]>
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 | |
946d1977 | 32 | static DEFINE_MUTEX(watchdog_mutex); |
ab992dc3 | 33 | |
05a4a952 NP |
34 | int __read_mostly nmi_watchdog_enabled; |
35 | ||
36 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) | |
37 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
38 | NMI_WATCHDOG_ENABLED; | |
84d56e66 | 39 | #else |
249e52e3 | 40 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
84d56e66 | 41 | #endif |
05a4a952 NP |
42 | |
43 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
05a4a952 NP |
44 | /* |
45 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
46 | */ | |
47 | unsigned int __read_mostly hardlockup_panic = | |
48 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
49 | /* | |
50 | * We may not want to enable hard lockup detection by default in all cases, | |
51 | * for example when running the kernel as a guest on a hypervisor. In these | |
52 | * cases this function can be called to disable hard lockup detection. This | |
53 | * function should only be executed once by the boot processor before the | |
54 | * kernel command line parameters are parsed, because otherwise it is not | |
55 | * possible to override this in hardlockup_panic_setup(). | |
56 | */ | |
7a355820 | 57 | void __init hardlockup_detector_disable(void) |
05a4a952 NP |
58 | { |
59 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
60 | } | |
61 | ||
62 | static int __init hardlockup_panic_setup(char *str) | |
63 | { | |
64 | if (!strncmp(str, "panic", 5)) | |
65 | hardlockup_panic = 1; | |
66 | else if (!strncmp(str, "nopanic", 7)) | |
67 | hardlockup_panic = 0; | |
68 | else if (!strncmp(str, "0", 1)) | |
69 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
70 | else if (!strncmp(str, "1", 1)) | |
71 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
72 | return 1; | |
73 | } | |
74 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
75 | ||
368a7e2c TG |
76 | # ifdef CONFIG_SMP |
77 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 78 | |
368a7e2c TG |
79 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
80 | { | |
81 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
82 | return 1; | |
83 | } | |
84 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
85 | # endif /* CONFIG_SMP */ | |
86 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 87 | |
84d56e66 | 88 | int __read_mostly watchdog_user_enabled; |
4eec42f3 | 89 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 90 | |
05a4a952 | 91 | struct cpumask watchdog_cpumask __read_mostly; |
fe4ba3c3 CM |
92 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); |
93 | ||
05a4a952 NP |
94 | /* |
95 | * These functions can be overridden if an architecture implements its | |
96 | * own hardlockup detector. | |
a10a842f NP |
97 | * |
98 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
99 | * softlockup watchdog threads start and stop. The arch must select the | |
100 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
101 | */ |
102 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
103 | { | |
104 | return 0; | |
105 | } | |
941154bd | 106 | |
05a4a952 NP |
107 | void __weak watchdog_nmi_disable(unsigned int cpu) |
108 | { | |
941154bd | 109 | hardlockup_detector_perf_disable(); |
05a4a952 NP |
110 | } |
111 | ||
a10a842f NP |
112 | /* |
113 | * watchdog_nmi_reconfigure can be implemented to be notified after any | |
114 | * watchdog configuration change. The arch hardlockup watchdog should | |
115 | * respond to the following variables: | |
116 | * - nmi_watchdog_enabled | |
117 | * - watchdog_thresh | |
118 | * - watchdog_cpumask | |
119 | * - sysctl_hardlockup_all_cpu_backtrace | |
120 | * - hardlockup_panic | |
a10a842f | 121 | */ |
2b9d7f23 | 122 | void __weak watchdog_nmi_reconfigure(void) { } |
a10a842f | 123 | |
05a4a952 NP |
124 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
125 | ||
2b9d7f23 TG |
126 | /* Global variables, exported for sysctl */ |
127 | unsigned int __read_mostly softlockup_panic = | |
128 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
129 | int __read_mostly soft_watchdog_enabled; | |
130 | ||
2eb2527f TG |
131 | struct cpumask watchdog_allowed_mask __read_mostly; |
132 | static bool softlockup_threads_initialized __read_mostly; | |
133 | ||
0f34c400 | 134 | static u64 __read_mostly sample_period; |
58687acb DZ |
135 | |
136 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
137 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
138 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
139 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 140 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
