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
29 #include <asm/irq_regs.h>
30 #include <linux/kvm_para.h>
31 #include <linux/kthread.h>
32
33 static DEFINE_MUTEX(watchdog_mutex);
34
35 #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG)
36 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED)
37 # define NMI_WATCHDOG_DEFAULT 1
38 #else
39 # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED)
40 # define NMI_WATCHDOG_DEFAULT 0
41 #endif
42
43 unsigned long __read_mostly watchdog_enabled;
44 int __read_mostly watchdog_user_enabled = 1;
45 int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT;
46 int __read_mostly soft_watchdog_user_enabled = 1;
47 int __read_mostly watchdog_thresh = 10;
48 int __read_mostly nmi_watchdog_available;
49
50 struct cpumask watchdog_allowed_mask __read_mostly;
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 * Should we panic when a soft-lockup or hard-lockup occurs:
58 */
59 unsigned int __read_mostly hardlockup_panic =
60 CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE;
61 /*
62 * We may not want to enable hard lockup detection by default in all cases,
63 * for example when running the kernel as a guest on a hypervisor. In these
64 * cases this function can be called to disable hard lockup detection. This
65 * function should only be executed once by the boot processor before the
66 * kernel command line parameters are parsed, because otherwise it is not
67 * possible to override this in hardlockup_panic_setup().
68 */
hardlockup_detector_disable(void)69 void __init hardlockup_detector_disable(void)
70 {
71 nmi_watchdog_user_enabled = 0;
72 }
73
hardlockup_panic_setup(char * str)74 static int __init hardlockup_panic_setup(char *str)
75 {
76 if (!strncmp(str, "panic", 5))
77 hardlockup_panic = 1;
78 else if (!strncmp(str, "nopanic", 7))
79 hardlockup_panic = 0;
80 else if (!strncmp(str, "0", 1))
81 nmi_watchdog_user_enabled = 0;
82 else if (!strncmp(str, "1", 1))
83 nmi_watchdog_user_enabled = 1;
84 return 1;
85 }
86 __setup("nmi_watchdog=", hardlockup_panic_setup);
87
88 # ifdef CONFIG_SMP
89 int __read_mostly sysctl_hardlockup_all_cpu_backtrace;
90
hardlockup_all_cpu_backtrace_setup(char * str)91 static int __init hardlockup_all_cpu_backtrace_setup(char *str)
92 {
93 sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
94 return 1;
95 }
96 __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup);
97 # endif /* CONFIG_SMP */
98 #endif /* CONFIG_HARDLOCKUP_DETECTOR */
99
100 /*
101 * These functions can be overridden if an architecture implements its
102 * own hardlockup detector.
103 *
104 * watchdog_nmi_enable/disable can be implemented to start and stop when
105 * softlockup watchdog threads start and stop. The arch must select the
106 * SOFTLOCKUP_DETECTOR Kconfig.
107 */
watchdog_nmi_enable(unsigned int cpu)108 int __weak watchdog_nmi_enable(unsigned int cpu)
109 {
110 hardlockup_detector_perf_enable();
111 return 0;
112 }
113
watchdog_nmi_disable(unsigned int cpu)114 void __weak watchdog_nmi_disable(unsigned int cpu)
115 {
116 hardlockup_detector_perf_disable();
117 }
118
119 /* Return 0, if a NMI watchdog is available. Error code otherwise */
watchdog_nmi_probe(void)120 int __weak __init watchdog_nmi_probe(void)
121 {
122 return hardlockup_detector_perf_init();
123 }
124
125 /**
126 * watchdog_nmi_stop - Stop the watchdog for reconfiguration
127 *
128 * The reconfiguration steps are:
129 * watchdog_nmi_stop();
130 * update_variables();
131 * watchdog_nmi_start();
132 */
watchdog_nmi_stop(void)133 void __weak watchdog_nmi_stop(void) { }
134
135 /**
136 * watchdog_nmi_start - Start the watchdog after reconfiguration
137 *
138 * Counterpart to watchdog_nmi_stop().
