]>
Commit | Line | Data |
---|---|---|
58687acb DZ |
1 | /* |
2 | * Detect hard and soft lockups on a system | |
3 | * | |
4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. | |
5 | * | |
86f5e6a7 FLVC |
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 | |
58687acb DZ |
9 | * to those contributors as well. |
10 | */ | |
11 | ||
5f92a7b0 | 12 | #define pr_fmt(fmt) "watchdog: " fmt |
4501980a | 13 | |
58687acb DZ |
14 | #include <linux/mm.h> |
15 | #include <linux/cpu.h> | |
16 | #include <linux/nmi.h> | |
17 | #include <linux/init.h> | |
58687acb DZ |
18 | #include <linux/module.h> |
19 | #include <linux/sysctl.h> | |
bcd951cf | 20 | #include <linux/smpboot.h> |
8bd75c77 | 21 | #include <linux/sched/rt.h> |
ae7e81c0 | 22 | #include <uapi/linux/sched/types.h> |
fe4ba3c3 | 23 | #include <linux/tick.h> |
82607adc | 24 | #include <linux/workqueue.h> |
e6017571 | 25 | #include <linux/sched/clock.h> |
b17b0153 | 26 | #include <linux/sched/debug.h> |
58687acb DZ |
27 | |
28 | #include <asm/irq_regs.h> | |
5d1c0f4a | 29 | #include <linux/kvm_para.h> |
81a4beef | 30 | #include <linux/kthread.h> |
58687acb | 31 | |
05a4a952 | 32 | /* Watchdog configuration */ |
ab992dc3 PZ |
33 | static DEFINE_MUTEX(watchdog_proc_mutex); |
34 | ||
05a4a952 NP |
35 | int __read_mostly nmi_watchdog_enabled; |
36 | ||
37 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) | |
38 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
39 | NMI_WATCHDOG_ENABLED; | |
84d56e66 | 40 | #else |
249e52e3 | 41 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
84d56e66 | 42 | #endif |
05a4a952 NP |
43 | |
44 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
45 | /* boot commands */ | |
46 | /* | |
47 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
48 | */ | |
49 | unsigned int __read_mostly hardlockup_panic = | |
50 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
51 | /* | |
52 | * We may not want to enable hard lockup detection by default in all cases, | |
53 | * for example when running the kernel as a guest on a hypervisor. In these | |
54 | * cases this function can be called to disable hard lockup detection. This | |
55 | * function should only be executed once by the boot processor before the | |
56 | * kernel command line parameters are parsed, because otherwise it is not | |
57 | * possible to override this in hardlockup_panic_setup(). | |
58 | */ | |
59 | void hardlockup_detector_disable(void) | |
60 | { | |
61 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
62 | } | |
63 | ||
64 | static int __init hardlockup_panic_setup(char *str) | |
65 | { | |
66 | if (!strncmp(str, "panic", 5)) | |
67 | hardlockup_panic = 1; | |
68 | else if (!strncmp(str, "nopanic", 7)) | |
69 | hardlockup_panic = 0; | |
70 | else if (!strncmp(str, "0", 1)) | |
71 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
72 | else if (!strncmp(str, "1", 1)) | |
73 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
74 | return 1; | |
75 | } | |
76 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
77 | ||
78 | #endif | |
79 | ||
80 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR | |
84d56e66 | 81 | int __read_mostly soft_watchdog_enabled; |
05a4a952 NP |
82 | #endif |
83 | ||
84d56e66 | 84 | int __read_mostly watchdog_user_enabled; |
4eec42f3 | 85 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 86 | |
ed235875 AT |
87 | #ifdef CONFIG_SMP |
88 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
55537871 | 89 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; |
ed235875 | 90 | #endif |
05a4a952 | 91 | struct cpumask watchdog_cpumask __read_mostly; |
fe4ba3c3 CM |
92 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); |
93 | ||
ec6a9066 UO |
94 | /* |
95 | * The 'watchdog_running' variable is set to 1 when the watchdog threads | |
96 | * are registered/started and is set to 0 when the watchdog threads are | |
97 | * unregistered/stopped, so it is an indicator whether the threads exist. | |
98 | */ | |
3c00ea82 | 99 | static int __read_mostly watchdog_running; |
ec6a9066 UO |
100 | /* |
101 | * If a subsystem has a need to deactivate the watchdog temporarily, it | |
102 | * can use the suspend/resume interface to achieve this. The content of | |
103 | * the 'watchdog_suspended' variable reflects this state. Existing threads | |
104 | * are parked/unparked by the lockup_detector_{suspend|resume} functions | |
105 | * (see comment blocks pertaining to those functions for further details). | |
