]>
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> |
78634061 | 27 | #include <linux/sched/isolation.h> |
58687acb DZ |
28 | |
29 | #include <asm/irq_regs.h> | |
5d1c0f4a | 30 | #include <linux/kvm_para.h> |
81a4beef | 31 | #include <linux/kthread.h> |
58687acb | 32 | |
946d1977 | 33 | static DEFINE_MUTEX(watchdog_mutex); |
ab992dc3 | 34 | |
05a4a952 | 35 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) |
09154985 TG |
36 | # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED | NMI_WATCHDOG_ENABLED) |
37 | # define NMI_WATCHDOG_DEFAULT 1 | |
84d56e66 | 38 | #else |
09154985 TG |
39 | # define WATCHDOG_DEFAULT (SOFT_WATCHDOG_ENABLED) |
40 | # define NMI_WATCHDOG_DEFAULT 0 | |
84d56e66 | 41 | #endif |
05a4a952 | 42 | |
09154985 TG |
43 | unsigned long __read_mostly watchdog_enabled; |
44 | int __read_mostly watchdog_user_enabled = 1; | |
45 | int __read_mostly nmi_watchdog_user_enabled = NMI_WATCHDOG_DEFAULT; | |
46 | int __read_mostly soft_watchdog_user_enabled = 1; | |
7feeb9cd | 47 | int __read_mostly watchdog_thresh = 10; |
a994a314 | 48 | int __read_mostly nmi_watchdog_available; |
7feeb9cd TG |
49 | |
50 | struct cpumask watchdog_allowed_mask __read_mostly; | |
7feeb9cd TG |
51 | |
52 | struct cpumask watchdog_cpumask __read_mostly; | |
53 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
54 | ||
05a4a952 | 55 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
05a4a952 NP |
56 | /* |
57 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
58 | */ | |
59 | unsigned int __read_mostly hardlockup_panic = | |
60 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
61 | /* | |
62 | * We may not want to enable hard lockup detection by default in all cases, | |
63 | * for example when running the kernel as a guest on a hypervisor. In these | |
64 | * cases this function can be called to disable hard lockup detection. This | |
65 | * function should only be executed once by the boot processor before the | |
66 | * kernel command line parameters are parsed, because otherwise it is not | |
67 | * possible to override this in hardlockup_panic_setup(). | |
68 | */ | |
7a355820 | 69 | void __init hardlockup_detector_disable(void) |
05a4a952 | 70 | { |
09154985 | 71 | nmi_watchdog_user_enabled = 0; |
05a4a952 NP |
72 | } |
73 | ||
74 | static int __init hardlockup_panic_setup(char *str) | |
75 | { | |
76 | if (!strncmp(str, "panic", 5)) | |
77 | hardlockup_panic = 1; | |
78 | else if (!strncmp(str, "nopanic", 7)) | |
79 | hardlockup_panic = 0; | |
80 | else if (!strncmp(str, "0", 1)) | |
09154985 | 81 | nmi_watchdog_user_enabled = 0; |
05a4a952 | 82 | else if (!strncmp(str, "1", 1)) |
09154985 | 83 | nmi_watchdog_user_enabled = 1; |
05a4a952 NP |
84 | return 1; |
85 | } | |
86 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
87 | ||
368a7e2c TG |
88 | # ifdef CONFIG_SMP |
89 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 90 | |
368a7e2c TG |
91 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
92 | { | |
93 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
94 | return 1; | |
95 | } | |
96 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
97 | # endif /* CONFIG_SMP */ | |
98 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 99 | |
05a4a952 NP |
100 | /* |
101 | * These functions can be overridden if an architecture implements its | |
102 | * own hardlockup detector. | |
a10a842f NP |
103 | * |
104 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
105 | * softlockup watchdog threads start and stop. The arch must select the | |
106 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
107 | */ |
108 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
109 | { | |
146c9d0e | 110 | hardlockup_detector_perf_enable(); |
05a4a952 NP |
111 | return 0; |
112 | } | |
941154bd | 113 | |
05a4a952 NP |
114 | void __weak watchdog_nmi_disable(unsigned int cpu) |
115 | { | |
941154bd | 116 | hardlockup_detector_perf_disable(); |
05a4a952 NP |
117 | } |
118 | ||
a994a314 TG |
119 | /* Return 0, if a NMI watchdog is available. Error code otherwise */ |
120 | int __weak __init watchdog_nmi_probe(void) | |
121 | { | |
122 | return hardlockup_detector_perf_init(); | |
123 | } | |
124 | ||
6592ad2f | 125 | /** |