141 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
142 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 143 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 144 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 145 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 146 | |
58687acb DZ |
147 | static int __init softlockup_panic_setup(char *str) |
148 | { | |
149 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
150 | return 1; |
151 | } | |
152 | __setup("softlockup_panic=", softlockup_panic_setup); | |
153 | ||
154 | static int __init nowatchdog_setup(char *str) | |
155 | { | |
195daf66 | 156 | watchdog_enabled = 0; |
58687acb DZ |
157 | return 1; |
158 | } | |
159 | __setup("nowatchdog", nowatchdog_setup); | |
160 | ||
58687acb DZ |
161 | static int __init nosoftlockup_setup(char *str) |
162 | { | |
195daf66 | 163 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
164 | return 1; |
165 | } | |
166 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 167 | |
ed235875 | 168 | #ifdef CONFIG_SMP |
368a7e2c TG |
169 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
170 | ||
ed235875 AT |
171 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
172 | { | |
368a7e2c | 173 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
174 | return 1; |
175 | } | |
176 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 177 | #endif |
58687acb | 178 | |
941154bd TG |
179 | static void __lockup_detector_cleanup(void); |
180 | ||
4eec42f3 MSB |
181 | /* |
182 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
183 | * lockups can have false positives under extreme conditions. So we generally | |
184 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
185 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
186 | * time the hard threshold is. | |
187 | */ | |
6e9101ae | 188 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
189 | { |
190 | return watchdog_thresh * 2; | |
191 | } | |
58687acb DZ |
192 | |
193 | /* | |
194 | * Returns seconds, approximately. We don't need nanosecond | |
195 | * resolution, and we don't need to waste time with a big divide when | |
196 | * 2^30ns == 1.074s. | |
197 | */ | |
c06b4f19 | 198 | static unsigned long get_timestamp(void) |
58687acb | 199 | { |
545a2bf7 | 200 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
201 | } |
202 | ||
0f34c400 | 203 | static void set_sample_period(void) |
58687acb DZ |
204 | { |
205 | /* | |
586692a5 | 206 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
207 | * the divide by 5 is to give hrtimer several chances (two |
208 | * or three with the current relation between the soft | |
209 | * and hard thresholds) to increment before the | |
210 | * hardlockup detector generates a warning | |
58687acb | 211 | */ |
0f34c400 | 212 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 213 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
214 | } |
215 | ||
216 | /* Commands for resetting the watchdog */ | |
217 | static void __touch_watchdog(void) | |
218 | { | |
c06b4f19 | 219 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
220 | } |
221 | ||
03e0d461 TH |
222 | /** |
223 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
224 | * | |
225 | * Call when the scheduler may have stalled for legitimate reasons | |
226 | * preventing the watchdog task from executing - e.g. the scheduler | |
227 | * entering idle state. This should only be used for scheduler events. | |
228 | * Use touch_softlockup_watchdog() for everything else. | |
229 | */ | |
230 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 231 | { |
7861144b AM |
232 | /* |
233 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
234 | * gets zeroed here, so use the raw_ operation. | |
235 | */ | |
236 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 237 | } |
03e0d461 TH |
238 | |
239 | void touch_softlockup_watchdog(void) | |
240 | { | |
241 | touch_softlockup_watchdog_sched(); | |
82607adc | 242 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 243 | } |
0167c781 | 244 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 245 | |
332fbdbc | 246 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
247 | { |
248 | int cpu; | |
249 | ||
250 | /* | |
d57108d4 TG |
251 | * watchdog_mutex cannpt be taken here, as this might be called |
252 | * from (soft)interrupt context, so the access to | |
253 | * watchdog_allowed_cpumask might race with a concurrent update. | |
254 | * | |
255 | * The watchdog time stamp can race against a concurrent real | |
256 | * update as well, the only side effect might be a cycle delay for | |
257 | * the softlockup check. | |
58687acb | 258 | */ |
d57108d4 | 259 | for_each_cpu(cpu, &watchdog_allowed_mask) |