106 | * | |
107 | * 'watchdog_suspended' also prevents threads from being registered/started | |
108 | * or unregistered/stopped via parameters in /proc/sys/kernel, so the state | |
109 | * of 'watchdog_running' cannot change while the watchdog is deactivated | |
110 | * temporarily (see related code in 'proc' handlers). | |
111 | */ | |
05a4a952 NP |
112 | int __read_mostly watchdog_suspended; |
113 | ||
114 | /* | |
115 | * These functions can be overridden if an architecture implements its | |
116 | * own hardlockup detector. | |
a10a842f NP |
117 | * |
118 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
119 | * softlockup watchdog threads start and stop. The arch must select the | |
120 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
121 | */ |
122 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
123 | { | |
124 | return 0; | |
125 | } | |
126 | void __weak watchdog_nmi_disable(unsigned int cpu) | |
127 | { | |
128 | } | |
129 | ||
a10a842f NP |
130 | /* |
131 | * watchdog_nmi_reconfigure can be implemented to be notified after any | |
132 | * watchdog configuration change. The arch hardlockup watchdog should | |
133 | * respond to the following variables: | |
134 | * - nmi_watchdog_enabled | |
135 | * - watchdog_thresh | |
136 | * - watchdog_cpumask | |
137 | * - sysctl_hardlockup_all_cpu_backtrace | |
138 | * - hardlockup_panic | |
139 | * - watchdog_suspended | |
140 | */ | |
141 | void __weak watchdog_nmi_reconfigure(void) | |
142 | { | |
143 | } | |
144 | ||
145 | ||
05a4a952 NP |
146 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
147 | ||
148 | /* Helper for online, unparked cpus. */ | |
149 | #define for_each_watchdog_cpu(cpu) \ | |
150 | for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask) | |
151 | ||
152 | atomic_t watchdog_park_in_progress = ATOMIC_INIT(0); | |
ec6a9066 | 153 | |
0f34c400 | 154 | static u64 __read_mostly sample_period; |
58687acb DZ |
155 | |
156 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
157 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
158 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
159 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 160 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
161 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
162 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 163 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 164 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 165 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 166 | |
58687acb DZ |
167 | unsigned int __read_mostly softlockup_panic = |
168 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
169 | ||
170 | static int __init softlockup_panic_setup(char *str) | |
171 | { | |
172 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
173 | ||
174 | return 1; | |
175 | } | |
176 | __setup("softlockup_panic=", softlockup_panic_setup); | |
177 | ||
178 | static int __init nowatchdog_setup(char *str) | |
179 | { | |
195daf66 | 180 | watchdog_enabled = 0; |
58687acb DZ |
181 | return 1; |
182 | } | |
183 | __setup("nowatchdog", nowatchdog_setup); | |
184 | ||
58687acb DZ |
185 | static int __init nosoftlockup_setup(char *str) |
186 | { | |
195daf66 | 187 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
188 | return 1; |
189 | } | |
190 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 191 | |
ed235875 AT |
192 | #ifdef CONFIG_SMP |
193 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
194 | { | |
195 | sysctl_softlockup_all_cpu_backtrace = | |
196 | !!simple_strtol(str, NULL, 0); | |
197 | return 1; | |
198 | } | |
199 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 200 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
55537871 JK |
201 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
202 | { | |
203 | sysctl_hardlockup_all_cpu_backtrace = | |
204 | !!simple_strtol(str, NULL, 0); | |
205 | return 1; | |
206 | } | |
207 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
ed235875 | 208 | #endif |
05a4a952 | 209 | #endif |
58687acb | 210 | |
4eec42f3 MSB |
211 | /* |
212 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
213 | * lockups can have false positives under extreme conditions. So we generally | |
214 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
215 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
216 | * time the hard threshold is. | |
217 | */ | |
6e9101ae | 218 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