6b9dc480 | 126 | * watchdog_nmi_stop - Stop the watchdog for reconfiguration |
6592ad2f | 127 | * |
6b9dc480 TG |
128 | * The reconfiguration steps are: |
129 | * watchdog_nmi_stop(); | |
6592ad2f | 130 | * update_variables(); |
6b9dc480 TG |
131 | * watchdog_nmi_start(); |
132 | */ | |
133 | void __weak watchdog_nmi_stop(void) { } | |
134 | ||
135 | /** | |
136 | * watchdog_nmi_start - Start the watchdog after reconfiguration | |
6592ad2f | 137 | * |
6b9dc480 TG |
138 | * Counterpart to watchdog_nmi_stop(). |
139 | * | |
140 | * The following variables have been updated in update_variables() and | |
141 | * contain the currently valid configuration: | |
7feeb9cd | 142 | * - watchdog_enabled |
a10a842f NP |
143 | * - watchdog_thresh |
144 | * - watchdog_cpumask | |
a10a842f | 145 | */ |
6b9dc480 | 146 | void __weak watchdog_nmi_start(void) { } |
a10a842f | 147 | |
09154985 TG |
148 | /** |
149 | * lockup_detector_update_enable - Update the sysctl enable bit | |
150 | * | |
151 | * Caller needs to make sure that the NMI/perf watchdogs are off, so this | |
152 | * can't race with watchdog_nmi_disable(). | |
153 | */ | |
154 | static void lockup_detector_update_enable(void) | |
155 | { | |
156 | watchdog_enabled = 0; | |
157 | if (!watchdog_user_enabled) | |
158 | return; | |
a994a314 | 159 | if (nmi_watchdog_available && nmi_watchdog_user_enabled) |
09154985 TG |
160 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; |
161 | if (soft_watchdog_user_enabled) | |
162 | watchdog_enabled |= SOFT_WATCHDOG_ENABLED; | |
163 | } | |
164 | ||
05a4a952 NP |
165 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
166 | ||
2b9d7f23 TG |
167 | /* Global variables, exported for sysctl */ |
168 | unsigned int __read_mostly softlockup_panic = | |
169 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
2eb2527f | 170 | |
0b62bf86 | 171 | static bool softlockup_threads_initialized __read_mostly; |
0f34c400 | 172 | static u64 __read_mostly sample_period; |
58687acb DZ |
173 | |
174 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
175 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
176 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
177 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 178 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
179 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
180 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 181 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 182 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 183 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 184 | |
58687acb DZ |
185 | static int __init softlockup_panic_setup(char *str) |
186 | { | |
187 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
188 | return 1; |
189 | } | |
190 | __setup("softlockup_panic=", softlockup_panic_setup); | |
191 | ||
192 | static int __init nowatchdog_setup(char *str) | |
193 | { | |
09154985 | 194 | watchdog_user_enabled = 0; |
58687acb DZ |
195 | return 1; |
196 | } | |
197 | __setup("nowatchdog", nowatchdog_setup); | |
198 | ||
58687acb DZ |
199 | static int __init nosoftlockup_setup(char *str) |
200 | { | |
09154985 | 201 | soft_watchdog_user_enabled = 0; |
58687acb DZ |
202 | return 1; |
203 | } | |
204 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 205 | |
ed235875 | 206 | #ifdef CONFIG_SMP |
368a7e2c TG |
207 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
208 | ||
ed235875 AT |
209 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
210 | { | |
368a7e2c | 211 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
212 | return 1; |
213 | } | |
214 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 215 | #endif |
58687acb | 216 | |
941154bd TG |
217 | static void __lockup_detector_cleanup(void); |
218 | ||
4eec42f3 MSB |
219 | /* |
220 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
221 | * lockups can have false positives under extreme conditions. So we generally | |
222 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
223 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
224 | * time the hard threshold is. | |
225 | */ | |
6e9101ae | 226 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
227 | { |
228 | return watchdog_thresh * 2; | |
229 | } | |
58687acb DZ |
230 | |
231 | /* | |