139 *
140 * The following variables have been updated in update_variables() and
141 * contain the currently valid configuration:
142 * - watchdog_enabled
143 * - watchdog_thresh
144 * - watchdog_cpumask
145 */
watchdog_nmi_start(void)146 void __weak watchdog_nmi_start(void) { }
147
148 /**
149 * lockup_detector_update_enable - Update the sysctl enable bit
150 *
151 * Caller needs to make sure that the NMI/perf watchdogs are off, so this
152 * can't race with watchdog_nmi_disable().
153 */
lockup_detector_update_enable(void)154 static void lockup_detector_update_enable(void)
155 {
156 watchdog_enabled = 0;
157 if (!watchdog_user_enabled)
158 return;
159 if (nmi_watchdog_available && nmi_watchdog_user_enabled)
160 watchdog_enabled |= NMI_WATCHDOG_ENABLED;
161 if (soft_watchdog_user_enabled)
162 watchdog_enabled |= SOFT_WATCHDOG_ENABLED;
163 }
164
165 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
166
167 #define SOFTLOCKUP_RESET ULONG_MAX
168
169 /* Global variables, exported for sysctl */
170 unsigned int __read_mostly softlockup_panic =
171 CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE;
172
173 static bool softlockup_threads_initialized __read_mostly;
174 static u64 __read_mostly sample_period;
175
176 static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts);
177 static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog);
178 static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer);
179 static DEFINE_PER_CPU(bool, softlockup_touch_sync);
180 static DEFINE_PER_CPU(bool, soft_watchdog_warn);
181 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts);
182 static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt);
183 static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved);
184 static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved);
185 static unsigned long soft_lockup_nmi_warn;
186
softlockup_panic_setup(char * str)187 static int __init softlockup_panic_setup(char *str)
188 {
189 softlockup_panic = simple_strtoul(str, NULL, 0);
190 return 1;
191 }
192 __setup("softlockup_panic=", softlockup_panic_setup);
193
nowatchdog_setup(char * str)194 static int __init nowatchdog_setup(char *str)
195 {
196 watchdog_user_enabled = 0;
197 return 1;
198 }
199 __setup("nowatchdog", nowatchdog_setup);
200
nosoftlockup_setup(char * str)201 static int __init nosoftlockup_setup(char *str)
202 {
203 soft_watchdog_user_enabled = 0;
204 return 1;
205 }
206 __setup("nosoftlockup", nosoftlockup_setup);
207
208 #ifdef CONFIG_SMP
209 int __read_mostly sysctl_softlockup_all_cpu_backtrace;
210
softlockup_all_cpu_backtrace_setup(char * str)211 static int __init softlockup_all_cpu_backtrace_setup(char *str)
212 {
213 sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0);
214 return 1;
215 }
216 __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup);
217 #endif
218
219 static void __lockup_detector_cleanup(void);
220
221 /*
222 * Hard-lockup warnings should be triggered after just a few seconds. Soft-
223 * lockups can have false positives under extreme conditions. So we generally
224 * want a higher threshold for soft lockups than for hard lockups. So we couple
225 * the thresholds with a factor: we make the soft threshold twice the amount of
226 * time the hard threshold is.
227 */
get_softlockup_thresh(void)228 static int get_softlockup_thresh(void)
229 {
230 return watchdog_thresh * 2;
231 }
232
233 /*
234 * Returns seconds, approximately. We don't need nanosecond
235 * resolution, and we don't need to waste time with a big divide when
236 * 2^30ns == 1.074s.