219 | { |
220 | return watchdog_thresh * 2; | |
221 | } | |
58687acb DZ |
222 | |
223 | /* | |
224 | * Returns seconds, approximately. We don't need nanosecond | |
225 | * resolution, and we don't need to waste time with a big divide when | |
226 | * 2^30ns == 1.074s. | |
227 | */ | |
c06b4f19 | 228 | static unsigned long get_timestamp(void) |
58687acb | 229 | { |
545a2bf7 | 230 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
231 | } |
232 | ||
0f34c400 | 233 | static void set_sample_period(void) |
58687acb DZ |
234 | { |
235 | /* | |
586692a5 | 236 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
237 | * the divide by 5 is to give hrtimer several chances (two |
238 | * or three with the current relation between the soft | |
239 | * and hard thresholds) to increment before the | |
240 | * hardlockup detector generates a warning | |
58687acb | 241 | */ |
0f34c400 | 242 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 243 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
244 | } |
245 | ||
246 | /* Commands for resetting the watchdog */ | |
247 | static void __touch_watchdog(void) | |
248 | { | |
c06b4f19 | 249 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
250 | } |
251 | ||
03e0d461 TH |
252 | /** |
253 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
254 | * | |
255 | * Call when the scheduler may have stalled for legitimate reasons | |
256 | * preventing the watchdog task from executing - e.g. the scheduler | |
257 | * entering idle state. This should only be used for scheduler events. | |
258 | * Use touch_softlockup_watchdog() for everything else. | |
259 | */ | |
260 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 261 | { |
7861144b AM |
262 | /* |
263 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
264 | * gets zeroed here, so use the raw_ operation. | |
265 | */ | |
266 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 267 | } |
03e0d461 TH |
268 | |
269 | void touch_softlockup_watchdog(void) | |
270 | { | |
271 | touch_softlockup_watchdog_sched(); | |
82607adc | 272 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 273 | } |
0167c781 | 274 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 275 | |
332fbdbc | 276 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
277 | { |
278 | int cpu; | |
279 | ||
280 | /* | |
281 | * this is done lockless | |
282 | * do we care if a 0 races with a timestamp? | |
283 | * all it means is the softlock check starts one cycle later | |
284 | */ | |
fe4ba3c3 | 285 | for_each_watchdog_cpu(cpu) |
58687acb | 286 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 287 | wq_watchdog_touch(-1); |
58687acb DZ |
288 | } |
289 | ||
58687acb DZ |
290 | void touch_softlockup_watchdog_sync(void) |
291 | { | |
f7f66b05 CL |
292 | __this_cpu_write(softlockup_touch_sync, true); |
293 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
294 | } |
295 | ||
26e09c6e | 296 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 297 | { |
c06b4f19 | 298 | unsigned long now = get_timestamp(); |
58687acb | 299 | |
39d2da21 | 300 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
301 | /* Warn about unreasonable delays. */ |
302 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
303 | return now - touch_ts; | |
304 | } | |
58687acb DZ |
305 | return 0; |
306 | } | |
307 | ||
05a4a952 NP |
308 | /* watchdog detector functions */ |
309 | bool is_hardlockup(void) | |
58687acb | 310 | { |
05a4a952 | 311 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 312 | |
05a4a952 NP |
313 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
314 | return true; | |
315 | ||
316 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
317 | return false; | |
73ce0511 | 318 | } |
05a4a952 NP |
319 | |
320 | static void watchdog_interrupt_count(void) | |
73ce0511 | 321 | { |
05a4a952 | 322 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 323 | } |
58687acb | 324 | |
58cf690a UO |
325 | static int watchdog_enable_all_cpus(void); |
326 | static void watchdog_disable_all_cpus(void); | |
327 | ||
58687acb DZ |
328 | /* watchdog kicker functions */ |
329 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
330 | { | |
909ea964 | 331 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
332 | struct pt_regs *regs = get_irq_regs(); |
333 | int duration; | |
ed235875 | 334 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 335 | |
6554fd8c TG |
336 | if (!watchdog_enabled || |
337 | atomic_read(&watchdog_park_in_progress) != 0) | |