232 | * Returns seconds, approximately. We don't need nanosecond | |
233 | * resolution, and we don't need to waste time with a big divide when | |
234 | * 2^30ns == 1.074s. | |
235 | */ | |
c06b4f19 | 236 | static unsigned long get_timestamp(void) |
58687acb | 237 | { |
545a2bf7 | 238 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
239 | } |
240 | ||
0f34c400 | 241 | static void set_sample_period(void) |
58687acb DZ |
242 | { |
243 | /* | |
586692a5 | 244 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
245 | * the divide by 5 is to give hrtimer several chances (two |
246 | * or three with the current relation between the soft | |
247 | * and hard thresholds) to increment before the | |
248 | * hardlockup detector generates a warning | |
58687acb | 249 | */ |
0f34c400 | 250 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 251 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
252 | } |
253 | ||
254 | /* Commands for resetting the watchdog */ | |
255 | static void __touch_watchdog(void) | |
256 | { | |
c06b4f19 | 257 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
258 | } |
259 | ||
03e0d461 TH |
260 | /** |
261 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
262 | * | |
263 | * Call when the scheduler may have stalled for legitimate reasons | |
264 | * preventing the watchdog task from executing - e.g. the scheduler | |
265 | * entering idle state. This should only be used for scheduler events. | |
266 | * Use touch_softlockup_watchdog() for everything else. | |
267 | */ | |
268 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 269 | { |
7861144b AM |
270 | /* |
271 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
272 | * gets zeroed here, so use the raw_ operation. | |
273 | */ | |
274 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 275 | } |
03e0d461 TH |
276 | |
277 | void touch_softlockup_watchdog(void) | |
278 | { | |
279 | touch_softlockup_watchdog_sched(); | |
82607adc | 280 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 281 | } |
0167c781 | 282 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 283 | |
332fbdbc | 284 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
285 | { |
286 | int cpu; | |
287 | ||
288 | /* | |
d57108d4 TG |
289 | * watchdog_mutex cannpt be taken here, as this might be called |
290 | * from (soft)interrupt context, so the access to | |
291 | * watchdog_allowed_cpumask might race with a concurrent update. | |
292 | * | |
293 | * The watchdog time stamp can race against a concurrent real | |
294 | * update as well, the only side effect might be a cycle delay for | |
295 | * the softlockup check. | |
58687acb | 296 | */ |
d57108d4 | 297 | for_each_cpu(cpu, &watchdog_allowed_mask) |
58687acb | 298 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 299 | wq_watchdog_touch(-1); |
58687acb DZ |
300 | } |
301 | ||
58687acb DZ |
302 | void touch_softlockup_watchdog_sync(void) |
303 | { | |
f7f66b05 CL |
304 | __this_cpu_write(softlockup_touch_sync, true); |
305 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
306 | } |
307 | ||
26e09c6e | 308 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 309 | { |
c06b4f19 | 310 | unsigned long now = get_timestamp(); |
58687acb | 311 | |
39d2da21 | 312 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
313 | /* Warn about unreasonable delays. */ |
314 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
315 | return now - touch_ts; | |
316 | } | |
58687acb DZ |
317 | return 0; |
318 | } | |
319 | ||
05a4a952 NP |
320 | /* watchdog detector functions */ |
321 | bool is_hardlockup(void) | |
58687acb | 322 | { |
05a4a952 | 323 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 324 | |
05a4a952 NP |
325 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
326 | return true; | |
327 | ||
328 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
329 | return false; | |
73ce0511 | 330 | } |
05a4a952 NP |
331 | |
332 | static void watchdog_interrupt_count(void) | |
73ce0511 | 333 | { |
05a4a952 | 334 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 335 | } |
58687acb DZ |
336 | |
337 | /* watchdog kicker functions */ | |
338 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
339 | { | |
909ea964 | 340 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
