171
172static bool softlockup_threads_initialized __read_mostly;
173static u64 __read_mostly sample_period;
174
175static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
176static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
177static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
178static DEFINE_PER_CPU(bool, softlockup_touch_sync);
179static DEFINE_PER_CPU(bool, soft_watchdog_warn);
180static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
181static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
182static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
183static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
184static unsigned long soft_lockup_nmi_warn;
185
186static int __init softlockup_panic_setup(char *str)
187{
188 softlockup_panic = simple_strtoul(str, NULL, 0);
189 return 1;
190}
191__setup("softlockup_panic=", softlockup_panic_setup);
192
193static int __init nowatchdog_setup(char *str)
194{
195 watchdog_user_enabled = 0;
196 return 1;
197}
198__setup("nowatchdog", nowatchdog_setup);
199
200static int __init nosoftlockup_setup(char *str)
201{
202 soft_watchdog_user_enabled = 0;
203 return 1;
204}
205__setup("nosoftlockup", nosoftlockup_setup);
206
207#ifdef CONFIG_SMP
208int __read_mostly sysctl_softlockup_all_cpu_backtrace;
209
210static int __init softlockup_all_cpu_backtrace_setup(char *str)
211{
212 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
213 return 1;
214}
215__setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
216#endif
217
218static void __lockup_detector_cleanup(void);
219
220/*
221 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
222 * lockups can have false positives under extreme conditions. So we generally
223 * want a higher threshold for soft lockups than for hard lockups. So we couple
224 * the thresholds with a factor: we make the soft threshold twice the amount of
225 * time the hard threshold is.
226 */
227static int get_softlockup_thresh(void)
228{
229 return watchdog_thresh * 2;
230}
231
232/*
233 * Returns seconds, approximately. We don't need nanosecond
234 * resolution, and we don't need to waste time with a big divide when
235 * 2^30ns == 1.074s.
236 */
237static unsigned long get_timestamp(void)
238{
239 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
240}
241
242static void set_sample_period(void)
243{
244 /*
245 * convert watchdog_thresh from seconds to ns
246 * the divide by 5 is to give hrtimer several chances (two
247 * or three with the current relation between the soft
248 * and hard thresholds) to increment before the
249 * hardlockup detector generates a warning
250 */
251 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
252 watchdog_update_hrtimer_threshold(sample_period);
253}
254
255/* Commands for resetting the watchdog */
256static void __touch_watchdog(void)
257{
258 __this_cpu_write(watchdog_touch_ts, get_timestamp());
259}
260
261/**
262 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
263 *
264 * Call when the scheduler may have stalled for legitimate reasons
265 * preventing the watchdog task from executing - e.g. the scheduler
266 * entering idle state. This should only be used for scheduler events.
267 * Use touch_softlockup_watchdog() for everything else.
268 */
269void touch_softlockup_watchdog_sched(void)
270{
271 /*
272 * Preemption can be enabled. It doesn't matter which CPU's timestamp
273 * gets zeroed here, so use the raw_ operation.
274 */
275 raw_cpu_write(watchdog_touch_ts, 0);
276}
277
278void touch_softlockup_watchdog(void)
279{
280 touch_softlockup_watchdog_sched();
281 wq_watchdog_touch(raw_smp_processor_id());
282}
283EXPORT_SYMBOL(touch_softlockup_watchdog);
284
285void touch_all_softlockup_watchdogs(void)
286{
287 int cpu;
288
289 /*
290 * watchdog_mutex cannpt be taken here, as this might be called
291 * from (soft)interrupt context, so the access to
292 * watchdog_allowed_cpumask might race with a concurrent update.
293 *
294 * The watchdog time stamp can race against a concurrent real
295 * update as well, the only side effect might be a cycle delay for
296 * the softlockup check.