b94f5118 DZ |
338 | return HRTIMER_NORESTART; |
339 | ||
58687acb DZ |
340 | /* kick the hardlockup detector */ |
341 | watchdog_interrupt_count(); | |
342 | ||
343 | /* kick the softlockup detector */ | |
909ea964 | 344 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
345 | |
346 | /* .. and repeat */ | |
0f34c400 | 347 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
348 | |
349 | if (touch_ts == 0) { | |
909ea964 | 350 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
351 | /* |
352 | * If the time stamp was touched atomically | |
353 | * make sure the scheduler tick is up to date. | |
354 | */ | |
909ea964 | 355 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
356 | sched_clock_tick(); |
357 | } | |
5d1c0f4a EM |
358 | |
359 | /* Clear the guest paused flag on watchdog reset */ | |
360 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
361 | __touch_watchdog(); |
362 | return HRTIMER_RESTART; | |
363 | } | |
364 | ||
365 | /* check for a softlockup | |
366 | * This is done by making sure a high priority task is | |
367 | * being scheduled. The task touches the watchdog to | |
368 | * indicate it is getting cpu time. If it hasn't then | |
369 | * this is a good indication some task is hogging the cpu | |
370 | */ | |
26e09c6e | 371 | duration = is_softlockup(touch_ts); |
58687acb | 372 | if (unlikely(duration)) { |
5d1c0f4a EM |
373 | /* |
374 | * If a virtual machine is stopped by the host it can look to | |
375 | * the watchdog like a soft lockup, check to see if the host | |
376 | * stopped the vm before we issue the warning | |
377 | */ | |
378 | if (kvm_check_and_clear_guest_paused()) | |
379 | return HRTIMER_RESTART; | |
380 | ||
58687acb | 381 | /* only warn once */ |
b1a8de1f | 382 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
383 | /* | |
384 | * When multiple processes are causing softlockups the | |
385 | * softlockup detector only warns on the first one | |
386 | * because the code relies on a full quiet cycle to | |
387 | * re-arm. The second process prevents the quiet cycle | |
388 | * and never gets reported. Use task pointers to detect | |
389 | * this. | |
390 | */ | |
391 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
392 | current) { | |
393 | __this_cpu_write(soft_watchdog_warn, false); | |
394 | __touch_watchdog(); | |
395 | } | |
58687acb | 396 | return HRTIMER_RESTART; |
b1a8de1f | 397 | } |
58687acb | 398 | |
ed235875 AT |
399 | if (softlockup_all_cpu_backtrace) { |
400 | /* Prevent multiple soft-lockup reports if one cpu is already | |
401 | * engaged in dumping cpu back traces | |
402 | */ | |
403 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
404 | /* Someone else will report us. Let's give up */ | |
405 | __this_cpu_write(soft_watchdog_warn, true); | |
406 | return HRTIMER_RESTART; | |
407 | } | |
408 | } | |
409 | ||
656c3b79 | 410 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 411 | smp_processor_id(), duration, |
58687acb | 412 | current->comm, task_pid_nr(current)); |
b1a8de1f | 413 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
414 | print_modules(); |
415 | print_irqtrace_events(current); | |
416 | if (regs) | |
417 | show_regs(regs); | |
418 | else | |
419 | dump_stack(); | |
420 | ||
ed235875 AT |
421 | if (softlockup_all_cpu_backtrace) { |
422 | /* Avoid generating two back traces for current | |
423 | * given that one is already made above | |
424 | */ | |
425 | trigger_allbutself_cpu_backtrace(); | |
426 | ||
427 | clear_bit(0, &soft_lockup_nmi_warn); | |
428 | /* Barrier to sync with other cpus */ | |
429 | smp_mb__after_atomic(); | |
430 | } | |
431 | ||
69361eef | 432 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
433 | if (softlockup_panic) |
434 | panic("softlockup: hung tasks"); | |
909ea964 | 435 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 436 | } else |
909ea964 | 437 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
438 | |
439 | return HRTIMER_RESTART; | |
440 | } | |
441 | ||
bcd951cf TG |
442 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
443 | { | |
444 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 445 | |
bcd951cf TG |
446 | sched_setscheduler(current, policy, ¶m); |
447 | } | |
448 | ||
449 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 450 | { |
f7f66b05 | 451 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 452 | |
3935e895 BM |
453 | /* kick off the timer for the hardlockup detector */ |