58687acb | 260 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 261 | wq_watchdog_touch(-1); |
58687acb DZ |
262 | } |
263 | ||
58687acb DZ |
264 | void touch_softlockup_watchdog_sync(void) |
265 | { | |
f7f66b05 CL |
266 | __this_cpu_write(softlockup_touch_sync, true); |
267 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
268 | } |
269 | ||
26e09c6e | 270 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 271 | { |
c06b4f19 | 272 | unsigned long now = get_timestamp(); |
58687acb | 273 | |
39d2da21 | 274 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
275 | /* Warn about unreasonable delays. */ |
276 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
277 | return now - touch_ts; | |
278 | } | |
58687acb DZ |
279 | return 0; |
280 | } | |
281 | ||
05a4a952 NP |
282 | /* watchdog detector functions */ |
283 | bool is_hardlockup(void) | |
58687acb | 284 | { |
05a4a952 | 285 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 286 | |
05a4a952 NP |
287 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
288 | return true; | |
289 | ||
290 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
291 | return false; | |
73ce0511 | 292 | } |
05a4a952 NP |
293 | |
294 | static void watchdog_interrupt_count(void) | |
73ce0511 | 295 | { |
05a4a952 | 296 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 297 | } |
58687acb DZ |
298 | |
299 | /* watchdog kicker functions */ | |
300 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
301 | { | |
909ea964 | 302 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
303 | struct pt_regs *regs = get_irq_regs(); |
304 | int duration; | |
ed235875 | 305 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 306 | |
01f0a027 | 307 | if (!watchdog_enabled) |
b94f5118 DZ |
308 | return HRTIMER_NORESTART; |
309 | ||
58687acb DZ |
310 | /* kick the hardlockup detector */ |
311 | watchdog_interrupt_count(); | |
312 | ||
313 | /* kick the softlockup detector */ | |
909ea964 | 314 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
315 | |
316 | /* .. and repeat */ | |
0f34c400 | 317 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
318 | |
319 | if (touch_ts == 0) { | |
909ea964 | 320 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
321 | /* |
322 | * If the time stamp was touched atomically | |
323 | * make sure the scheduler tick is up to date. | |
324 | */ | |
909ea964 | 325 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
326 | sched_clock_tick(); |
327 | } | |
5d1c0f4a EM |
328 | |
329 | /* Clear the guest paused flag on watchdog reset */ | |
330 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
331 | __touch_watchdog(); |
332 | return HRTIMER_RESTART; | |
333 | } | |
334 | ||
335 | /* check for a softlockup | |
336 | * This is done by making sure a high priority task is | |
337 | * being scheduled. The task touches the watchdog to | |
338 | * indicate it is getting cpu time. If it hasn't then | |
339 | * this is a good indication some task is hogging the cpu | |
340 | */ | |
26e09c6e | 341 | duration = is_softlockup(touch_ts); |
58687acb | 342 | if (unlikely(duration)) { |
5d1c0f4a EM |
343 | /* |
344 | * If a virtual machine is stopped by the host it can look to | |
345 | * the watchdog like a soft lockup, check to see if the host | |
346 | * stopped the vm before we issue the warning | |
347 | */ | |
348 | if (kvm_check_and_clear_guest_paused()) | |
349 | return HRTIMER_RESTART; | |
350 | ||
58687acb | 351 | /* only warn once */ |
b1a8de1f | 352 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
353 | /* | |
354 | * When multiple processes are causing softlockups the | |
355 | * softlockup detector only warns on the first one | |
356 | * because the code relies on a full quiet cycle to | |
357 | * re-arm. The second process prevents the quiet cycle | |
358 | * and never gets reported. Use task pointers to detect | |
359 | * this. | |
360 | */ | |
361 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
362 | current) { | |
363 | __this_cpu_write(soft_watchdog_warn, false); | |
364 | __touch_watchdog(); | |
365 | } | |
58687acb | 366 | return HRTIMER_RESTART; |
b1a8de1f | 367 | } |
58687acb | 368 | |
ed235875 AT |
369 | if (softlockup_all_cpu_backtrace) { |
370 | /* Prevent multiple soft-lockup reports if one cpu is already | |
371 | * engaged in dumping cpu back traces | |
372 | */ | |
373 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
374 | /* Someone else will report us. Let's give up */ | |
375 | __this_cpu_write(soft_watchdog_warn, true); | |
376 | return HRTIMER_RESTART; | |
377 | } | |
378 | } | |
379 | ||