237 */
get_timestamp(void)238 static unsigned long get_timestamp(void)
239 {
240 return running_clock() >> 30LL; /* 2^30 ~= 10^9 */
241 }
242
set_sample_period(void)243 static void set_sample_period(void)
244 {
245 /*
246 * convert watchdog_thresh from seconds to ns
247 * the divide by 5 is to give hrtimer several chances (two
248 * or three with the current relation between the soft
249 * and hard thresholds) to increment before the
250 * hardlockup detector generates a warning
251 */
252 sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5);
253 watchdog_update_hrtimer_threshold(sample_period);
254 }
255
256 /* Commands for resetting the watchdog */
__touch_watchdog(void)257 static void __touch_watchdog(void)
258 {
259 __this_cpu_write(watchdog_touch_ts, get_timestamp());
260 }
261
262 /**
263 * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls
264 *
265 * Call when the scheduler may have stalled for legitimate reasons
266 * preventing the watchdog task from executing - e.g. the scheduler
267 * entering idle state. This should only be used for scheduler events.
268 * Use touch_softlockup_watchdog() for everything else.
269 */
touch_softlockup_watchdog_sched(void)270 notrace void touch_softlockup_watchdog_sched(void)
271 {
272 /*
273 * Preemption can be enabled. It doesn't matter which CPU's timestamp
274 * gets zeroed here, so use the raw_ operation.
275 */
276 raw_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
277 }
278
touch_softlockup_watchdog(void)279 notrace void touch_softlockup_watchdog(void)
280 {
281 touch_softlockup_watchdog_sched();
282 wq_watchdog_touch(raw_smp_processor_id());
283 }
284 EXPORT_SYMBOL(touch_softlockup_watchdog);
285
touch_all_softlockup_watchdogs(void)286 void touch_all_softlockup_watchdogs(void)
287 {
288 int cpu;
289
290 /*
291 * watchdog_mutex cannpt be taken here, as this might be called
292 * from (soft)interrupt context, so the access to
293 * watchdog_allowed_cpumask might race with a concurrent update.
294 *
295 * The watchdog time stamp can race against a concurrent real
296 * update as well, the only side effect might be a cycle delay for
297 * the softlockup check.
298 */
299 for_each_cpu(cpu, &watchdog_allowed_mask)
300 per_cpu(watchdog_touch_ts, cpu) = SOFTLOCKUP_RESET;
301 wq_watchdog_touch(-1);
302 }
303
touch_softlockup_watchdog_sync(void)304 void touch_softlockup_watchdog_sync(void)
305 {
306 __this_cpu_write(softlockup_touch_sync, true);
307 __this_cpu_write(watchdog_touch_ts, SOFTLOCKUP_RESET);
308 }
309
is_softlockup(unsigned long touch_ts)310 static int is_softlockup(unsigned long touch_ts)
311 {
312 unsigned long now = get_timestamp();
313
314 if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){
315 /* Warn about unreasonable delays. */
316 if (time_after(now, touch_ts + get_softlockup_thresh()))
317 return now - touch_ts;
318 }
319 return 0;
320 }
321
322 /* watchdog detector functions */
is_hardlockup(void)323 bool is_hardlockup(void)
324 {
325 unsigned long hrint = __this_cpu_read(hrtimer_interrupts);
326
327 if (__this_cpu_read(hrtimer_interrupts_saved) == hrint)
328 return true;
329
330 __this_cpu_write(hrtimer_interrupts_saved, hrint);
331 return false;
332 }
333
watchdog_interrupt_count(void)334 static void watchdog_interrupt_count(void)
335 {
336 __this_cpu_inc(hrtimer_interrupts);
337 }
338
339 /* watchdog kicker functions */
watchdog_timer_fn(struct hrtimer * hrtimer)340 static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer)
341 {
342 unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts);
343 struct pt_regs *regs = get_irq_regs();
344 int duration;
345 int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace;
346
347 if (!watchdog_enabled)
348 return HRTIMER_NORESTART;
349
350 /* kick the hardlockup detector */
351 watchdog_interrupt_count();
352
353 /* kick the softlockup detector */
354 wake_up_process(__this_cpu_read(softlockup_watchdog));
355
356 /* .. and repeat */
357 hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period));
358
359 if (touch_ts == SOFTLOCKUP_RESET) {
360 if (unlikely(__this_cpu_read(softlockup_touch_sync))) {
361 /*
362 * If the time stamp was touched atomically
363 * make sure the scheduler tick is up to date.