454 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
455 | hrtimer->function = watchdog_timer_fn; | |
456 | ||
bcd951cf TG |
457 | /* Enable the perf event */ |
458 | watchdog_nmi_enable(cpu); | |
58687acb | 459 | |
58687acb | 460 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
0f34c400 | 461 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
58687acb DZ |
462 | HRTIMER_MODE_REL_PINNED); |
463 | ||
bcd951cf TG |
464 | /* initialize timestamp */ |
465 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
466 | __touch_watchdog(); | |
467 | } | |
58687acb | 468 | |
bcd951cf TG |
469 | static void watchdog_disable(unsigned int cpu) |
470 | { | |
f7f66b05 | 471 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 472 | |
bcd951cf TG |
473 | watchdog_set_prio(SCHED_NORMAL, 0); |
474 | hrtimer_cancel(hrtimer); | |
475 | /* disable the perf event */ | |
476 | watchdog_nmi_disable(cpu); | |
58687acb DZ |
477 | } |
478 | ||
b8900bc0 FW |
479 | static void watchdog_cleanup(unsigned int cpu, bool online) |
480 | { | |
481 | watchdog_disable(cpu); | |
482 | } | |
483 | ||
bcd951cf TG |
484 | static int watchdog_should_run(unsigned int cpu) |
485 | { | |
486 | return __this_cpu_read(hrtimer_interrupts) != | |
487 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
488 | } | |
489 | ||
490 | /* | |
491 | * The watchdog thread function - touches the timestamp. | |
492 | * | |
0f34c400 | 493 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
494 | * default) to reset the softlockup timestamp. If this gets delayed |
495 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
496 | * triggers in watchdog_timer_fn(). | |
497 | */ | |
498 | static void watchdog(unsigned int cpu) | |
499 | { | |
500 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
501 | __this_cpu_read(hrtimer_interrupts)); | |
502 | __touch_watchdog(); | |
bcfba4f4 UO |
503 | |
504 | /* | |
505 | * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the | |
506 | * failure path. Check for failures that can occur asynchronously - | |
507 | * for example, when CPUs are on-lined - and shut down the hardware | |
508 | * perf event on each CPU accordingly. | |
509 | * | |
510 | * The only non-obvious place this bit can be cleared is through | |
511 | * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a | |
512 | * pr_info here would be too noisy as it would result in a message | |
513 | * every few seconds if the hardlockup was disabled but the softlockup | |
514 | * enabled. | |
515 | */ | |
516 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
517 | watchdog_nmi_disable(cpu); | |
bcd951cf | 518 | } |
58687acb | 519 | |
b8900bc0 FW |
520 | static 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 | ||
81a4beef UO |
531 | /* |
532 | * park all watchdog threads that are specified in 'watchdog_cpumask' | |
ee7fed54 UO |
533 | * |
534 | * This function returns an error if kthread_park() of a watchdog thread | |
535 | * fails. In this situation, the watchdog threads of some CPUs can already | |
536 | * be parked and the watchdog threads of other CPUs can still be runnable. | |
537 | * Callers are expected to handle this special condition as appropriate in | |
538 | * their context. | |
a2a45b85 UO |
539 | * |
540 | * This function may only be called in a context that is protected against | |
541 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
542 | */ |
543 | static int watchdog_park_threads(void) | |
544 | { | |
545 | int cpu, ret = 0; | |
546 | ||
b94f5118 DZ |
547 | atomic_set(&watchdog_park_in_progress, 1); |
548 | ||
81a4beef UO |
549 | for_each_watchdog_cpu(cpu) { |
550 | ret = kthread_park(per_cpu(softlockup_watchdog, cpu)); | |
551 | if (ret) | |
552 | break; | |
553 | } | |
81a4beef | 554 | |
b94f5118 DZ |
555 | atomic_set(&watchdog_park_in_progress, 0); |
556 | ||
81a4beef UO |
557 | return ret; |
558 | } | |
559 | ||
560 | /* | |
561 | * unpark all watchdog threads that are specified in 'watchdog_cpumask' | |
a2a45b85 UO |
562 | * |
563 | * This function may only be called in a context that is protected against | |
564 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
565 | */ |
566 | static void watchdog_unpark_threads(void) | |
567 | { | |
568 | int cpu; | |
569 | ||
81a4beef UO |
570 | for_each_watchdog_cpu(cpu) |
571 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); | |
81a4beef UO |
572 | } |
573 | ||
b43cb43c | 574 | static int update_watchdog_all_cpus(void) |
9809b18f | 575 | { |
b43cb43c UO |
