]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Detect hard and soft lockups on a system | |
3 | * | |
4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. | |
5 | * | |
6 | * Note: Most of this code is borrowed heavily from the original softlockup | |
7 | * detector, so thanks to Ingo for the initial implementation. | |
8 | * Some chunks also taken from the old x86-specific nmi watchdog code, thanks | |
9 | * to those contributors as well. | |
10 | */ | |
11 | ||
12 | #define pr_fmt(fmt) "watchdog: " fmt | |
13 | ||
14 | #include <linux/mm.h> | |
15 | #include <linux/cpu.h> | |
16 | #include <linux/nmi.h> | |
17 | #include <linux/init.h> | |
18 | #include <linux/module.h> | |
19 | #include <linux/sysctl.h> | |
20 | #include <linux/smpboot.h> | |
21 | #include <linux/sched/rt.h> | |
22 | #include <uapi/linux/sched/types.h> | |
23 | #include <linux/tick.h> | |
24 | #include <linux/workqueue.h> | |
25 | #include <linux/sched/clock.h> | |
26 | #include <linux/sched/debug.h> | |
27 | ||
28 | #include <asm/irq_regs.h> | |
29 | #include <linux/kvm_para.h> | |
30 | #include <linux/kthread.h> | |
31 | ||
32 | static DEFINE_MUTEX(watchdog_mutex); | |
33 | ||
34 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) | |
35 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
36 | NMI_WATCHDOG_ENABLED; | |
37 | #else | |
38 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; | |
39 | #endif | |
40 | ||
41 | int __read_mostly nmi_watchdog_user_enabled; | |
42 | int __read_mostly soft_watchdog_user_enabled; | |
43 | int __read_mostly watchdog_user_enabled; | |
44 | int __read_mostly watchdog_thresh = 10; | |
45 | ||
46 | struct cpumask watchdog_allowed_mask __read_mostly; | |
47 | static bool softlockup_threads_initialized __read_mostly; | |
48 | ||
49 | struct cpumask watchdog_cpumask __read_mostly; | |
50 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
51 | ||
52 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
53 | /* | |
54 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
55 | */ | |
56 | unsigned int __read_mostly hardlockup_panic = | |
57 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
58 | /* | |
59 | * We may not want to enable hard lockup detection by default in all cases, | |
60 | * for example when running the kernel as a guest on a hypervisor. In these | |
61 | * cases this function can be called to disable hard lockup detection. This | |
62 | * function should only be executed once by the boot processor before the | |
63 | * kernel command line parameters are parsed, because otherwise it is not | |
64 | * possible to override this in hardlockup_panic_setup(). | |
65 | */ | |
66 | void __init hardlockup_detector_disable(void) | |
67 | { | |
68 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
69 | } | |
70 | ||
71 | static int __init hardlockup_panic_setup(char *str) | |
72 | { | |
73 | if (!strncmp(str, "panic", 5)) | |
74 | hardlockup_panic = 1; | |
75 | else if (!strncmp(str, "nopanic", 7)) | |
76 | hardlockup_panic = 0; | |
77 | else if (!strncmp(str, "0", 1)) | |
78 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
79 | else if (!strncmp(str, "1", 1)) | |
80 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
81 | return 1; | |
82 | } | |
83 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
84 | ||
85 | # ifdef CONFIG_SMP | |
86 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
87 | ||
88 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) | |
89 | { | |
90 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
91 | return 1; | |
92 | } | |
93 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
94 | # endif /* CONFIG_SMP */ | |
95 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
96 | ||
97 | /* | |
98 | * These functions can be overridden if an architecture implements its | |
99 | * own hardlockup detector. | |
100 | * | |
101 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
102 | * softlockup watchdog threads start and stop. The arch must select the | |
103 | * SOFTLOCKUP_DETECTOR Kconfig. | |
104 | */ | |
105 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
106 | { | |
107 | return 0; | |
108 | } | |
109 | ||
110 | void __weak watchdog_nmi_disable(unsigned int cpu) | |
111 | { | |
112 | hardlockup_detector_perf_disable(); | |
113 | } | |
114 | ||
115 | /** | |
116 | * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs | |
117 | * @run: If false stop the watchdogs on all enabled CPUs | |
118 | * If true start the watchdogs on all enabled CPUs | |
119 | * | |
120 | * The core call order is: | |
121 | * watchdog_nmi_reconfigure(false); | |