656c3b79 | 380 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 381 | smp_processor_id(), duration, |
58687acb | 382 | current->comm, task_pid_nr(current)); |
b1a8de1f | 383 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
384 | print_modules(); |
385 | print_irqtrace_events(current); | |
386 | if (regs) | |
387 | show_regs(regs); | |
388 | else | |
389 | dump_stack(); | |
390 | ||
ed235875 AT |
391 | if (softlockup_all_cpu_backtrace) { |
392 | /* Avoid generating two back traces for current | |
393 | * given that one is already made above | |
394 | */ | |
395 | trigger_allbutself_cpu_backtrace(); | |
396 | ||
397 | clear_bit(0, &soft_lockup_nmi_warn); | |
398 | /* Barrier to sync with other cpus */ | |
399 | smp_mb__after_atomic(); | |
400 | } | |
401 | ||
69361eef | 402 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
403 | if (softlockup_panic) |
404 | panic("softlockup: hung tasks"); | |
909ea964 | 405 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 406 | } else |
909ea964 | 407 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
408 | |
409 | return HRTIMER_RESTART; | |
410 | } | |
411 | ||
bcd951cf TG |
412 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
413 | { | |
414 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 415 | |
bcd951cf TG |
416 | sched_setscheduler(current, policy, ¶m); |
417 | } | |
418 | ||
419 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 420 | { |
01f0a027 | 421 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 422 | |
01f0a027 TG |
423 | /* |
424 | * Start the timer first to prevent the NMI watchdog triggering | |
425 | * before the timer has a chance to fire. | |
426 | */ | |
3935e895 BM |
427 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
428 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
429 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
430 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 431 | |
01f0a027 TG |
432 | /* Initialize timestamp */ |
433 | __touch_watchdog(); | |
bcd951cf TG |
434 | /* Enable the perf event */ |
435 | watchdog_nmi_enable(cpu); | |
58687acb | 436 | |
bcd951cf | 437 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 438 | } |
58687acb | 439 | |
bcd951cf TG |
440 | static void watchdog_disable(unsigned int cpu) |
441 | { | |
01f0a027 | 442 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 443 | |
bcd951cf | 444 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
445 | /* |
446 | * Disable the perf event first. That prevents that a large delay | |
447 | * between disabling the timer and disabling the perf event causes | |
448 | * the perf NMI to detect a false positive. | |
449 | */ | |
bcd951cf | 450 | watchdog_nmi_disable(cpu); |
01f0a027 | 451 | hrtimer_cancel(hrtimer); |
58687acb DZ |
452 | } |
453 | ||
b8900bc0 FW |
454 | static void watchdog_cleanup(unsigned int cpu, bool online) |
455 | { | |
456 | watchdog_disable(cpu); | |
457 | } | |
458 | ||
bcd951cf TG |
459 | static int watchdog_should_run(unsigned int cpu) |
460 | { | |
461 | return __this_cpu_read(hrtimer_interrupts) != | |
462 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
463 | } | |
464 | ||
465 | /* | |
466 | * The watchdog thread function - touches the timestamp. | |
467 | * | |
0f34c400 | 468 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
469 | * default) to reset the softlockup timestamp. If this gets delayed |
470 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
471 | * triggers in watchdog_timer_fn(). | |
472 | */ | |
473 | static void watchdog(unsigned int cpu) | |
474 | { | |
475 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
476 | __this_cpu_read(hrtimer_interrupts)); | |
477 | __touch_watchdog(); | |
478 | } | |
58687acb | 479 | |
b8900bc0 FW |
480 | static struct smp_hotplug_thread watchdog_threads = { |
481 | .store = &softlockup_watchdog, | |
482 | .thread_should_run = watchdog_should_run, | |
483 | .thread_fn = watchdog, | |
484 | .thread_comm = "watchdog/%u", | |
485 | .setup = watchdog_enable, | |
486 | .cleanup = watchdog_cleanup, | |
487 | .park = watchdog_disable, | |
488 | .unpark = watchdog_enable, | |
489 | }; | |
490 | ||
2eb2527f TG |
491 | static void softlockup_update_smpboot_threads(void) |
492 | { | |
493 | lockdep_assert_held(&watchdog_mutex); | |
494 | ||
495 | if (!softlockup_threads_initialized) | |
496 | return; | |
497 | ||
498 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
499 | &watchdog_allowed_mask); | |
500 | __lockup_detector_cleanup(); | |
501 | } | |
502 | ||
503 | /* Temporarily park all watchdog threads */ | |
504 | static void softlockup_park_all_threads(void) | |
505 | { | |
506 | cpumask_clear(&watchdog_allowed_mask); | |
507 | softlockup_update_smpboot_threads(); | |