576 | int ret; |
577 | ||
578 | ret = watchdog_park_threads(); | |
579 | if (ret) | |
580 | return ret; | |
581 | ||
d4bdd0b2 | 582 | watchdog_unpark_threads(); |
b43cb43c UO |
583 | |
584 | return 0; | |
9809b18f MH |
585 | } |
586 | ||
b2f57c3a | 587 | static int watchdog_enable_all_cpus(void) |
58687acb | 588 | { |
b8900bc0 | 589 | int err = 0; |
58687acb | 590 | |
3c00ea82 | 591 | if (!watchdog_running) { |
230ec939 FW |
592 | err = smpboot_register_percpu_thread_cpumask(&watchdog_threads, |
593 | &watchdog_cpumask); | |
b8900bc0 FW |
594 | if (err) |
595 | pr_err("Failed to create watchdog threads, disabled\n"); | |
230ec939 | 596 | else |
3c00ea82 | 597 | watchdog_running = 1; |
b2f57c3a UO |
598 | } else { |
599 | /* | |
600 | * Enable/disable the lockup detectors or | |
601 | * change the sample period 'on the fly'. | |
602 | */ | |
b43cb43c UO |
603 | err = update_watchdog_all_cpus(); |
604 | ||
605 | if (err) { | |
606 | watchdog_disable_all_cpus(); | |
607 | pr_err("Failed to update lockup detectors, disabled\n"); | |
608 | } | |
bcd951cf | 609 | } |
b8900bc0 | 610 | |
b43cb43c UO |
611 | if (err) |
612 | watchdog_enabled = 0; | |
613 | ||
b8900bc0 | 614 | return err; |
58687acb DZ |
615 | } |
616 | ||
617 | static void watchdog_disable_all_cpus(void) | |
618 | { | |
3c00ea82 FW |
619 | if (watchdog_running) { |
620 | watchdog_running = 0; | |
b8900bc0 | 621 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 622 | } |
58687acb DZ |
623 | } |
624 | ||
a10a842f NP |
625 | #ifdef CONFIG_SYSCTL |
626 | static int watchdog_update_cpus(void) | |
627 | { | |
628 | return smpboot_update_cpumask_percpu_thread( | |
629 | &watchdog_threads, &watchdog_cpumask); | |
630 | } | |
631 | #endif | |
632 | ||
05a4a952 NP |
633 | #else /* SOFTLOCKUP */ |
634 | static int watchdog_park_threads(void) | |
635 | { | |
636 | return 0; | |
637 | } | |
638 | ||
639 | static void watchdog_unpark_threads(void) | |
640 | { | |
641 | } | |
642 | ||
643 | static int watchdog_enable_all_cpus(void) | |
644 | { | |
645 | return 0; | |
646 | } | |
647 | ||
648 | static void watchdog_disable_all_cpus(void) | |
649 | { | |
650 | } | |
651 | ||
a10a842f NP |
652 | #ifdef CONFIG_SYSCTL |
653 | static int watchdog_update_cpus(void) | |
654 | { | |
655 | return 0; | |
656 | } | |
657 | #endif | |
658 | ||
05a4a952 NP |
659 | static void set_sample_period(void) |
660 | { | |
661 | } | |
662 | #endif /* SOFTLOCKUP */ | |
663 | ||
6554fd8c TG |
664 | /** |
665 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
666 | * | |
667 | * Special interface for parisc. It prevents lockup detector warnings from | |
668 | * the default pm_poweroff() function which busy loops forever. | |
669 | */ | |
670 | void lockup_detector_soft_poweroff(void) | |
671 | { | |
672 | watchdog_enabled = 0; | |
673 | } | |
674 | ||
05a4a952 NP |
675 | /* |
676 | * Suspend the hard and soft lockup detector by parking the watchdog threads. | |
677 | */ | |
678 | int lockup_detector_suspend(void) | |
679 | { | |
680 | int ret = 0; | |
681 | ||
682 | get_online_cpus(); | |
683 | mutex_lock(&watchdog_proc_mutex); | |
684 | /* | |
685 | * Multiple suspend requests can be active in parallel (counted by | |
686 | * the 'watchdog_suspended' variable). If the watchdog threads are | |
687 | * running, the first caller takes care that they will be parked. | |
688 | * The state of 'watchdog_running' cannot change while a suspend | |
689 | * request is active (see related code in 'proc' handlers). | |
690 | */ | |
691 | if (watchdog_running && !watchdog_suspended) | |
692 | ret = watchdog_park_threads(); | |
693 | ||
694 | if (ret == 0) | |
695 | watchdog_suspended++; | |
696 | else { | |
697 | watchdog_disable_all_cpus(); | |
698 | pr_err("Failed to suspend lockup detectors, disabled\n"); | |
699 | watchdog_enabled = 0; | |
700 | } | |
701 | ||
a10a842f NP |
702 | watchdog_nmi_reconfigure(); |
703 | ||
05a4a952 NP |
704 | mutex_unlock(&watchdog_proc_mutex); |
705 | ||
706 | return ret; | |
707 | } | |
708 | ||
709 | /* | |
710 | * Resume the hard and soft lockup detector by unparking the watchdog threads. | |
711 | */ | |
712 | void lockup_detector_resume(void) | |
713 | { | |
714 | mutex_lock(&watchdog_proc_mutex); | |
715 | ||
716 | watchdog_suspended--; | |
717 | /* | |
718 | * The watchdog threads are unparked if they were previously running | |
719 | * and if there is no more active suspend request. | |
720 | */ | |