122 | * update_variables(); | |
123 | * watchdog_nmi_reconfigure(true); | |
124 | * | |
125 | * The second call which starts the watchdogs again guarantees that the | |
126 | * following variables are stable across the call. | |
127 | * - watchdog_enabled | |
128 | * - watchdog_thresh | |
129 | * - watchdog_cpumask | |
130 | * | |
131 | * After the call the variables can be changed again. | |
132 | */ | |
133 | void __weak watchdog_nmi_reconfigure(bool run) { } | |
134 | ||
135 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR | |
136 | ||
137 | /* Global variables, exported for sysctl */ | |
138 | unsigned int __read_mostly softlockup_panic = | |
139 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
140 | ||
141 | static u64 __read_mostly sample_period; | |
142 | ||
143 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
144 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
145 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
146 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
147 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); | |
148 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); | |
149 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
150 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); | |
151 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); | |
152 | static unsigned long soft_lockup_nmi_warn; | |
153 | ||
154 | static int __init softlockup_panic_setup(char *str) | |
155 | { | |
156 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
157 | return 1; | |
158 | } | |
159 | __setup("softlockup_panic=", softlockup_panic_setup); | |
160 | ||
161 | static int __init nowatchdog_setup(char *str) | |
162 | { | |
163 | watchdog_enabled = 0; | |
164 | return 1; | |
165 | } | |
166 | __setup("nowatchdog", nowatchdog_setup); | |
167 | ||
168 | static int __init nosoftlockup_setup(char *str) | |
169 | { | |
170 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; | |
171 | return 1; | |
172 | } | |
173 | __setup("nosoftlockup", nosoftlockup_setup); | |
174 | ||
175 | #ifdef CONFIG_SMP | |
176 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
177 | ||
178 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
179 | { | |
180 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
181 | return 1; | |
182 | } | |
183 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
184 | #endif | |
185 | ||
186 | static void __lockup_detector_cleanup(void); | |
187 | ||
188 | /* | |
189 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
190 | * lockups can have false positives under extreme conditions. So we generally | |
191 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
192 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
193 | * time the hard threshold is. | |
194 | */ | |
195 | static int get_softlockup_thresh(void) | |
196 | { | |
197 | return watchdog_thresh * 2; | |
198 | } | |
199 | ||
200 | /* | |
201 | * Returns seconds, approximately. We don't need nanosecond | |
202 | * resolution, and we don't need to waste time with a big divide when | |
203 | * 2^30ns == 1.074s. | |
204 | */ | |
205 | static unsigned long get_timestamp(void) | |
206 | { | |
207 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ | |
208 | } | |
209 | ||
210 | static void set_sample_period(void) | |
211 | { | |
212 | /* | |
213 | * convert watchdog_thresh from seconds to ns | |
214 | * the divide by 5 is to give hrtimer several chances (two | |
215 | * or three with the current relation between the soft | |
216 | * and hard thresholds) to increment before the | |
217 | * hardlockup detector generates a warning | |
218 | */ | |
219 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); | |
220 | watchdog_update_hrtimer_threshold(sample_period); | |
221 | } | |
222 | ||
223 | /* Commands for resetting the watchdog */ | |
224 | static void __touch_watchdog(void) | |
225 | { | |
226 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); | |
227 | } | |
228 | ||
229 | /** | |
230 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
231 | * | |
232 | * Call when the scheduler may have stalled for legitimate reasons | |
233 | * preventing the watchdog task from executing - e.g. the scheduler | |
234 | * entering idle state. This should only be used for scheduler events. | |
235 | * Use touch_softlockup_watchdog() for everything else. | |
236 | */ | |
237 | void touch_softlockup_watchdog_sched(void) | |
238 | { | |
239 | /* | |
240 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
241 | * gets zeroed here, so use the raw_ operation. | |
242 | */ | |