364 */
365 __this_cpu_write(softlockup_touch_sync, false);
366 sched_clock_tick();
367 }
368
369 /* Clear the guest paused flag on watchdog reset */
370 kvm_check_and_clear_guest_paused();
371 __touch_watchdog();
372 return HRTIMER_RESTART;
373 }
374
375 /* check for a softlockup
376 * This is done by making sure a high priority task is
377 * being scheduled. The task touches the watchdog to
378 * indicate it is getting cpu time. If it hasn't then
379 * this is a good indication some task is hogging the cpu
380 */
381 duration = is_softlockup(touch_ts);
382 if (unlikely(duration)) {
383 /*
384 * If a virtual machine is stopped by the host it can look to
385 * the watchdog like a soft lockup, check to see if the host
386 * stopped the vm before we issue the warning
387 */
388 if (kvm_check_and_clear_guest_paused())
389 return HRTIMER_RESTART;
390
391 /* only warn once */
392 if (__this_cpu_read(soft_watchdog_warn) == true) {
393 /*
394 * When multiple processes are causing softlockups the
395 * softlockup detector only warns on the first one
396 * because the code relies on a full quiet cycle to
397 * re-arm. The second process prevents the quiet cycle
398 * and never gets reported. Use task pointers to detect
399 * this.
400 */
401 if (__this_cpu_read(softlockup_task_ptr_saved) !=
402 current) {
403 __this_cpu_write(soft_watchdog_warn, false);
404 __touch_watchdog();
405 }
406 return HRTIMER_RESTART;
407 }
408
409 if (softlockup_all_cpu_backtrace) {
410 /* Prevent multiple soft-lockup reports if one cpu is already
411 * engaged in dumping cpu back traces
412 */
413 if (test_and_set_bit(0, &soft_lockup_nmi_warn)) {
414 /* Someone else will report us. Let's give up */
415 __this_cpu_write(soft_watchdog_warn, true);
416 return HRTIMER_RESTART;
417 }
418 }
419
420 pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n",
421 smp_processor_id(), duration,
422 current->comm, task_pid_nr(current));
423 __this_cpu_write(softlockup_task_ptr_saved, current);
424 print_modules();
425 print_irqtrace_events(current);
426 if (regs)
427 show_regs(regs);
428 else
429 dump_stack();
430
431 if (softlockup_all_cpu_backtrace) {
432 /* Avoid generating two back traces for current
433 * given that one is already made above
434 */
435 trigger_allbutself_cpu_backtrace();
436
437 clear_bit(0, &soft_lockup_nmi_warn);
438 /* Barrier to sync with other cpus */
439 smp_mb__after_atomic();
440 }
441
442 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
443 if (softlockup_panic)
444 panic("softlockup: hung tasks");
445 __this_cpu_write(soft_watchdog_warn, true);
446 } else
447 __this_cpu_write(soft_watchdog_warn, false);
448
449 return HRTIMER_RESTART;
450 }
451
watchdog_set_prio(unsigned int policy,unsigned int prio)452 static void watchdog_set_prio(unsigned int policy, unsigned int prio)
453 {
454 struct sched_param param = { .sched_priority = prio };
455
456 sched_setscheduler(current, policy, ¶m);
457 }
458
watchdog_enable(unsigned int cpu)459 static void watchdog_enable(unsigned int cpu)
460 {
461 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
462
463 /*
464 * Start the timer first to prevent the NMI watchdog triggering
465 * before the timer has a chance to fire.