721 | if (watchdog_running && !watchdog_suspended) | |
722 | watchdog_unpark_threads(); | |
723 | ||
a10a842f NP |
724 | watchdog_nmi_reconfigure(); |
725 | ||
05a4a952 NP |
726 | mutex_unlock(&watchdog_proc_mutex); |
727 | put_online_cpus(); | |
728 | } | |
729 | ||
58cf690a UO |
730 | #ifdef CONFIG_SYSCTL |
731 | ||
58687acb | 732 | /* |
a0c9cbb9 UO |
733 | * Update the run state of the lockup detectors. |
734 | */ | |
735 | static int proc_watchdog_update(void) | |
736 | { | |
737 | int err = 0; | |
738 | ||
739 | /* | |
740 | * Watchdog threads won't be started if they are already active. | |
741 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes | |
742 | * care of this. If those threads are already active, the sample | |
743 | * period will be updated and the lockup detectors will be enabled | |
744 | * or disabled 'on the fly'. | |
745 | */ | |
746 | if (watchdog_enabled && watchdog_thresh) | |
b2f57c3a | 747 | err = watchdog_enable_all_cpus(); |
a0c9cbb9 UO |
748 | else |
749 | watchdog_disable_all_cpus(); | |
750 | ||
a10a842f NP |
751 | watchdog_nmi_reconfigure(); |
752 | ||
a0c9cbb9 UO |
753 | return err; |
754 | ||
755 | } | |
756 | ||
ef246a21 UO |
757 | /* |
758 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
759 | * | |
760 | * caller | table->data points to | 'which' contains the flag(s) | |
761 | * -------------------|-----------------------|----------------------------- | |
762 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
763 | * | | with SOFT_WATCHDOG_ENABLED | |
764 | * -------------------|-----------------------|----------------------------- | |
765 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
766 | * -------------------|-----------------------|----------------------------- | |
767 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
768 | */ | |
769 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
770 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
771 | { | |
772 | int err, old, new; | |
773 | int *watchdog_param = (int *)table->data; | |
774 | ||
8614ddef | 775 | get_online_cpus(); |
ef246a21 UO |
776 | mutex_lock(&watchdog_proc_mutex); |
777 | ||
8c073d27 UO |
778 | if (watchdog_suspended) { |
779 | /* no parameter changes allowed while watchdog is suspended */ | |
780 | err = -EAGAIN; | |
781 | goto out; | |
782 | } | |
783 | ||
ef246a21 UO |
784 | /* |
785 | * If the parameter is being read return the state of the corresponding | |
786 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
787 | * run state of the lockup detectors. | |
788 | */ | |
789 | if (!write) { | |
790 | *watchdog_param = (watchdog_enabled & which) != 0; | |
791 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
792 | } else { | |
793 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
794 | if (err) | |
795 | goto out; | |
796 | ||
797 | /* | |
798 | * There is a race window between fetching the current value | |
799 | * from 'watchdog_enabled' and storing the new value. During | |
800 | * this race window, watchdog_nmi_enable() can sneak in and | |
801 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
802 | * The 'cmpxchg' detects this race and the loop retries. | |
803 | */ | |
804 | do { | |
805 | old = watchdog_enabled; | |
806 | /* | |
807 | * If the parameter value is not zero set the | |
808 | * corresponding bit(s), else clear it(them). | |
809 | */ | |
810 | if (*watchdog_param) | |
811 | new = old | which; | |
812 | else | |
813 | new = old & ~which; | |
814 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
815 | ||
816 | /* | |
b43cb43c UO |
817 | * Update the run state of the lockup detectors. There is _no_ |
818 | * need to check the value returned by proc_watchdog_update() | |
819 | * and to restore the previous value of 'watchdog_enabled' as | |
820 | * both lockup detectors are disabled if proc_watchdog_update() | |
821 | * returns an error. | |
ef246a21 | 822 | */ |
a1ee1932 JH |
823 | if (old == new) |
824 | goto out; | |
825 | ||
ef246a21 | 826 | err = proc_watchdog_update(); |
ef246a21 UO |
827 | } |
828 | out: | |
829 | mutex_unlock(&watchdog_proc_mutex); | |
8614ddef | 830 | put_online_cpus(); |
ef246a21 UO |
831 | return err; |
832 | } | |
833 | ||
83a80a39 UO |
834 | /* |
835 | * /proc/sys/kernel/watchdog | |
836 | */ | |
837 | int proc_watchdog(struct ctl_table *table, int write, | |
838 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
839 | { | |
840 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