243 | raw_cpu_write(watchdog_touch_ts, 0); | |
244 | } | |
245 | ||
246 | void touch_softlockup_watchdog(void) | |
247 | { | |
248 | touch_softlockup_watchdog_sched(); | |
249 | wq_watchdog_touch(raw_smp_processor_id()); | |
250 | } | |
251 | EXPORT_SYMBOL(touch_softlockup_watchdog); | |
252 | ||
253 | void touch_all_softlockup_watchdogs(void) | |
254 | { | |
255 | int cpu; | |
256 | ||
257 | /* | |
258 | * watchdog_mutex cannpt be taken here, as this might be called | |
259 | * from (soft)interrupt context, so the access to | |
260 | * watchdog_allowed_cpumask might race with a concurrent update. | |
261 | * | |
262 | * The watchdog time stamp can race against a concurrent real | |
263 | * update as well, the only side effect might be a cycle delay for | |
264 | * the softlockup check. | |
265 | */ | |
266 | for_each_cpu(cpu, &watchdog_allowed_mask) | |
267 | per_cpu(watchdog_touch_ts, cpu) = 0; | |
268 | wq_watchdog_touch(-1); | |
269 | } | |
270 | ||
271 | void touch_softlockup_watchdog_sync(void) | |
272 | { | |
273 | __this_cpu_write(softlockup_touch_sync, true); | |
274 | __this_cpu_write(watchdog_touch_ts, 0); | |
275 | } | |
276 | ||
277 | static int is_softlockup(unsigned long touch_ts) | |
278 | { | |
279 | unsigned long now = get_timestamp(); | |
280 | ||
281 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ | |
282 | /* Warn about unreasonable delays. */ | |
283 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
284 | return now - touch_ts; | |
285 | } | |
286 | return 0; | |
287 | } | |
288 | ||
289 | /* watchdog detector functions */ | |
290 | bool is_hardlockup(void) | |
291 | { | |
292 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); | |
293 | ||
294 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) | |
295 | return true; | |
296 | ||
297 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
298 | return false; | |
299 | } | |
300 | ||
301 | static void watchdog_interrupt_count(void) | |
302 | { | |
303 | __this_cpu_inc(hrtimer_interrupts); | |
304 | } | |
305 | ||
306 | /* watchdog kicker functions */ | |
307 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
308 | { | |
309 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); | |
310 | struct pt_regs *regs = get_irq_regs(); | |
311 | int duration; | |
312 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; | |
313 | ||
314 | if (!watchdog_enabled) | |
315 | return HRTIMER_NORESTART; | |
316 | ||
317 | /* kick the hardlockup detector */ | |
318 | watchdog_interrupt_count(); | |
319 | ||
320 | /* kick the softlockup detector */ | |
321 | wake_up_process(__this_cpu_read(softlockup_watchdog)); | |
322 | ||
323 | /* .. and repeat */ | |
324 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); | |
325 | ||
326 | if (touch_ts == 0) { | |
327 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { | |
328 | /* | |
329 | * If the time stamp was touched atomically | |
330 | * make sure the scheduler tick is up to date. | |
331 | */ | |
332 | __this_cpu_write(softlockup_touch_sync, false); | |
333 | sched_clock_tick(); | |
334 | } | |
335 | ||
336 | /* Clear the guest paused flag on watchdog reset */ | |
337 | kvm_check_and_clear_guest_paused(); | |
338 | __touch_watchdog(); | |
339 | return HRTIMER_RESTART; | |
340 | } | |
341 | ||
342 | /* check for a softlockup | |
343 | * This is done by making sure a high priority task is | |
344 | * being scheduled. The task touches the watchdog to | |
345 | * indicate it is getting cpu time. If it hasn't then | |
346 | * this is a good indication some task is hogging the cpu | |
347 | */ | |
348 | duration = is_softlockup(touch_ts); | |
349 | if (unlikely(duration)) { | |
350 | /* | |
351 | * If a virtual machine is stopped by the host it can look to | |
352 | * the watchdog like a soft lockup, check to see if the host | |
353 | * stopped the vm before we issue the warning | |
354 | */ | |
355 | if (kvm_check_and_clear_guest_paused()) | |
356 | return HRTIMER_RESTART; | |
357 | ||
358 | /* only warn once */ | |
359 | if (__this_cpu_read(soft_watchdog_warn) == true) { | |
360 | /* | |
361 | * When multiple processes are causing softlockups the | |
362 | * softlockup detector only warns on the first one | |
363 | * because the code relies on a full quiet cycle to | |
364 | * re-arm. The second process prevents the quiet cycle | |
365 | * and never gets reported. Use task pointers to detect | |
366 | * this. | |
367 | */ | |
368 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
369 | current) { | |
370 | __this_cpu_write(soft_watchdog_warn, false); | |
371 | __touch_watchdog(); | |
