]>
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 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) |
35 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
36 | NMI_WATCHDOG_ENABLED; | |
84d56e66 | 37 | #else |
249e52e3 | 38 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
84d56e66 | 39 | #endif |
05a4a952 | 40 | |
7feeb9cd TG |
41 | int __read_mostly nmi_watchdog_user_enabled; |
42 | int __read_mostly soft_watchdog_user_enabled; | |
43 | int __read_mostly watchdog_user_enabled; | |
44 | int __read_mostly watchdog_thresh = 10; | |
45 | ||
46 | struct cpumask watchdog_allowed_mask __read_mostly; | |
47 | static bool softlockup_threads_initialized __read_mostly; | |
48 | ||
49 | struct cpumask watchdog_cpumask __read_mostly; | |
50 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
51 | ||
05a4a952 | 52 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
05a4a952 NP |
53 | /* |
54 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
55 | */ | |
56 | unsigned int __read_mostly hardlockup_panic = | |
57 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
58 | /* | |
59 | * We may not want to enable hard lockup detection by default in all cases, | |
60 | * for example when running the kernel as a guest on a hypervisor. In these | |
61 | * cases this function can be called to disable hard lockup detection. This | |
62 | * function should only be executed once by the boot processor before the | |
63 | * kernel command line parameters are parsed, because otherwise it is not | |
64 | * possible to override this in hardlockup_panic_setup(). | |
65 | */ | |
7a355820 | 66 | void __init hardlockup_detector_disable(void) |
05a4a952 NP |
67 | { |
68 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
69 | } | |
70 | ||
71 | static int __init hardlockup_panic_setup(char *str) | |
72 | { | |
73 | if (!strncmp(str, "panic", 5)) | |
74 | hardlockup_panic = 1; | |
75 | else if (!strncmp(str, "nopanic", 7)) | |
76 | hardlockup_panic = 0; | |
77 | else if (!strncmp(str, "0", 1)) | |
78 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
79 | else if (!strncmp(str, "1", 1)) | |
80 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
81 | return 1; | |
82 | } | |
83 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
84 | ||
368a7e2c TG |
85 | # ifdef CONFIG_SMP |
86 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 87 | |
368a7e2c TG |
88 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
89 | { | |
90 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
91 | return 1; | |
92 | } | |
93 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
94 | # endif /* CONFIG_SMP */ | |
95 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 96 | |
05a4a952 NP |
97 | /* |
98 | * These functions can be overridden if an architecture implements its | |
99 | * own hardlockup detector. | |
a10a842f NP |
100 | * |
101 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
102 | * softlockup watchdog threads start and stop. The arch must select the | |
103 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
104 | */ |
105 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
106 | { | |
107 | return 0; | |
108 | } | |
941154bd | 109 | |
05a4a952 NP |
110 | void __weak watchdog_nmi_disable(unsigned int cpu) |
111 | { | |
941154bd | 112 | hardlockup_detector_perf_disable(); |
05a4a952 NP |
113 | } |
114 | ||
a10a842f NP |
115 | /* |
116 | * watchdog_nmi_reconfigure can be implemented to be notified after any | |
117 | * watchdog configuration change. The arch hardlockup watchdog should | |
118 | * respond to the following variables: | |
7feeb9cd | 119 | * - watchdog_enabled |
a10a842f NP |
120 | * - watchdog_thresh |
121 | * - watchdog_cpumask | |
122 | * - sysctl_hardlockup_all_cpu_backtrace | |
123 | * - hardlockup_panic | |
a10a842f | 124 | */ |
2b9d7f23 | 125 | void __weak watchdog_nmi_reconfigure(void) { } |
a10a842f | 126 | |
05a4a952 NP |
127 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
128 | ||
2b9d7f23 TG |
129 | /* Global variables, exported for sysctl */ |
130 | unsigned int __read_mostly softlockup_panic = | |
131 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
2eb2527f | 132 | |
0f34c400 | 133 | static u64 __read_mostly sample_period; |
58687acb DZ |
134 | |
135 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
136 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
137 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
138 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 139 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
140 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