841 | table, write, buffer, lenp, ppos); | |
842 | } | |
843 | ||
844 | /* | |
845 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 846 | */ |
83a80a39 UO |
847 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
848 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
849 | { | |
850 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
851 | table, write, buffer, lenp, ppos); | |
852 | } | |
853 | ||
854 | /* | |
855 | * /proc/sys/kernel/soft_watchdog | |
856 | */ | |
857 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
858 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
859 | { | |
860 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
861 | table, write, buffer, lenp, ppos); | |
862 | } | |
58687acb | 863 | |
83a80a39 UO |
864 | /* |
865 | * /proc/sys/kernel/watchdog_thresh | |
866 | */ | |
867 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
868 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 869 | { |
a1ee1932 | 870 | int err, old, new; |
58687acb | 871 | |
8614ddef | 872 | get_online_cpus(); |
359e6fab | 873 | mutex_lock(&watchdog_proc_mutex); |
bcd951cf | 874 | |
8c073d27 UO |
875 | if (watchdog_suspended) { |
876 | /* no parameter changes allowed while watchdog is suspended */ | |
877 | err = -EAGAIN; | |
878 | goto out; | |
879 | } | |
880 | ||
83a80a39 | 881 | old = ACCESS_ONCE(watchdog_thresh); |
b8900bc0 | 882 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 883 | |
b8900bc0 | 884 | if (err || !write) |
359e6fab | 885 | goto out; |
e04ab2bc | 886 | |
b66a2356 | 887 | /* |
d283c640 | 888 | * Update the sample period. Restore on failure. |
b66a2356 | 889 | */ |
a1ee1932 JH |
890 | new = ACCESS_ONCE(watchdog_thresh); |
891 | if (old == new) | |
892 | goto out; | |
893 | ||
83a80a39 UO |
894 | set_sample_period(); |
895 | err = proc_watchdog_update(); | |
d283c640 | 896 | if (err) { |
83a80a39 | 897 | watchdog_thresh = old; |
d283c640 UO |
898 | set_sample_period(); |
899 | } | |
359e6fab MH |
900 | out: |
901 | mutex_unlock(&watchdog_proc_mutex); | |
8614ddef | 902 | put_online_cpus(); |
b8900bc0 | 903 | return err; |
58687acb | 904 | } |
fe4ba3c3 CM |
905 | |
906 | /* | |
907 | * The cpumask is the mask of possible cpus that the watchdog can run | |
908 | * on, not the mask of cpus it is actually running on. This allows the | |
909 | * user to specify a mask that will include cpus that have not yet | |
910 | * been brought online, if desired. | |
911 | */ | |
912 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
913 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
914 | { | |
915 | int err; | |
916 | ||
8614ddef | 917 | get_online_cpus(); |
fe4ba3c3 | 918 | mutex_lock(&watchdog_proc_mutex); |
8c073d27 UO |
919 | |
920 | if (watchdog_suspended) { | |
921 | /* no parameter changes allowed while watchdog is suspended */ | |
922 | err = -EAGAIN; | |
923 | goto out; | |
924 | } | |
925 | ||
fe4ba3c3 CM |
926 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
927 | if (!err && write) { | |
928 | /* Remove impossible cpus to keep sysctl output cleaner. */ | |
929 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, | |
930 | cpu_possible_mask); | |
931 | ||
932 | if (watchdog_running) { | |
933 | /* | |
934 | * Failure would be due to being unable to allocate | |
935 | * a temporary cpumask, so we are likely not in a | |
936 | * position to do much else to make things better. | |
937 | */ | |
a10a842f | 938 | if (watchdog_update_cpus() != 0) |
fe4ba3c3 CM |
939 | pr_err("cpumask update failed\n"); |
940 | } | |
a10a842f NP |
941 | |
942 | watchdog_nmi_reconfigure(); | |
fe4ba3c3 | 943 | } |
8c073d27 | 944 | out: |
fe4ba3c3 | 945 | mutex_unlock(&watchdog_proc_mutex); |
8614ddef | 946 | put_online_cpus(); |
fe4ba3c3 CM |
947 | return err; |
948 | } | |
949 | ||
58687acb DZ |
950 | #endif /* CONFIG_SYSCTL */ |
951 | ||
004417a6 | 952 | void __init lockup_detector_init(void) |
58687acb | 953 | { |
0f34c400 | 954 | set_sample_period(); |
b8900bc0 | 955 | |
fe4ba3c3 CM |
956 | #ifdef CONFIG_NO_HZ_FULL |
957 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
958 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
959 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
960 | } else |
961 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
962 | #else | |
963 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
964 | #endif | |
965 | ||
195daf66 | 966 | if (watchdog_enabled) |
b2f57c3a | 967 | watchdog_enable_all_cpus(); |
58687acb | 968 | } |