372 | } | |
373 | return HRTIMER_RESTART; | |
374 | } | |
375 | ||
376 | if (softlockup_all_cpu_backtrace) { | |
377 | /* Prevent multiple soft-lockup reports if one cpu is already | |
378 | * engaged in dumping cpu back traces | |
379 | */ | |
380 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
381 | /* Someone else will report us. Let's give up */ | |
382 | __this_cpu_write(soft_watchdog_warn, true); | |
383 | return HRTIMER_RESTART; | |
384 | } | |
385 | } | |
386 | ||
387 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", | |
388 | smp_processor_id(), duration, | |
389 | current->comm, task_pid_nr(current)); | |
390 | __this_cpu_write(softlockup_task_ptr_saved, current); | |
391 | print_modules(); | |
392 | print_irqtrace_events(current); | |
393 | if (regs) | |
394 | show_regs(regs); | |
395 | else | |
396 | dump_stack(); | |
397 | ||
398 | if (softlockup_all_cpu_backtrace) { | |
399 | /* Avoid generating two back traces for current | |
400 | * given that one is already made above | |
401 | */ | |
402 | trigger_allbutself_cpu_backtrace(); | |
403 | ||
404 | clear_bit(0, &soft_lockup_nmi_warn); | |
405 | /* Barrier to sync with other cpus */ | |
406 | smp_mb__after_atomic(); | |
407 | } | |
408 | ||
409 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); | |
410 | if (softlockup_panic) | |
411 | panic("softlockup: hung tasks"); | |
412 | __this_cpu_write(soft_watchdog_warn, true); | |
413 | } else | |
414 | __this_cpu_write(soft_watchdog_warn, false); | |
415 | ||
416 | return HRTIMER_RESTART; | |
417 | } | |
418 | ||
419 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) | |
420 | { | |
421 | struct sched_param param = { .sched_priority = prio }; | |
422 | ||
423 | sched_setscheduler(current, policy, ¶m); | |
424 | } | |
425 | ||
426 | static void watchdog_enable(unsigned int cpu) | |
427 | { | |
428 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); | |
429 | ||
430 | /* | |
431 | * Start the timer first to prevent the NMI watchdog triggering | |
432 | * before the timer has a chance to fire. | |
433 | */ | |
434 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
435 | hrtimer->function = watchdog_timer_fn; | |
436 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), | |
437 | HRTIMER_MODE_REL_PINNED); | |
438 | ||
439 | /* Initialize timestamp */ | |
440 | __touch_watchdog(); | |
441 | /* Enable the perf event */ | |
442 | watchdog_nmi_enable(cpu); | |
443 | ||
444 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
445 | } | |
446 | ||
447 | static void watchdog_disable(unsigned int cpu) | |
448 | { | |
449 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); | |
450 | ||
451 | watchdog_set_prio(SCHED_NORMAL, 0); | |
452 | /* | |
453 | * Disable the perf event first. That prevents that a large delay | |
454 | * between disabling the timer and disabling the perf event causes | |
455 | * the perf NMI to detect a false positive. | |
456 | */ | |
457 | watchdog_nmi_disable(cpu); | |
458 | hrtimer_cancel(hrtimer); | |
459 | } | |
460 | ||
461 | static void watchdog_cleanup(unsigned int cpu, bool online) | |
462 | { | |
463 | watchdog_disable(cpu); | |
464 | } | |
465 | ||
466 | static int watchdog_should_run(unsigned int cpu) | |
467 | { | |
468 | return __this_cpu_read(hrtimer_interrupts) != | |
469 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
470 | } | |
471 | ||
472 | /* | |
473 | * The watchdog thread function - touches the timestamp. | |
474 | * | |
475 | * It only runs once every sample_period seconds (4 seconds by | |
476 | * default) to reset the softlockup timestamp. If this gets delayed | |
477 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
478 | * triggers in watchdog_timer_fn(). | |
479 | */ | |
480 | static void watchdog(unsigned int cpu) | |
481 | { | |
482 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
483 | __this_cpu_read(hrtimer_interrupts)); | |
484 | __touch_watchdog(); | |
485 | } | |
486 | ||
487 | static struct smp_hotplug_thread watchdog_threads = { | |
488 | .store = &softlockup_watchdog, | |
489 | .thread_should_run = watchdog_should_run, | |
490 | .thread_fn = watchdog, | |
491 | .thread_comm = "watchdog/%u", | |
492 | .setup = watchdog_enable, | |
493 | .cleanup = watchdog_cleanup, | |
494 | .park = watchdog_disable, | |
495 | .unpark = watchdog_enable, | |
496 | }; | |
497 | ||
498 | static void softlockup_update_smpboot_threads(void) | |
499 | { | |
500 | lockdep_assert_held(&watchdog_mutex); | |
501 | ||
502 | if (!softlockup_threads_initialized) | |
503 | return; | |
504 | ||
505 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