341 | struct pt_regs *regs = get_irq_regs(); |
342 | int duration; | |
ed235875 | 343 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 344 | |
01f0a027 | 345 | if (!watchdog_enabled) |
b94f5118 DZ |
346 | return HRTIMER_NORESTART; |
347 | ||
58687acb DZ |
348 | /* kick the hardlockup detector */ |
349 | watchdog_interrupt_count(); | |
350 | ||
351 | /* kick the softlockup detector */ | |
909ea964 | 352 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
353 | |
354 | /* .. and repeat */ | |
0f34c400 | 355 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
356 | |
357 | if (touch_ts == 0) { | |
909ea964 | 358 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
359 | /* |
360 | * If the time stamp was touched atomically | |
361 | * make sure the scheduler tick is up to date. | |
362 | */ | |
909ea964 | 363 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
364 | sched_clock_tick(); |
365 | } | |
5d1c0f4a EM |
366 | |
367 | /* Clear the guest paused flag on watchdog reset */ | |
368 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
369 | __touch_watchdog(); |
370 | return HRTIMER_RESTART; | |
371 | } | |
372 | ||
373 | /* check for a softlockup | |
374 | * This is done by making sure a high priority task is | |
375 | * being scheduled. The task touches the watchdog to | |
376 | * indicate it is getting cpu time. If it hasn't then | |
377 | * this is a good indication some task is hogging the cpu | |
378 | */ | |
26e09c6e | 379 | duration = is_softlockup(touch_ts); |
58687acb | 380 | if (unlikely(duration)) { |
5d1c0f4a EM |
381 | /* |
382 | * If a virtual machine is stopped by the host it can look to | |
383 | * the watchdog like a soft lockup, check to see if the host | |
384 | * stopped the vm before we issue the warning | |
385 | */ | |
386 | if (kvm_check_and_clear_guest_paused()) | |
387 | return HRTIMER_RESTART; | |
388 | ||
58687acb | 389 | /* only warn once */ |
b1a8de1f | 390 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
391 | /* | |
392 | * When multiple processes are causing softlockups the | |
393 | * softlockup detector only warns on the first one | |
394 | * because the code relies on a full quiet cycle to | |
395 | * re-arm. The second process prevents the quiet cycle | |
396 | * and never gets reported. Use task pointers to detect | |
397 | * this. | |
398 | */ | |
399 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
400 | current) { | |
401 | __this_cpu_write(soft_watchdog_warn, false); | |
402 | __touch_watchdog(); | |
403 | } | |
58687acb | 404 | return HRTIMER_RESTART; |
b1a8de1f | 405 | } |
58687acb | 406 | |
ed235875 AT |
407 | if (softlockup_all_cpu_backtrace) { |
408 | /* Prevent multiple soft-lockup reports if one cpu is already | |
409 | * engaged in dumping cpu back traces | |
410 | */ | |
411 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
412 | /* Someone else will report us. Let's give up */ | |
413 | __this_cpu_write(soft_watchdog_warn, true); | |
414 | return HRTIMER_RESTART; | |
415 | } | |
416 | } | |
417 | ||
656c3b79 | 418 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 419 | smp_processor_id(), duration, |
58687acb | 420 | current->comm, task_pid_nr(current)); |
b1a8de1f | 421 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
422 | print_modules(); |
423 | print_irqtrace_events(current); | |
424 | if (regs) | |
425 | show_regs(regs); | |
426 | else | |
427 | dump_stack(); | |
428 | ||
ed235875 AT |
429 | if (softlockup_all_cpu_backtrace) { |
430 | /* Avoid generating two back traces for current | |
431 | * given that one is already made above | |
432 | */ | |
433 | trigger_allbutself_cpu_backtrace(); | |
434 | ||
435 | clear_bit(0, &soft_lockup_nmi_warn); | |
436 | /* Barrier to sync with other cpus */ | |
437 | smp_mb__after_atomic(); | |
438 | } | |
439 | ||
69361eef | 440 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
441 | if (softlockup_panic) |
442 | panic("softlockup: hung tasks"); | |
909ea964 | 443 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 444 | } else |
909ea964 | 445 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
446 | |
447 | return HRTIMER_RESTART; | |
448 | } | |
449 | ||
bcd951cf TG |
450 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