141 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 142 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 143 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 144 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 145 | |
58687acb DZ |
146 | static int __init softlockup_panic_setup(char *str) |
147 | { | |
148 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
149 | return 1; |
150 | } | |
151 | __setup("softlockup_panic=", softlockup_panic_setup); | |
152 | ||
153 | static int __init nowatchdog_setup(char *str) | |
154 | { | |
195daf66 | 155 | watchdog_enabled = 0; |
58687acb DZ |
156 | return 1; |
157 | } | |
158 | __setup("nowatchdog", nowatchdog_setup); | |
159 | ||
58687acb DZ |
160 | static int __init nosoftlockup_setup(char *str) |
161 | { | |
195daf66 | 162 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
163 | return 1; |
164 | } | |
165 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 166 | |
ed235875 | 167 | #ifdef CONFIG_SMP |
368a7e2c TG |
168 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
169 | ||
ed235875 AT |
170 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
171 | { | |
368a7e2c | 172 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
173 | return 1; |
174 | } | |
175 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 176 | #endif |
58687acb | 177 | |
941154bd TG |
178 | static void __lockup_detector_cleanup(void); |
179 | ||
4eec42f3 MSB |
180 | /* |
181 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
182 | * lockups can have false positives under extreme conditions. So we generally | |
183 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
184 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
185 | * time the hard threshold is. | |
186 | */ | |
6e9101ae | 187 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
188 | { |
189 | return watchdog_thresh * 2; | |
190 | } | |
58687acb DZ |
191 | |
192 | /* | |
193 | * Returns seconds, approximately. We don't need nanosecond | |
194 | * resolution, and we don't need to waste time with a big divide when | |
195 | * 2^30ns == 1.074s. | |
196 | */ | |
c06b4f19 | 197 | static unsigned long get_timestamp(void) |
58687acb | 198 | { |
545a2bf7 | 199 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
200 | } |
201 | ||
0f34c400 | 202 | static void set_sample_period(void) |
58687acb DZ |
203 | { |
204 | /* | |
586692a5 | 205 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
206 | * the divide by 5 is to give hrtimer several chances (two |
207 | * or three with the current relation between the soft | |
208 | * and hard thresholds) to increment before the | |
209 | * hardlockup detector generates a warning | |
58687acb | 210 | */ |
0f34c400 | 211 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 212 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
213 | } |
214 | ||
215 | /* Commands for resetting the watchdog */ | |
216 | static void __touch_watchdog(void) | |
217 | { | |
c06b4f19 | 218 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
219 | } |
220 | ||
03e0d461 TH |
221 | /** |
222 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
223 | * | |
224 | * Call when the scheduler may have stalled for legitimate reasons | |
225 | * preventing the watchdog task from executing - e.g. the scheduler | |
226 | * entering idle state. This should only be used for scheduler events. | |
227 | * Use touch_softlockup_watchdog() for everything else. | |
228 | */ | |
229 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 230 | { |
7861144b AM |
231 | /* |
232 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
233 | * gets zeroed here, so use the raw_ operation. | |
234 | */ | |
235 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 236 | } |
03e0d461 TH |
237 | |
238 | void touch_softlockup_watchdog(void) | |
239 | { | |
240 | touch_softlockup_watchdog_sched(); | |
82607adc | 241 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 242 | } |
0167c781 | 243 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 244 | |
332fbdbc | 245 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
246 | { |
247 | int cpu; | |
248 | ||
249 | /* | |
d57108d4 TG |
250 | * watchdog_mutex cannpt be taken here, as this might be called |
251 | * from (soft)interrupt context, so the access to | |
252 | * watchdog_allowed_cpumask might race with a concurrent update. | |
253 | * | |
254 | * The watchdog time stamp can race against a concurrent real | |
255 | * update as well, the only side effect might be a cycle delay for | |
256 | * the softlockup check. | |
58687acb | 257 | */ |
d57108d4 | 258 | for_each_cpu(cpu, &watchdog_allowed_mask) |