508 | } | |
509 | ||
e8b62b2d TG |
510 | /* Unpark enabled threads */ |
511 | static void softlockup_unpark_threads(void) | |
2eb2527f TG |
512 | { |
513 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
514 | softlockup_update_smpboot_threads(); | |
515 | } | |
516 | ||
517 | static void softlockup_reconfigure_threads(bool enabled) | |
518 | { | |
519 | softlockup_park_all_threads(); | |
520 | set_sample_period(); | |
521 | if (enabled) | |
e8b62b2d | 522 | softlockup_unpark_threads(); |
2eb2527f TG |
523 | } |
524 | ||
525 | /* | |
526 | * Create the watchdog thread infrastructure. | |
527 | * | |
528 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
529 | * the threads are sucessfully initialized, take the proper locks and | |
530 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
531 | */ | |
532 | static __init void softlockup_init_threads(void) | |
533 | { | |
534 | int ret; | |
535 | ||
536 | /* | |
537 | * If sysctl is off and watchdog got disabled on the command line, | |
538 | * nothing to do here. | |
539 | */ | |
540 | if (!IS_ENABLED(CONFIG_SYSCTL) && | |
541 | !(watchdog_enabled && watchdog_thresh)) | |
542 | return; | |
543 | ||
544 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
545 | &watchdog_allowed_mask); | |
546 | if (ret) { | |
547 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
548 | return; | |
549 | } | |
550 | ||
551 | mutex_lock(&watchdog_mutex); | |
552 | softlockup_threads_initialized = true; | |
553 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); | |
554 | mutex_unlock(&watchdog_mutex); | |
555 | } | |
556 | ||
2b9d7f23 TG |
557 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
558 | static inline int watchdog_park_threads(void) { return 0; } | |
559 | static inline void watchdog_unpark_threads(void) { } | |
560 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
561 | static inline void watchdog_disable_all_cpus(void) { } | |
2eb2527f | 562 | static inline void softlockup_init_threads(void) { } |
2eb2527f | 563 | static inline void softlockup_reconfigure_threads(bool enabled) { } |
2b9d7f23 | 564 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 565 | |
941154bd TG |
566 | static void __lockup_detector_cleanup(void) |
567 | { | |
568 | lockdep_assert_held(&watchdog_mutex); | |
569 | hardlockup_detector_perf_cleanup(); | |
570 | } | |
571 | ||
572 | /** | |
573 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
574 | * | |
575 | * Caller must not hold the cpu hotplug rwsem. | |
576 | */ | |
577 | void lockup_detector_cleanup(void) | |
578 | { | |
579 | mutex_lock(&watchdog_mutex); | |
580 | __lockup_detector_cleanup(); | |
581 | mutex_unlock(&watchdog_mutex); | |
582 | } | |
583 | ||
6554fd8c TG |
584 | /** |
585 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
586 | * | |
587 | * Special interface for parisc. It prevents lockup detector warnings from | |
588 | * the default pm_poweroff() function which busy loops forever. | |
589 | */ | |
590 | void lockup_detector_soft_poweroff(void) | |
591 | { | |
592 | watchdog_enabled = 0; | |
593 | } | |
594 | ||
58cf690a UO |
595 | #ifdef CONFIG_SYSCTL |
596 | ||
e8b62b2d | 597 | /* Propagate any changes to the watchdog threads */ |
d57108d4 | 598 | static void proc_watchdog_update(void) |
a0c9cbb9 | 599 | { |
e8b62b2d TG |
600 | /* Remove impossible cpus to keep sysctl output clean. */ |
601 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
d57108d4 | 602 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); |
a10a842f | 603 | watchdog_nmi_reconfigure(); |
a0c9cbb9 UO |
604 | } |
605 | ||
ef246a21 UO |
606 | /* |
607 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
608 | * | |
609 | * caller | table->data points to | 'which' contains the flag(s) | |
610 | * -------------------|-----------------------|----------------------------- | |
611 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
612 | * | | with SOFT_WATCHDOG_ENABLED | |
613 | * -------------------|-----------------------|----------------------------- | |
614 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
615 | * -------------------|-----------------------|----------------------------- | |
616 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
617 | */ | |
618 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
619 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
620 | { | |
621 | int err, old, new; | |
622 | int *watchdog_param = (int *)table->data; | |
623 | ||
b7a34981 | 624 | cpu_hotplug_disable(); |
946d1977 | 625 | mutex_lock(&watchdog_mutex); |
ef246a21 UO |
626 | |
627 | /* | |
628 | * If the parameter is being read return the state of the corresponding | |