451 | { | |
452 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 453 | |
bcd951cf TG |
454 | sched_setscheduler(current, policy, ¶m); |
455 | } | |
456 | ||
457 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 458 | { |
01f0a027 | 459 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 460 | |
01f0a027 TG |
461 | /* |
462 | * Start the timer first to prevent the NMI watchdog triggering | |
463 | * before the timer has a chance to fire. | |
464 | */ | |
3935e895 BM |
465 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
466 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
467 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
468 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 469 | |
01f0a027 TG |
470 | /* Initialize timestamp */ |
471 | __touch_watchdog(); | |
bcd951cf | 472 | /* Enable the perf event */ |
146c9d0e TG |
473 | if (watchdog_enabled & NMI_WATCHDOG_ENABLED) |
474 | watchdog_nmi_enable(cpu); | |
58687acb | 475 | |
bcd951cf | 476 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 477 | } |
58687acb | 478 | |
bcd951cf TG |
479 | static void watchdog_disable(unsigned int cpu) |
480 | { | |
01f0a027 | 481 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 482 | |
bcd951cf | 483 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
484 | /* |
485 | * Disable the perf event first. That prevents that a large delay | |
486 | * between disabling the timer and disabling the perf event causes | |
487 | * the perf NMI to detect a false positive. | |
488 | */ | |
bcd951cf | 489 | watchdog_nmi_disable(cpu); |
01f0a027 | 490 | hrtimer_cancel(hrtimer); |
58687acb DZ |
491 | } |
492 | ||
b8900bc0 FW |
493 | static void watchdog_cleanup(unsigned int cpu, bool online) |
494 | { | |
495 | watchdog_disable(cpu); | |
496 | } | |
497 | ||
bcd951cf TG |
498 | static int watchdog_should_run(unsigned int cpu) |
499 | { | |
500 | return __this_cpu_read(hrtimer_interrupts) != | |
501 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
502 | } | |
503 | ||
504 | /* | |
505 | * The watchdog thread function - touches the timestamp. | |
506 | * | |
0f34c400 | 507 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
508 | * default) to reset the softlockup timestamp. If this gets delayed |
509 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
510 | * triggers in watchdog_timer_fn(). | |
511 | */ | |
512 | static void watchdog(unsigned int cpu) | |
513 | { | |
514 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
515 | __this_cpu_read(hrtimer_interrupts)); | |
516 | __touch_watchdog(); | |
517 | } | |
58687acb | 518 | |
b8900bc0 FW |
519 | static struct smp_hotplug_thread watchdog_threads = { |
520 | .store = &softlockup_watchdog, | |
521 | .thread_should_run = watchdog_should_run, | |
522 | .thread_fn = watchdog, | |
523 | .thread_comm = "watchdog/%u", | |
524 | .setup = watchdog_enable, | |
525 | .cleanup = watchdog_cleanup, | |
526 | .park = watchdog_disable, | |
527 | .unpark = watchdog_enable, | |
528 | }; | |
529 | ||
2eb2527f TG |
530 | static void softlockup_update_smpboot_threads(void) |
531 | { | |
532 | lockdep_assert_held(&watchdog_mutex); | |
533 | ||
534 | if (!softlockup_threads_initialized) | |
535 | return; | |
536 | ||
537 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
538 | &watchdog_allowed_mask); | |
2eb2527f TG |
539 | } |
540 | ||
541 | /* Temporarily park all watchdog threads */ | |
542 | static void softlockup_park_all_threads(void) | |
543 | { | |
544 | cpumask_clear(&watchdog_allowed_mask); | |
545 | softlockup_update_smpboot_threads(); | |
546 | } | |
547 | ||
e8b62b2d TG |
548 | /* Unpark enabled threads */ |
549 | static void softlockup_unpark_threads(void) | |
2eb2527f TG |
550 | { |
551 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
552 | softlockup_update_smpboot_threads(); | |
553 | } | |
554 | ||
5587185d | 555 | static void lockup_detector_reconfigure(void) |
2eb2527f | 556 | { |
e31d6883 | 557 | cpus_read_lock(); |
6b9dc480 | 558 | watchdog_nmi_stop(); |
2eb2527f TG |
559 | softlockup_park_all_threads(); |
560 | set_sample_period(); | |
09154985 TG |
561 | lockup_detector_update_enable(); |
562 | if (watchdog_enabled && watchdog_thresh) | |
e8b62b2d | 563 | softlockup_unpark_threads(); |