297 */
298 for_each_cpu(cpu, &watchdog_allowed_mask)
299 per_cpu(watchdog_touch_ts, cpu) = 0;
300 wq_watchdog_touch(-1);
301}
302
303void touch_softlockup_watchdog_sync(void)
304{
305 __this_cpu_write(softlockup_touch_sync, true);
306 __this_cpu_write(watchdog_touch_ts, 0);
307}
308
309static int is_softlockup(unsigned long touch_ts)
310{
311 unsigned long now = get_timestamp();
312
313 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
314 /* Warn about unreasonable delays. */
315 if (time_after(now, touch_ts + get_softlockup_thresh()))
316 return now - touch_ts;
317 }
318 return 0;
319}
320
321/* watchdog detector functions */
322bool is_hardlockup(void)
323{
324 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
325
326 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
327 return true;
328
329 __this_cpu_write(hrtimer_interrupts_saved, hrint);
330 return false;
331}
332
333static void watchdog_interrupt_count(void)
334{
335 __this_cpu_inc(hrtimer_interrupts);
336}
337
338/* watchdog kicker functions */
339static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
340{
341 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
342 struct pt_regs *regs = get_irq_regs();
343 int duration;
344 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
345
346 if (!watchdog_enabled)
347 return HRTIMER_NORESTART;
348
349 /* kick the hardlockup detector */
350 watchdog_interrupt_count();
351
352 /* kick the softlockup detector */
353 wake_up_process(__this_cpu_read(softlockup_watchdog));
354
355 /* .. and repeat */
356 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
357
358 if (touch_ts == 0) {
359 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
360 /*
361 * If the time stamp was touched atomically
362 * make sure the scheduler tick is up to date.
363 */
364 __this_cpu_write(softlockup_touch_sync, false);
365 sched_clock_tick();
366 }
367
368 /* Clear the guest paused flag on watchdog reset */
369 kvm_check_and_clear_guest_paused();
370 __touch_watchdog();
371 return HRTIMER_RESTART;
372 }
373
374 /* check for a softlockup
375 * This is done by making sure a high priority task is
376 * being scheduled. The task touches the watchdog to
377 * indicate it is getting cpu time. If it hasn't then
378 * this is a good indication some task is hogging the cpu
379 */
380 duration = is_softlockup(touch_ts);
381 if (unlikely(duration)) {
382 /*
383 * If a virtual machine is stopped by the host it can look to
384 * the watchdog like a soft lockup, check to see if the host
385 * stopped the vm before we issue the warning
386 */
387 if (kvm_check_and_clear_guest_paused())
388 return HRTIMER_RESTART;
389
390 /* only warn once */
391 if (__this_cpu_read(soft_watchdog_warn) == true) {
392 /*
393 * When multiple processes are causing softlockups the
394 * softlockup detector only warns on the first one
395 * because the code relies on a full quiet cycle to
396 * re-arm. The second process prevents the quiet cycle
397 * and never gets reported. Use task pointers to detect
398 * this.
399 */
400 if (__this_cpu_read(softlockup_task_ptr_saved) !=
401 current) {
402 __this_cpu_write(soft_watchdog_warn, false);
403 __touch_watchdog();
404 }
405 return HRTIMER_RESTART;
406 }
407
408 if (softlockup_all_cpu_backtrace) {
409 /* Prevent multiple soft-lockup reports if one cpu is already
410 * engaged in dumping cpu back traces
411 */
412 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
413 /* Someone else will report us. Let's give up */
414 __this_cpu_write(soft_watchdog_warn, true);
415 return HRTIMER_RESTART;
416 }
417 }
418
419 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
420 smp_processor_id(), duration,
421 current->comm, task_pid_nr(current));
422 __this_cpu_write(softlockup_task_ptr_saved, current);
423 print_modules();
424 print_irqtrace_events(current);
425 if (regs)
426 show_regs(regs);
427 else
428 dump_stack();
429
430 if (softlockup_all_cpu_backtrace) {
431 /* Avoid generating two back traces for current
432 * given that one is already made above
433 */
434 trigger_allbutself_cpu_backtrace();
435
436 clear_bit(0, &soft_lockup_nmi_warn);
437 /* Barrier to sync with other cpus */
438 smp_mb__after_atomic();
439 }
440
441 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
442 if (softlockup_panic)
443 panic("softlockup: hung tasks");
444 __this_cpu_write(soft_watchdog_warn, true);
445 } else
446 __this_cpu_write(soft_watchdog_warn, false);
447
448 return HRTIMER_RESTART;
449}
450
451static void watchdog_set_prio(unsigned int policy, unsigned int prio)
452{
453 struct sched_param param = { .sched_priority = prio };
454
455 sched_setscheduler(current, policy, ¶m);
456}
457
458static void watchdog_enable(unsigned int cpu)
459{
460 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
461
462 /*
463 * Start the timer first to prevent the NMI watchdog triggering
464 * before the timer has a chance to fire.
465 */
466 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
467 hrtimer->function = watchdog_timer_fn;
468 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
469 HRTIMER_MODE_REL_PINNED);
470
471 /* Initialize timestamp */
472 __touch_watchdog();
473 /* Enable the perf event */
474 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
475 watchdog_nmi_enable(cpu);
476
477 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
478}
479
480static void watchdog_disable(unsigned int cpu)
481{
482 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
483
484 watchdog_set_prio(SCHED_NORMAL, 0);
485 /*
486 * Disable the perf event first. That prevents that a large delay
487 * between disabling the timer and disabling the perf event causes
488 * the perf NMI to detect a false positive.