466 */
467 hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
468 hrtimer->function = watchdog_timer_fn;
469 hrtimer_start(hrtimer, ns_to_ktime(sample_period),
470 HRTIMER_MODE_REL_PINNED);
471
472 /* Initialize timestamp */
473 __touch_watchdog();
474 /* Enable the perf event */
475 if (watchdog_enabled & NMI_WATCHDOG_ENABLED)
476 watchdog_nmi_enable(cpu);
477
478 watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1);
479 }
480
watchdog_disable(unsigned int cpu)481 static void watchdog_disable(unsigned int cpu)
482 {
483 struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer);
484
485 watchdog_set_prio(SCHED_NORMAL, 0);
486 /*
487 * Disable the perf event first. That prevents that a large delay
488 * between disabling the timer and disabling the perf event causes
489 * the perf NMI to detect a false positive.
490 */
491 watchdog_nmi_disable(cpu);
492 hrtimer_cancel(hrtimer);
493 }
494
watchdog_cleanup(unsigned int cpu,bool online)495 static void watchdog_cleanup(unsigned int cpu, bool online)
496 {
497 watchdog_disable(cpu);
498 }
499
watchdog_should_run(unsigned int cpu)500 static int watchdog_should_run(unsigned int cpu)
501 {
502 return __this_cpu_read(hrtimer_interrupts) !=
503 __this_cpu_read(soft_lockup_hrtimer_cnt);
504 }
505
506 /*
507 * The watchdog thread function - touches the timestamp.
508 *
509 * It only runs once every sample_period seconds (4 seconds by
510 * default) to reset the softlockup timestamp. If this gets delayed
511 * for more than 2*watchdog_thresh seconds then the debug-printout
512 * triggers in watchdog_timer_fn().
513 */
watchdog(unsigned int cpu)514 static void watchdog(unsigned int cpu)
515 {
516 __this_cpu_write(soft_lockup_hrtimer_cnt,
517 __this_cpu_read(hrtimer_interrupts));
518 __touch_watchdog();
519 }
520
521 static struct smp_hotplug_thread watchdog_threads = {
522 .store = &softlockup_watchdog,
523 .thread_should_run = watchdog_should_run,
524 .thread_fn = watchdog,
525 .thread_comm = "watchdog/%u",
526 .setup = watchdog_enable,
527 .cleanup = watchdog_cleanup,
528 .park = watchdog_disable,
529 .unpark = watchdog_enable,
530 };
531
softlockup_update_smpboot_threads(void)532 static void softlockup_update_smpboot_threads(void)
533 {
534 lockdep_assert_held(&watchdog_mutex);
535
536 if (!softlockup_threads_initialized)
537 return;
538
539 smpboot_update_cpumask_percpu_thread(&watchdog_threads,
540 &watchdog_allowed_mask);
541 }
542
543 /* Temporarily park all watchdog threads */
softlockup_park_all_threads(void)544 static void softlockup_park_all_threads(void)
545 {
546 cpumask_clear(&watchdog_allowed_mask);
547 softlockup_update_smpboot_threads();
548 }
549
550 /* Unpark enabled threads */
softlockup_unpark_threads(void)551 static void softlockup_unpark_threads(void)
552 {
553 cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask);
554 softlockup_update_smpboot_threads();
555 }
556
lockup_detector_reconfigure(void)557 static void lockup_detector_reconfigure(void)
558 {
559 cpus_read_lock();
560 watchdog_nmi_stop();
561 softlockup_park_all_threads();
562 set_sample_period();
563 lockup_detector_update_enable();
564 if (watchdog_enabled && watchdog_thresh)
565 softlockup_unpark_threads();
566 watchdog_nmi_start();
567 cpus_read_unlock();
568 /*
569 * Must be called outside the cpus locked section to prevent
570 * recursive locking in the perf code.
571 */
572 __lockup_detector_cleanup();
573 }
574
575 /*
576 * Create the watchdog thread infrastructure and configure the detector(s).