6b9dc480 | 564 | watchdog_nmi_start(); |
e31d6883 TG |
565 | cpus_read_unlock(); |
566 | /* | |
567 | * Must be called outside the cpus locked section to prevent | |
568 | * recursive locking in the perf code. | |
569 | */ | |
570 | __lockup_detector_cleanup(); | |
2eb2527f TG |
571 | } |
572 | ||
573 | /* | |
5587185d | 574 | * Create the watchdog thread infrastructure and configure the detector(s). |
2eb2527f TG |
575 | * |
576 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
577 | * the threads are sucessfully initialized, take the proper locks and | |
578 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
579 | */ | |
5587185d | 580 | static __init void lockup_detector_setup(void) |
2eb2527f TG |
581 | { |
582 | int ret; | |
583 | ||
584 | /* | |
585 | * If sysctl is off and watchdog got disabled on the command line, | |
586 | * nothing to do here. | |
587 | */ | |
09154985 TG |
588 | lockup_detector_update_enable(); |
589 | ||
2eb2527f TG |
590 | if (!IS_ENABLED(CONFIG_SYSCTL) && |
591 | !(watchdog_enabled && watchdog_thresh)) | |
592 | return; | |
593 | ||
594 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
595 | &watchdog_allowed_mask); | |
596 | if (ret) { | |
597 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
598 | return; | |
599 | } | |
600 | ||
601 | mutex_lock(&watchdog_mutex); | |
602 | softlockup_threads_initialized = true; | |
5587185d | 603 | lockup_detector_reconfigure(); |
2eb2527f TG |
604 | mutex_unlock(&watchdog_mutex); |
605 | } | |
606 | ||
2b9d7f23 TG |
607 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
608 | static inline int watchdog_park_threads(void) { return 0; } | |
609 | static inline void watchdog_unpark_threads(void) { } | |
610 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
611 | static inline void watchdog_disable_all_cpus(void) { } | |
5587185d | 612 | static void lockup_detector_reconfigure(void) |
6592ad2f | 613 | { |
e31d6883 | 614 | cpus_read_lock(); |
6b9dc480 | 615 | watchdog_nmi_stop(); |
09154985 | 616 | lockup_detector_update_enable(); |
6b9dc480 | 617 | watchdog_nmi_start(); |
e31d6883 | 618 | cpus_read_unlock(); |
6592ad2f | 619 | } |
5587185d | 620 | static inline void lockup_detector_setup(void) |
34ddaa3e | 621 | { |
5587185d | 622 | lockup_detector_reconfigure(); |
34ddaa3e | 623 | } |
2b9d7f23 | 624 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 625 | |
941154bd TG |
626 | static void __lockup_detector_cleanup(void) |
627 | { | |
628 | lockdep_assert_held(&watchdog_mutex); | |
629 | hardlockup_detector_perf_cleanup(); | |
630 | } | |
631 | ||
632 | /** | |
633 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
634 | * | |
635 | * Caller must not hold the cpu hotplug rwsem. | |
636 | */ | |
637 | void lockup_detector_cleanup(void) | |
638 | { | |
639 | mutex_lock(&watchdog_mutex); | |
640 | __lockup_detector_cleanup(); | |
641 | mutex_unlock(&watchdog_mutex); | |
642 | } | |
643 | ||
6554fd8c TG |
644 | /** |
645 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
646 | * | |
647 | * Special interface for parisc. It prevents lockup detector warnings from | |
648 | * the default pm_poweroff() function which busy loops forever. | |
649 | */ | |
650 | void lockup_detector_soft_poweroff(void) | |
651 | { | |
652 | watchdog_enabled = 0; | |
653 | } | |
654 | ||
58cf690a UO |
655 | #ifdef CONFIG_SYSCTL |
656 | ||
e8b62b2d | 657 | /* Propagate any changes to the watchdog threads */ |
d57108d4 | 658 | static void proc_watchdog_update(void) |
a0c9cbb9 | 659 | { |
e8b62b2d TG |
660 | /* Remove impossible cpus to keep sysctl output clean. */ |
661 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
5587185d | 662 | lockup_detector_reconfigure(); |
a0c9cbb9 UO |
663 | } |
664 | ||
ef246a21 UO |
665 | /* |
666 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
667 | * | |
7feeb9cd TG |
668 | * caller | table->data points to | 'which' |
669 | * -------------------|----------------------------|-------------------------- | |
670 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED | | |
671 | * | | SOFT_WATCHDOG_ENABLED | |
672 | * -------------------|----------------------------|-------------------------- | |
673 | * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED | |
674 | * -------------------|----------------------------|-------------------------- | |
675 | * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED | |
ef246a21 UO |
676 | */ |
677 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
678 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
679 | { | |
09154985 | 680 | int err, old, *param = table->data; |
ef246a21 | 681 | |
946d1977 | 682 | mutex_lock(&watchdog_mutex); |
ef246a21 | 683 | |
ef246a21 | 684 | if (!write) { |
09154985 TG |
685 | /* |
686 | * On read synchronize the userspace interface. This is a | |
687 | * racy snapshot. | |
688 | */ | |
689 | *param = (watchdog_enabled & which) != 0; | |
ef246a21 UO |
690 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
691 | } else { | |
09154985 | 692 | old = READ_ONCE(*param); |
ef246a21 | 693 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
09154985 | 694 | if (!err && old != READ_ONCE(*param)) |
d57108d4 | 695 | proc_watchdog_update(); |
ef246a21 | 696 | } |
946d1977 | 697 | mutex_unlock(&watchdog_mutex); |
ef246a21 UO |
698 | return err; |
699 | } | |
700 | ||
83a80a39 UO |
701 | /* |
702 | * /proc/sys/kernel/watchdog | |
703 | */ | |
704 | int proc_watchdog(struct ctl_table *table, int write, | |
705 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
706 | { | |
707 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
708 | table, write, buffer, lenp, ppos); | |
709 | } | |
710 | ||
711 | /* | |
712 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 713 | */ |
83a80a39 UO |
714 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
715 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
716 | { | |
a994a314 TG |
717 | if (!nmi_watchdog_available && write) |
718 | return -ENOTSUPP; | |
83a80a39 UO |
719 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, |
720 | table, write, buffer, lenp, ppos); | |
721 | } | |
722 | ||
723 | /* | |
724 | * /proc/sys/kernel/soft_watchdog | |
725 | */ | |
726 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
727 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
728 | { | |
729 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
730 | table, write, buffer, lenp, ppos); | |
731 | } | |
58687acb | 732 | |
83a80a39 UO |
733 | /* |
734 | * /proc/sys/kernel/watchdog_thresh | |
735 | */ | |
736 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
737 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 738 | { |
d57108d4 | 739 | int err, old; |
58687acb | 740 | |
946d1977 | 741 | mutex_lock(&watchdog_mutex); |
bcd951cf | 742 | |
d57108d4 | 743 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 744 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 745 | |
d57108d4 TG |
746 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
747 | proc_watchdog_update(); | |
e04ab2bc | 748 | |
946d1977 | 749 | mutex_unlock(&watchdog_mutex); |
b8900bc0 | 750 | return err; |
58687acb | 751 | } |
fe4ba3c3 CM |
752 | |
753 | /* | |
754 | * The cpumask is the mask of possible cpus that the watchdog can run | |
755 | * on, not the mask of cpus it is actually running on. This allows the | |
756 | * user to specify a mask that will include cpus that have not yet | |
757 | * been brought online, if desired. | |
758 | */ | |
759 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
760 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
761 | { | |
762 | int err; | |
763 | ||
946d1977 | 764 | mutex_lock(&watchdog_mutex); |
8c073d27 | 765 | |
fe4ba3c3 | 766 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 767 | if (!err && write) |
e8b62b2d | 768 | proc_watchdog_update(); |
5490125d | 769 | |
946d1977 | 770 | mutex_unlock(&watchdog_mutex); |
fe4ba3c3 CM |
771 | return err; |
772 | } | |
58687acb DZ |
773 | #endif /* CONFIG_SYSCTL */ |
774 | ||
004417a6 | 775 | void __init lockup_detector_init(void) |
58687acb | 776 | { |
13316b31 | 777 | if (tick_nohz_full_enabled()) |
314b08ff | 778 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
13316b31 FW |
779 | |
780 | cpumask_copy(&watchdog_cpumask, housekeeping_cpumask()); | |
fe4ba3c3 | 781 | |
a994a314 TG |
782 | if (!watchdog_nmi_probe()) |
783 | nmi_watchdog_available = true; | |
5587185d | 784 | lockup_detector_setup(); |
58687acb | 785 | } |