629 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
630 | * run state of the lockup detectors. | |
631 | */ | |
632 | if (!write) { | |
633 | *watchdog_param = (watchdog_enabled & which) != 0; | |
634 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
635 | } else { | |
636 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
637 | if (err) | |
638 | goto out; | |
639 | ||
640 | /* | |
641 | * There is a race window between fetching the current value | |
642 | * from 'watchdog_enabled' and storing the new value. During | |
643 | * this race window, watchdog_nmi_enable() can sneak in and | |
644 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
645 | * The 'cmpxchg' detects this race and the loop retries. | |
646 | */ | |
647 | do { | |
648 | old = watchdog_enabled; | |
649 | /* | |
650 | * If the parameter value is not zero set the | |
651 | * corresponding bit(s), else clear it(them). | |
652 | */ | |
653 | if (*watchdog_param) | |
654 | new = old | which; | |
655 | else | |
656 | new = old & ~which; | |
657 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
658 | ||
d57108d4 TG |
659 | if (old != new) |
660 | proc_watchdog_update(); | |
ef246a21 UO |
661 | } |
662 | out: | |
946d1977 | 663 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 664 | cpu_hotplug_enable(); |
ef246a21 UO |
665 | return err; |
666 | } | |
667 | ||
83a80a39 UO |
668 | /* |
669 | * /proc/sys/kernel/watchdog | |
670 | */ | |
671 | int proc_watchdog(struct ctl_table *table, int write, | |
672 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
673 | { | |
674 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
675 | table, write, buffer, lenp, ppos); | |
676 | } | |
677 | ||
678 | /* | |
679 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 680 | */ |
83a80a39 UO |
681 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
682 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
683 | { | |
684 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
685 | table, write, buffer, lenp, ppos); | |
686 | } | |
687 | ||
688 | /* | |
689 | * /proc/sys/kernel/soft_watchdog | |
690 | */ | |
691 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
692 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
693 | { | |
694 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
695 | table, write, buffer, lenp, ppos); | |
696 | } | |
58687acb | 697 | |
83a80a39 UO |
698 | /* |
699 | * /proc/sys/kernel/watchdog_thresh | |
700 | */ | |
701 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
702 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 703 | { |
d57108d4 | 704 | int err, old; |
58687acb | 705 | |
b7a34981 | 706 | cpu_hotplug_disable(); |
946d1977 | 707 | mutex_lock(&watchdog_mutex); |
bcd951cf | 708 | |
d57108d4 | 709 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 710 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 711 | |
d57108d4 TG |
712 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
713 | proc_watchdog_update(); | |
e04ab2bc | 714 | |
946d1977 | 715 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 716 | cpu_hotplug_enable(); |
b8900bc0 | 717 | return err; |
58687acb | 718 | } |
fe4ba3c3 CM |
719 | |
720 | /* | |
721 | * The cpumask is the mask of possible cpus that the watchdog can run | |
722 | * on, not the mask of cpus it is actually running on. This allows the | |
723 | * user to specify a mask that will include cpus that have not yet | |
724 | * been brought online, if desired. | |
725 | */ | |
726 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
727 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
728 | { | |
729 | int err; | |
730 | ||
b7a34981 | 731 | cpu_hotplug_disable(); |
946d1977 | 732 | mutex_lock(&watchdog_mutex); |
8c073d27 | 733 | |
fe4ba3c3 | 734 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 735 | if (!err && write) |
e8b62b2d | 736 | proc_watchdog_update(); |
5490125d | 737 | |
946d1977 | 738 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 739 | cpu_hotplug_enable(); |
fe4ba3c3 CM |
740 | return err; |
741 | } | |
58687acb DZ |
742 | #endif /* CONFIG_SYSCTL */ |
743 | ||
004417a6 | 744 | void __init lockup_detector_init(void) |
58687acb | 745 | { |
fe4ba3c3 CM |
746 | #ifdef CONFIG_NO_HZ_FULL |
747 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
748 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
749 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
750 | } else |
751 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
752 | #else | |
753 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
754 | #endif | |
755 | ||
d57108d4 | 756 | softlockup_init_threads(); |
58687acb | 757 | } |