577 *
578 * The threads are not unparked as watchdog_allowed_mask is empty. When
579 * the threads are sucessfully initialized, take the proper locks and
580 * unpark the threads in the watchdog_cpumask if the watchdog is enabled.
581 */
lockup_detector_setup(void)582 static __init void lockup_detector_setup(void)
583 {
584 int ret;
585
586 /*
587 * If sysctl is off and watchdog got disabled on the command line,
588 * nothing to do here.
589 */
590 lockup_detector_update_enable();
591
592 if (!IS_ENABLED(CONFIG_SYSCTL) &&
593 !(watchdog_enabled && watchdog_thresh))
594 return;
595
596 ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads,
597 &watchdog_allowed_mask);
598 if (ret) {
599 pr_err("Failed to initialize soft lockup detector threads\n");
600 return;
601 }
602
603 mutex_lock(&watchdog_mutex);
604 softlockup_threads_initialized = true;
605 lockup_detector_reconfigure();
606 mutex_unlock(&watchdog_mutex);
607 }
608
609 #else /* CONFIG_SOFTLOCKUP_DETECTOR */
watchdog_park_threads(void)610 static inline int watchdog_park_threads(void) { return 0; }
watchdog_unpark_threads(void)611 static inline void watchdog_unpark_threads(void) { }
watchdog_enable_all_cpus(void)612 static inline int watchdog_enable_all_cpus(void) { return 0; }
watchdog_disable_all_cpus(void)613 static inline void watchdog_disable_all_cpus(void) { }
lockup_detector_reconfigure(void)614 static void lockup_detector_reconfigure(void)
615 {
616 cpus_read_lock();
617 watchdog_nmi_stop();
618 lockup_detector_update_enable();
619 watchdog_nmi_start();
620 cpus_read_unlock();
621 }
lockup_detector_setup(void)622 static inline void lockup_detector_setup(void)
623 {
624 lockup_detector_reconfigure();
625 }
626 #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */
627
__lockup_detector_cleanup(void)628 static void __lockup_detector_cleanup(void)
629 {
630 lockdep_assert_held(&watchdog_mutex);
631 hardlockup_detector_perf_cleanup();
632 }
633
634 /**
635 * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes
636 *
637 * Caller must not hold the cpu hotplug rwsem.
638 */
lockup_detector_cleanup(void)639 void lockup_detector_cleanup(void)
640 {
641 mutex_lock(&watchdog_mutex);
642 __lockup_detector_cleanup();
643 mutex_unlock(&watchdog_mutex);
644 }
645
646 /**
647 * lockup_detector_soft_poweroff - Interface to stop lockup detector(s)
648 *
649 * Special interface for parisc. It prevents lockup detector warnings from
650 * the default pm_poweroff() function which busy loops forever.
651 */
lockup_detector_soft_poweroff(void)652 void lockup_detector_soft_poweroff(void)
653 {
654 watchdog_enabled = 0;
655 }
656
657 #ifdef CONFIG_SYSCTL
658
659 /* Propagate any changes to the watchdog threads */
proc_watchdog_update(void)660 static void proc_watchdog_update(void)
661 {
662 /* Remove impossible cpus to keep sysctl output clean. */
663 cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask);
664 lockup_detector_reconfigure();
665 }
666
667 /*
668 * common function for watchdog, nmi_watchdog and soft_watchdog parameter
669 *
670 * caller | table->data points to | 'which'
671 * -------------------|----------------------------|--------------------------
672 * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED |
673 * | | SOFT_WATCHDOG_ENABLED
674 * -------------------|----------------------------|--------------------------
675 * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED
676 * -------------------|----------------------------|--------------------------
677 * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED
678 */
proc_watchdog_common(int which,struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)679 static int proc_watchdog_common(int which, struct ctl_table *table, int write,
680 void __user *buffer, size_t *lenp, loff_t *ppos)
681 {
682 int err, old, *param = table->data;
683
684 mutex_lock(&watchdog_mutex);
685
686 if (!write) {
687 /*
688 * On read synchronize the userspace interface. This is a
689 * racy snapshot.