58687acb | 259 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 260 | wq_watchdog_touch(-1); |
58687acb DZ |
261 | } |
262 | ||
58687acb DZ |
263 | void touch_softlockup_watchdog_sync(void) |
264 | { | |
f7f66b05 CL |
265 | __this_cpu_write(softlockup_touch_sync, true); |
266 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
267 | } |
268 | ||
26e09c6e | 269 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 270 | { |
c06b4f19 | 271 | unsigned long now = get_timestamp(); |
58687acb | 272 | |
39d2da21 | 273 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
274 | /* Warn about unreasonable delays. */ |
275 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
276 | return now - touch_ts; | |
277 | } | |
58687acb DZ |
278 | return 0; |
279 | } | |
280 | ||
05a4a952 NP |
281 | /* watchdog detector functions */ |
282 | bool is_hardlockup(void) | |
58687acb | 283 | { |
05a4a952 | 284 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 285 | |
05a4a952 NP |
286 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
287 | return true; | |
288 | ||
289 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
290 | return false; | |
73ce0511 | 291 | } |
05a4a952 NP |
292 | |
293 | static void watchdog_interrupt_count(void) | |
73ce0511 | 294 | { |
05a4a952 | 295 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 296 | } |
58687acb DZ |
297 | |
298 | /* watchdog kicker functions */ | |
299 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
300 | { | |
909ea964 | 301 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
302 | struct pt_regs *regs = get_irq_regs(); |
303 | int duration; | |
ed235875 | 304 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 305 | |
01f0a027 | 306 | if (!watchdog_enabled) |
b94f5118 DZ |
307 | return HRTIMER_NORESTART; |
308 | ||
58687acb DZ |
309 | /* kick the hardlockup detector */ |
310 | watchdog_interrupt_count(); | |
311 | ||
312 | /* kick the softlockup detector */ | |
909ea964 | 313 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
314 | |
315 | /* .. and repeat */ | |
0f34c400 | 316 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
317 | |
318 | if (touch_ts == 0) { | |
909ea964 | 319 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
320 | /* |
321 | * If the time stamp was touched atomically | |
322 | * make sure the scheduler tick is up to date. | |
323 | */ | |
909ea964 | 324 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
325 | sched_clock_tick(); |
326 | } | |
5d1c0f4a EM |
327 | |
328 | /* Clear the guest paused flag on watchdog reset */ | |
329 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
330 | __touch_watchdog(); |
331 | return HRTIMER_RESTART; | |
332 | } | |
333 | ||
334 | /* check for a softlockup | |
335 | * This is done by making sure a high priority task is | |
336 | * being scheduled. The task touches the watchdog to | |
337 | * indicate it is getting cpu time. If it hasn't then | |
338 | * this is a good indication some task is hogging the cpu | |
339 | */ | |
26e09c6e | 340 | duration = is_softlockup(touch_ts); |
58687acb | 341 | if (unlikely(duration)) { |
5d1c0f4a EM |
342 | /* |
343 | * If a virtual machine is stopped by the host it can look to | |
344 | * the watchdog like a soft lockup, check to see if the host | |
345 | * stopped the vm before we issue the warning | |
346 | */ | |
347 | if (kvm_check_and_clear_guest_paused()) | |
348 | return HRTIMER_RESTART; | |
349 | ||
58687acb | 350 | /* only warn once */ |
b1a8de1f | 351 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
352 | /* | |
353 | * When multiple processes are causing softlockups the | |
354 | * softlockup detector only warns on the first one | |
355 | * because the code relies on a full quiet cycle to | |
356 | * re-arm. The second process prevents the quiet cycle | |
357 | * and never gets reported. Use task pointers to detect | |
358 | * this. | |
359 | */ | |
360 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
361 | current) { | |
362 | __this_cpu_write(soft_watchdog_warn, false); | |
363 | __touch_watchdog(); | |
364 | } | |
58687acb | 365 | return HRTIMER_RESTART; |
b1a8de1f | 366 | } |
58687acb | 367 | |
ed235875 AT |
368 | if (softlockup_all_cpu_backtrace) { |
369 | /* Prevent multiple soft-lockup reports if one cpu is already | |
370 | * engaged in dumping cpu back traces | |
371 | */ | |
372 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
373 | /* Someone else will report us. Let's give up */ | |
374 | __this_cpu_write(soft_watchdog_warn, true); | |
375 | return HRTIMER_RESTART; | |
376 | } | |
377 | } | |
378 | ||
