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