489 */
490 watchdog_nmi_disable(cpu);
491 hrtimer_cancel(hrtimer);
492}
493
494static void watchdog_cleanup(unsigned int cpu, bool online)
495{
496 watchdog_disable(cpu);
497}
498
499static int watchdog_should_run(unsigned int cpu)
500{
501 return __this_cpu_read(hrtimer_interrupts) !=
502 __this_cpu_read(soft_lockup_hrtimer_cnt);
503}
504
505/*
506 * The watchdog thread function - touches the timestamp.
507 *
508 * It only runs once every sample_period seconds (4 seconds by
509 * default) to reset the softlockup timestamp. If this gets delayed
510 * for more than 2*watchdog_thresh seconds then the debug-printout
511 * triggers in watchdog_timer_fn().
512 */
513static void watchdog(unsigned int cpu)
514{
515 __this_cpu_write(soft_lockup_hrtimer_cnt,
516 __this_cpu_read(hrtimer_interrupts));
517 __touch_watchdog();
518}
519
520static struct smp_hotplug_thread watchdog_threads = {
521 .store = &softlockup_watchdog,
522 .thread_should_run = watchdog_should_run,
523 .thread_fn = watchdog,
524 .thread_comm = "watchdog/%u",
525 .setup = watchdog_enable,
526 .cleanup = watchdog_cleanup,
527 .park = watchdog_disable,
528 .unpark = watchdog_enable,
529};
530
531static void softlockup_update_smpboot_threads(void)
532{
533 lockdep_assert_held(&watchdog_mutex);
534
535 if (!softlockup_threads_initialized)
536 return;
537
538 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
539 &watchdog_allowed_mask);
540}
541
542/* Temporarily park all watchdog threads */
543static void softlockup_park_all_threads(void)
544{
545 cpumask_clear(&watchdog_allowed_mask);
546 softlockup_update_smpboot_threads();
547}
548
549/* Unpark enabled threads */
550static void softlockup_unpark_threads(void)
551{
552 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
553 softlockup_update_smpboot_threads();
554}
555
556static void lockup_detector_reconfigure(void)
557{
558 cpus_read_lock();
559 watchdog_nmi_stop();
560 softlockup_park_all_threads();
561 set_sample_period();
562 lockup_detector_update_enable();
563 if (watchdog_enabled && watchdog_thresh)
564 softlockup_unpark_threads();
565 watchdog_nmi_start();
566 cpus_read_unlock();
567 /*
568 * Must be called outside the cpus locked section to prevent
569 * recursive locking in the perf code.
570 */
571 __lockup_detector_cleanup();
572}
573
574/*
575 * Create the watchdog thread infrastructure and configure the detector(s).
576 *
577 * The threads are not unparked as watchdog_allowed_mask is empty. When
578 * the threads are sucessfully initialized, take the proper locks and
579 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
580 */
581static __init void lockup_detector_setup(void)
582{
583 int ret;
584
585 /*
586 * If sysctl is off and watchdog got disabled on the command line,
587 * nothing to do here.
588 */
589 lockup_detector_update_enable();
590
591 if (!IS_ENABLED(CONFIG_SYSCTL) &&
592 !(watchdog_enabled && watchdog_thresh))
593 return;
594
595 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
596 &watchdog_allowed_mask);
597 if (ret) {
598 pr_err("Failed to initialize soft lockup detector threads\n");
599 return;
600 }
601
602 mutex_lock(&watchdog_mutex);
603 softlockup_threads_initialized = true;
604 lockup_detector_reconfigure();
605 mutex_unlock(&watchdog_mutex);
606}
607
608#else /* CONFIG_SOFTLOCKUP_DETECTOR */
609static inline int watchdog_park_threads(void) { return 0; }
610static inline void watchdog_unpark_threads(void) { }
611static inline int watchdog_enable_all_cpus(void) { return 0; }
612static inline void watchdog_disable_all_cpus(void) { }
613static void lockup_detector_reconfigure(void)
614{
615 cpus_read_lock();
616 watchdog_nmi_stop();
617 lockup_detector_update_enable();
618 watchdog_nmi_start();
619 cpus_read_unlock();
620}
621static inline void lockup_detector_setup(void)
622{
623 lockup_detector_reconfigure();
624}
625#endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
626
627static void __lockup_detector_cleanup(void)
628{
629 lockdep_assert_held(&watchdog_mutex);
630 hardlockup_detector_perf_cleanup();
631}
632
633/**
634 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
635 *
636 * Caller must not hold the cpu hotplug rwsem.