506 | &watchdog_allowed_mask); | |
507 | __lockup_detector_cleanup(); | |
508 | } | |
509 | ||
510 | /* Temporarily park all watchdog threads */ | |
511 | static void softlockup_park_all_threads(void) | |
512 | { | |
513 | cpumask_clear(&watchdog_allowed_mask); | |
514 | softlockup_update_smpboot_threads(); | |
515 | } | |
516 | ||
517 | /* Unpark enabled threads */ | |
518 | static void softlockup_unpark_threads(void) | |
519 | { | |
520 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
521 | softlockup_update_smpboot_threads(); | |
522 | } | |
523 | ||
524 | static void softlockup_reconfigure_threads(bool enabled) | |
525 | { | |
526 | watchdog_nmi_reconfigure(false); | |
527 | softlockup_park_all_threads(); | |
528 | set_sample_period(); | |
529 | if (enabled) | |
530 | softlockup_unpark_threads(); | |
531 | watchdog_nmi_reconfigure(true); | |
532 | } | |
533 | ||
534 | /* | |
535 | * Create the watchdog thread infrastructure. | |
536 | * | |
537 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
538 | * the threads are sucessfully initialized, take the proper locks and | |
539 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
540 | */ | |
541 | static __init void softlockup_init_threads(void) | |
542 | { | |
543 | int ret; | |
544 | ||
545 | /* | |
546 | * If sysctl is off and watchdog got disabled on the command line, | |
547 | * nothing to do here. | |
548 | */ | |
549 | if (!IS_ENABLED(CONFIG_SYSCTL) && | |
550 | !(watchdog_enabled && watchdog_thresh)) | |
551 | return; | |
552 | ||
553 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
554 | &watchdog_allowed_mask); | |
555 | if (ret) { | |
556 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
557 | return; | |
558 | } | |
559 | ||
560 | mutex_lock(&watchdog_mutex); | |
561 | softlockup_threads_initialized = true; | |
562 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); | |
563 | mutex_unlock(&watchdog_mutex); | |
564 | } | |
565 | ||
566 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ | |
567 | static inline int watchdog_park_threads(void) { return 0; } | |
568 | static inline void watchdog_unpark_threads(void) { } | |
569 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
570 | static inline void watchdog_disable_all_cpus(void) { } | |
571 | static inline void softlockup_init_threads(void) { } | |
572 | static void softlockup_reconfigure_threads(bool enabled) | |
573 | { | |
574 | watchdog_nmi_reconfigure(false); | |
575 | watchdog_nmi_reconfigure(true); | |
576 | } | |
577 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ | |
578 | ||
579 | static void __lockup_detector_cleanup(void) | |
580 | { | |
581 | lockdep_assert_held(&watchdog_mutex); | |
582 | hardlockup_detector_perf_cleanup(); | |
583 | } | |
584 | ||
585 | /** | |
586 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
587 | * | |
588 | * Caller must not hold the cpu hotplug rwsem. | |
589 | */ | |
590 | void lockup_detector_cleanup(void) | |
591 | { | |
592 | mutex_lock(&watchdog_mutex); | |
593 | __lockup_detector_cleanup(); | |
594 | mutex_unlock(&watchdog_mutex); | |
595 | } | |
596 | ||
597 | /** | |
598 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
599 | * | |
600 | * Special interface for parisc. It prevents lockup detector warnings from | |
601 | * the default pm_poweroff() function which busy loops forever. | |
602 | */ | |
603 | void lockup_detector_soft_poweroff(void) | |
604 | { | |
605 | watchdog_enabled = 0; | |
606 | } | |
607 | ||
608 | #ifdef CONFIG_SYSCTL | |
609 | ||
610 | /* Propagate any changes to the watchdog threads */ | |
611 | static void proc_watchdog_update(void) | |
612 | { | |
613 | /* Remove impossible cpus to keep sysctl output clean. */ | |
614 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
615 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); | |
616 | } | |
617 | ||
618 | /* | |
619 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
620 | * | |
621 | * caller | table->data points to | 'which' | |
622 | * -------------------|----------------------------|-------------------------- | |
623 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED | | |
624 | * | | SOFT_WATCHDOG_ENABLED | |
625 | * -------------------|----------------------------|-------------------------- | |
626 | * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED | |
627 | * -------------------|----------------------------|-------------------------- | |
628 | * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED | |
629 | */ | |
630 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