656c3b79 | 379 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 380 | smp_processor_id(), duration, |
58687acb | 381 | current->comm, task_pid_nr(current)); |
b1a8de1f | 382 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
383 | print_modules(); |
384 | print_irqtrace_events(current); | |
385 | if (regs) | |
386 | show_regs(regs); | |
387 | else | |
388 | dump_stack(); | |
389 | ||
ed235875 AT |
390 | if (softlockup_all_cpu_backtrace) { |
391 | /* Avoid generating two back traces for current | |
392 | * given that one is already made above | |
393 | */ | |
394 | trigger_allbutself_cpu_backtrace(); | |
395 | ||
396 | clear_bit(0, &soft_lockup_nmi_warn); | |
397 | /* Barrier to sync with other cpus */ | |
398 | smp_mb__after_atomic(); | |
399 | } | |
400 | ||
69361eef | 401 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
402 | if (softlockup_panic) |
403 | panic("softlockup: hung tasks"); | |
909ea964 | 404 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 405 | } else |
909ea964 | 406 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
407 | |
408 | return HRTIMER_RESTART; | |
409 | } | |
410 | ||
bcd951cf TG |
411 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
412 | { | |
413 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 414 | |
bcd951cf TG |
415 | sched_setscheduler(current, policy, ¶m); |
416 | } | |
417 | ||
418 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 419 | { |
01f0a027 | 420 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 421 | |
01f0a027 TG |
422 | /* |
423 | * Start the timer first to prevent the NMI watchdog triggering | |
424 | * before the timer has a chance to fire. | |
425 | */ | |
3935e895 BM |
426 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
427 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
428 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
429 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 430 | |
01f0a027 TG |
431 | /* Initialize timestamp */ |
432 | __touch_watchdog(); | |
bcd951cf TG |
433 | /* Enable the perf event */ |
434 | watchdog_nmi_enable(cpu); | |
58687acb | 435 | |
bcd951cf | 436 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 437 | } |
58687acb | 438 | |
bcd951cf TG |
439 | static void watchdog_disable(unsigned int cpu) |
440 | { | |
01f0a027 | 441 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 442 | |
bcd951cf | 443 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
444 | /* |
445 | * Disable the perf event first. That prevents that a large delay | |
446 | * between disabling the timer and disabling the perf event causes | |
447 | * the perf NMI to detect a false positive. | |
448 | */ | |
bcd951cf | 449 | watchdog_nmi_disable(cpu); |
01f0a027 | 450 | hrtimer_cancel(hrtimer); |
58687acb DZ |
451 | } |
452 | ||
b8900bc0 FW |
453 | static void watchdog_cleanup(unsigned int cpu, bool online) |
454 | { | |
455 | watchdog_disable(cpu); | |
456 | } | |
457 | ||
bcd951cf TG |
458 | static int watchdog_should_run(unsigned int cpu) |
459 | { | |
460 | return __this_cpu_read(hrtimer_interrupts) != | |
461 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
462 | } | |
463 | ||
464 | /* | |
465 | * The watchdog thread function - touches the timestamp. | |
466 | * | |
0f34c400 | 467 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
468 | * default) to reset the softlockup timestamp. If this gets delayed |
469 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
470 | * triggers in watchdog_timer_fn(). | |
471 | */ | |
472 | static void watchdog(unsigned int cpu) | |
473 | { | |
474 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
475 | __this_cpu_read(hrtimer_interrupts)); | |
476 | __touch_watchdog(); | |
477 | } | |
58687acb | 478 | |
b8900bc0 FW |
479 | static struct smp_hotplug_thread watchdog_threads = { |
480 | .store = &softlockup_watchdog, | |
481 | .thread_should_run = watchdog_should_run, | |
482 | .thread_fn = watchdog, | |
483 | .thread_comm = "watchdog/%u", | |
484 | .setup = watchdog_enable, | |
485 | .cleanup = watchdog_cleanup, | |
486 | .park = watchdog_disable, | |
487 | .unpark = watchdog_enable, | |
488 | }; | |
489 | ||
2eb2527f TG |
490 | static void softlockup_update_smpboot_threads(void) |
491 | { | |
492 | lockdep_assert_held(&watchdog_mutex); | |
493 | ||
494 | if (!softlockup_threads_initialized) | |
495 | return; | |
496 | ||
497 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
498 | &watchdog_allowed_mask); | |
499 | __lockup_detector_cleanup(); | |
500 | } | |
501 | ||
502 | /* Temporarily park all watchdog threads */ | |
503 | static void softlockup_park_all_threads(void) | |