690 */
691 *param = (watchdog_enabled & which) != 0;
692 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
693 } else {
694 old = READ_ONCE(*param);
695 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
696 if (!err && old != READ_ONCE(*param))
697 proc_watchdog_update();
698 }
699 mutex_unlock(&watchdog_mutex);
700 return err;
701 }
702
703 /*
704 * /proc/sys/kernel/watchdog
705 */
proc_watchdog(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)706 int proc_watchdog(struct ctl_table *table, int write,
707 void __user *buffer, size_t *lenp, loff_t *ppos)
708 {
709 return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED,
710 table, write, buffer, lenp, ppos);
711 }
712
713 /*
714 * /proc/sys/kernel/nmi_watchdog
715 */
proc_nmi_watchdog(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)716 int proc_nmi_watchdog(struct ctl_table *table, int write,
717 void __user *buffer, size_t *lenp, loff_t *ppos)
718 {
719 if (!nmi_watchdog_available && write)
720 return -ENOTSUPP;
721 return proc_watchdog_common(NMI_WATCHDOG_ENABLED,
722 table, write, buffer, lenp, ppos);
723 }
724
725 /*
726 * /proc/sys/kernel/soft_watchdog
727 */
proc_soft_watchdog(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)728 int proc_soft_watchdog(struct ctl_table *table, int write,
729 void __user *buffer, size_t *lenp, loff_t *ppos)
730 {
731 return proc_watchdog_common(SOFT_WATCHDOG_ENABLED,
732 table, write, buffer, lenp, ppos);
733 }
734
735 /*
736 * /proc/sys/kernel/watchdog_thresh
737 */
proc_watchdog_thresh(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)738 int proc_watchdog_thresh(struct ctl_table *table, int write,
739 void __user *buffer, size_t *lenp, loff_t *ppos)
740 {
741 int err, old;
742
743 mutex_lock(&watchdog_mutex);
744
745 old = READ_ONCE(watchdog_thresh);
746 err = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
747
748 if (!err && write && old != READ_ONCE(watchdog_thresh))
749 proc_watchdog_update();
750
751 mutex_unlock(&watchdog_mutex);
752 return err;
753 }
754
755 /*
756 * The cpumask is the mask of possible cpus that the watchdog can run
757 * on, not the mask of cpus it is actually running on. This allows the
758 * user to specify a mask that will include cpus that have not yet
759 * been brought online, if desired.
760 */
proc_watchdog_cpumask(struct ctl_table * table,int write,void __user * buffer,size_t * lenp,loff_t * ppos)761 int proc_watchdog_cpumask(struct ctl_table *table, int write,
762 void __user *buffer, size_t *lenp, loff_t *ppos)
763 {
764 int err;
765
766 mutex_lock(&watchdog_mutex);
767
768 err = proc_do_large_bitmap(table, write, buffer, lenp, ppos);
769 if (!err && write)
770 proc_watchdog_update();
771
772 mutex_unlock(&watchdog_mutex);
773 return err;
774 }
775 #endif /* CONFIG_SYSCTL */
776
lockup_detector_init(void)777 void __init lockup_detector_init(void)
778 {
779 #ifdef CONFIG_NO_HZ_FULL
780 if (tick_nohz_full_enabled()) {
781 pr_info("Disabling watchdog on nohz_full cores by default\n");
782 cpumask_copy(&watchdog_cpumask, housekeeping_mask);
783 } else
784 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
785 #else
786 cpumask_copy(&watchdog_cpumask, cpu_possible_mask);
787 #endif
788
789 if (!watchdog_nmi_probe())
790 nmi_watchdog_available = true;
791 lockup_detector_setup();
792 }
793