631 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
632 | { | |
633 | int err, old, new; | |
634 | int *watchdog_param = (int *)table->data; | |
635 | ||
636 | cpu_hotplug_disable(); | |
637 | mutex_lock(&watchdog_mutex); | |
638 | ||
639 | /* | |
640 | * If the parameter is being read return the state of the corresponding | |
641 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
642 | * run state of the lockup detectors. | |
643 | */ | |
644 | if (!write) { | |
645 | *watchdog_param = (watchdog_enabled & which) != 0; | |
646 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
647 | } else { | |
648 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
649 | if (err) | |
650 | goto out; | |
651 | ||
652 | /* | |
653 | * There is a race window between fetching the current value | |
654 | * from 'watchdog_enabled' and storing the new value. During | |
655 | * this race window, watchdog_nmi_enable() can sneak in and | |
656 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
657 | * The 'cmpxchg' detects this race and the loop retries. | |
658 | */ | |
659 | do { | |
660 | old = watchdog_enabled; | |
661 | /* | |
662 | * If the parameter value is not zero set the | |
663 | * corresponding bit(s), else clear it(them). | |
664 | */ | |
665 | if (*watchdog_param) | |
666 | new = old | which; | |
667 | else | |
668 | new = old & ~which; | |
669 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
670 | ||
671 | if (old != new) | |
672 | proc_watchdog_update(); | |
673 | } | |
674 | out: | |
675 | mutex_unlock(&watchdog_mutex); | |
676 | cpu_hotplug_enable(); | |
677 | return err; | |
678 | } | |
679 | ||
680 | /* | |
681 | * /proc/sys/kernel/watchdog | |
682 | */ | |
683 | int proc_watchdog(struct ctl_table *table, int write, | |
684 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
685 | { | |
686 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
687 | table, write, buffer, lenp, ppos); | |
688 | } | |
689 | ||
690 | /* | |
691 | * /proc/sys/kernel/nmi_watchdog | |
692 | */ | |
693 | int proc_nmi_watchdog(struct ctl_table *table, int write, | |
694 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
695 | { | |
696 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
697 | table, write, buffer, lenp, ppos); | |
698 | } | |
699 | ||
700 | /* | |
701 | * /proc/sys/kernel/soft_watchdog | |
702 | */ | |
703 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
704 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
705 | { | |
706 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
707 | table, write, buffer, lenp, ppos); | |
708 | } | |
709 | ||
710 | /* | |
711 | * /proc/sys/kernel/watchdog_thresh | |
712 | */ | |
713 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
714 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
715 | { | |
716 | int err, old; | |
717 | ||
718 | cpu_hotplug_disable(); | |
719 | mutex_lock(&watchdog_mutex); | |
720 | ||
721 | old = READ_ONCE(watchdog_thresh); | |
722 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
723 | ||
724 | if (!err && write && old != READ_ONCE(watchdog_thresh)) | |
725 | proc_watchdog_update(); | |
726 | ||
727 | mutex_unlock(&watchdog_mutex); | |
728 | cpu_hotplug_enable(); | |
729 | return err; | |
730 | } | |
731 | ||
732 | /* | |
733 | * The cpumask is the mask of possible cpus that the watchdog can run | |
734 | * on, not the mask of cpus it is actually running on. This allows the | |
735 | * user to specify a mask that will include cpus that have not yet | |
736 | * been brought online, if desired. | |
737 | */ | |
738 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
739 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
740 | { | |
741 | int err; | |
742 | ||
743 | cpu_hotplug_disable(); | |
744 | mutex_lock(&watchdog_mutex); | |
745 | ||
746 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); | |
747 | if (!err && write) | |
748 | proc_watchdog_update(); | |
749 | ||
750 | mutex_unlock(&watchdog_mutex); | |
751 | cpu_hotplug_enable(); | |
752 | return err; | |
753 | } | |
754 | #endif /* CONFIG_SYSCTL */ | |
755 | ||
756 | void __init lockup_detector_init(void) | |
757 | { | |
758 | #ifdef CONFIG_NO_HZ_FULL | |
759 | if (tick_nohz_full_enabled()) { | |
760 | pr_info("Disabling watchdog on nohz_full cores by default\n"); | |
761 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
762 | } else | |
763 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
764 | #else | |
765 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
766 | #endif | |
767 | ||
768 | softlockup_init_threads(); | |
769 | } |