504 | { | |
505 | cpumask_clear(&watchdog_allowed_mask); | |
506 | softlockup_update_smpboot_threads(); | |
507 | } | |
508 | ||
e8b62b2d TG |
509 | /* Unpark enabled threads */ |
510 | static void softlockup_unpark_threads(void) | |
2eb2527f TG |
511 | { |
512 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
513 | softlockup_update_smpboot_threads(); | |
514 | } | |
515 | ||
516 | static void softlockup_reconfigure_threads(bool enabled) | |
517 | { | |
518 | softlockup_park_all_threads(); | |
519 | set_sample_period(); | |
520 | if (enabled) | |
e8b62b2d | 521 | softlockup_unpark_threads(); |
2eb2527f TG |
522 | } |
523 | ||
524 | /* | |
525 | * Create the watchdog thread infrastructure. | |
526 | * | |
527 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
528 | * the threads are sucessfully initialized, take the proper locks and | |
529 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
530 | */ | |
531 | static __init void softlockup_init_threads(void) | |
532 | { | |
533 | int ret; | |
534 | ||
535 | /* | |
536 | * If sysctl is off and watchdog got disabled on the command line, | |
537 | * nothing to do here. | |
538 | */ | |
539 | if (!IS_ENABLED(CONFIG_SYSCTL) && | |
540 | !(watchdog_enabled && watchdog_thresh)) | |
541 | return; | |
542 | ||
543 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
544 | &watchdog_allowed_mask); | |
545 | if (ret) { | |
546 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
547 | return; | |
548 | } | |
549 | ||
550 | mutex_lock(&watchdog_mutex); | |
551 | softlockup_threads_initialized = true; | |
552 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); | |
553 | mutex_unlock(&watchdog_mutex); | |
554 | } | |
555 | ||
2b9d7f23 TG |
556 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
557 | static inline int watchdog_park_threads(void) { return 0; } | |
558 | static inline void watchdog_unpark_threads(void) { } | |
559 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
560 | static inline void watchdog_disable_all_cpus(void) { } | |
2eb2527f | 561 | static inline void softlockup_init_threads(void) { } |
2eb2527f | 562 | static inline void softlockup_reconfigure_threads(bool enabled) { } |
2b9d7f23 | 563 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 564 | |
941154bd TG |
565 | static void __lockup_detector_cleanup(void) |
566 | { | |
567 | lockdep_assert_held(&watchdog_mutex); | |
568 | hardlockup_detector_perf_cleanup(); | |
569 | } | |
570 | ||
571 | /** | |
572 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
573 | * | |
574 | * Caller must not hold the cpu hotplug rwsem. | |
575 | */ | |
576 | void lockup_detector_cleanup(void) | |
577 | { | |
578 | mutex_lock(&watchdog_mutex); | |
579 | __lockup_detector_cleanup(); | |
580 | mutex_unlock(&watchdog_mutex); | |
581 | } | |
582 | ||
6554fd8c TG |
583 | /** |
584 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
585 | * | |
586 | * Special interface for parisc. It prevents lockup detector warnings from | |
587 | * the default pm_poweroff() function which busy loops forever. | |
588 | */ | |
589 | void lockup_detector_soft_poweroff(void) | |
590 | { | |
591 | watchdog_enabled = 0; | |
592 | } | |
593 | ||
58cf690a UO |
594 | #ifdef CONFIG_SYSCTL |
595 | ||
e8b62b2d | 596 | /* Propagate any changes to the watchdog threads */ |
d57108d4 | 597 | static void proc_watchdog_update(void) |
a0c9cbb9 | 598 | { |
e8b62b2d TG |
599 | /* Remove impossible cpus to keep sysctl output clean. */ |
600 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
d57108d4 | 601 | softlockup_reconfigure_threads(watchdog_enabled && watchdog_thresh); |
a10a842f | 602 | watchdog_nmi_reconfigure(); |
a0c9cbb9 UO |
603 | } |
604 | ||
ef246a21 UO |
605 | /* |
606 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
607 | * | |
7feeb9cd TG |
608 | * caller | table->data points to | 'which' |
609 | * -------------------|----------------------------|-------------------------- | |
610 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED | | |
611 | * | | SOFT_WATCHDOG_ENABLED | |
612 | * -------------------|----------------------------|-------------------------- | |
613 | * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED | |
614 | * -------------------|----------------------------|-------------------------- | |
615 | * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED | |
ef246a21 UO |
616 | */ |
617 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
618 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
619 | { | |
620 | int err, old, new; | |
621 | int *watchdog_param = (int *)table->data; | |
622 | ||