637 */
638void lockup_detector_cleanup(void)
639{
640 mutex_lock(&watchdog_mutex);
641 __lockup_detector_cleanup();
642 mutex_unlock(&watchdog_mutex);
643}
644
645/**
646 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
647 *
648 * Special interface for parisc. It prevents lockup detector warnings from
649 * the default pm_poweroff() function which busy loops forever.
650 */
651void lockup_detector_soft_poweroff(void)
652{
653 watchdog_enabled = 0;
654}
655
656#ifdef CONFIG_SYSCTL
657
658/* Propagate any changes to the watchdog threads */
659static void proc_watchdog_update(void)
660{
661 /* Remove impossible cpus to keep sysctl output clean. */
662 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
663 lockup_detector_reconfigure();
664}
665
666/*
667 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
668 *
669 * caller | table->data points to | 'which'
670 * -------------------|----------------------------|--------------------------
671 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
672 * | | SOFT_WATCHDOG_ENABLED
673 * -------------------|----------------------------|--------------------------
674 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
675 * -------------------|----------------------------|--------------------------
676 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
677 */
678static int proc_watchdog_common(int which, struct ctl_table *table, int write,
679 void __user *buffer, size_t *lenp, loff_t *ppos)
680{
681 int err, old, *param = table->data;
682
683 mutex_lock(&watchdog_mutex);
684
685 if (!write) {
686 /*
687 * On read synchronize the userspace interface. This is a
688 * racy snapshot.
689 */
690 *param = (watchdog_enabled & which) != 0;
691 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
692 } else {
693 old = READ_ONCE(*param);
694 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
695 if (!err && old != READ_ONCE(*param))
696 proc_watchdog_update();
697 }
698 mutex_unlock(&watchdog_mutex);
699 return err;
700}
701
702/*
703 * /proc/sys/kernel/watchdog
704 */
705int proc_watchdog(struct ctl_table *table, int write,
706 void __user *buffer, size_t *lenp, loff_t *ppos)
707{
708 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
709 table, write, buffer, lenp, ppos);
710}
711
712/*
713 * /proc/sys/kernel/nmi_watchdog
714 */
715int proc_nmi_watchdog(struct ctl_table *table, int write,
716 void __user *buffer, size_t *lenp, loff_t *ppos)
717{
718 if (!nmi_watchdog_available && write)
719 return -ENOTSUPP;
720 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
721 table, write, buffer, lenp, ppos);
722}
723
724/*
725 * /proc/sys/kernel/soft_watchdog
726 */
727int proc_soft_watchdog(struct ctl_table *table, int write,
728 void __user *buffer, size_t *lenp, loff_t *ppos)
729{
730 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
731 table, write, buffer, lenp, ppos);
732}
733
734/*
735 * /proc/sys/kernel/watchdog_thresh
736 */
737int proc_watchdog_thresh(struct ctl_table *table, int write,
738 void __user *buffer, size_t *lenp, loff_t *ppos)
739{
740 int err, old;
741
742 mutex_lock(&watchdog_mutex);
743
744 old = READ_ONCE(watchdog_thresh);
745 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
746
747 if (!err && write && old != READ_ONCE(watchdog_thresh))
748 proc_watchdog_update();
749
750 mutex_unlock(&watchdog_mutex);
751 return err;
752}
753
754/*
755 * The cpumask is the mask of possible cpus that the watchdog can run
756 * on, not the mask of cpus it is actually running on. This allows the
757 * user to specify a mask that will include cpus that have not yet
758 * been brought online, if desired.
759 */
760int proc_watchdog_cpumask(struct ctl_table *table, int write,
761 void __user *buffer, size_t *lenp, loff_t *ppos)
762{
763 int err;
764
765 mutex_lock(&watchdog_mutex);
766
767 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
768 if (!err && write)
769 proc_watchdog_update();
770
771 mutex_unlock(&watchdog_mutex);
772 return err;
773}
774#endif /* CONFIG_SYSCTL */
775
776void __init lockup_detector_init(void)
777{
778 if (tick_nohz_full_enabled())
779 pr_info("Disabling watchdog on nohz_full cores by default\n");
780
781 cpumask_copy(&watchdog_cpumask,
782 housekeeping_cpumask(HK_FLAG_TIMER));
783
784 if (!watchdog_nmi_probe())
785 nmi_watchdog_available = true;
786 lockup_detector_setup();
787}