b7a34981 | 623 | cpu_hotplug_disable(); |
946d1977 | 624 | mutex_lock(&watchdog_mutex); |
ef246a21 UO |
625 | |
626 | /* | |
627 | * If the parameter is being read return the state of the corresponding | |
628 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
629 | * run state of the lockup detectors. | |
630 | */ | |
631 | if (!write) { | |
632 | *watchdog_param = (watchdog_enabled & which) != 0; | |
633 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
634 | } else { | |
635 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
636 | if (err) | |
637 | goto out; | |
638 | ||
639 | /* | |
640 | * There is a race window between fetching the current value | |
641 | * from 'watchdog_enabled' and storing the new value. During | |
642 | * this race window, watchdog_nmi_enable() can sneak in and | |
643 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
644 | * The 'cmpxchg' detects this race and the loop retries. | |
645 | */ | |
646 | do { | |
647 | old = watchdog_enabled; | |
648 | /* | |
649 | * If the parameter value is not zero set the | |
650 | * corresponding bit(s), else clear it(them). | |
651 | */ | |
652 | if (*watchdog_param) | |
653 | new = old | which; | |
654 | else | |
655 | new = old & ~which; | |
656 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
657 | ||
d57108d4 TG |
658 | if (old != new) |
659 | proc_watchdog_update(); | |
ef246a21 UO |
660 | } |
661 | out: | |
946d1977 | 662 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 663 | cpu_hotplug_enable(); |
ef246a21 UO |
664 | return err; |
665 | } | |
666 | ||
83a80a39 UO |
667 | /* |
668 | * /proc/sys/kernel/watchdog | |
669 | */ | |
670 | int proc_watchdog(struct ctl_table *table, int write, | |
671 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
672 | { | |
673 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
674 | table, write, buffer, lenp, ppos); | |
675 | } | |
676 | ||
677 | /* | |
678 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 679 | */ |
83a80a39 UO |
680 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
681 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
682 | { | |
683 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
684 | table, write, buffer, lenp, ppos); | |
685 | } | |
686 | ||
687 | /* | |
688 | * /proc/sys/kernel/soft_watchdog | |
689 | */ | |
690 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
691 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
692 | { | |
693 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
694 | table, write, buffer, lenp, ppos); | |
695 | } | |
58687acb | 696 | |
83a80a39 UO |
697 | /* |
698 | * /proc/sys/kernel/watchdog_thresh | |
699 | */ | |
700 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
701 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 702 | { |
d57108d4 | 703 | int err, old; |
58687acb | 704 | |
b7a34981 | 705 | cpu_hotplug_disable(); |
946d1977 | 706 | mutex_lock(&watchdog_mutex); |
bcd951cf | 707 | |
d57108d4 | 708 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 709 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 710 | |
d57108d4 TG |
711 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
712 | proc_watchdog_update(); | |
e04ab2bc | 713 | |
946d1977 | 714 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 715 | cpu_hotplug_enable(); |
b8900bc0 | 716 | return err; |
58687acb | 717 | } |
fe4ba3c3 CM |
718 | |
719 | /* | |
720 | * The cpumask is the mask of possible cpus that the watchdog can run | |
721 | * on, not the mask of cpus it is actually running on. This allows the | |
722 | * user to specify a mask that will include cpus that have not yet | |
723 | * been brought online, if desired. | |
724 | */ | |
725 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
726 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
727 | { | |
728 | int err; | |
729 | ||
b7a34981 | 730 | cpu_hotplug_disable(); |
946d1977 | 731 | mutex_lock(&watchdog_mutex); |
8c073d27 | 732 | |
fe4ba3c3 | 733 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 734 | if (!err && write) |
e8b62b2d | 735 | proc_watchdog_update(); |
5490125d | 736 | |
946d1977 | 737 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 738 | cpu_hotplug_enable(); |
fe4ba3c3 CM |
739 | return err; |
740 | } | |
58687acb DZ |
741 | #endif /* CONFIG_SYSCTL */ |
742 | ||
004417a6 | 743 | void __init lockup_detector_init(void) |
58687acb | 744 | { |
fe4ba3c3 CM |
745 | #ifdef CONFIG_NO_HZ_FULL |
746 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
747 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
748 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
749 | } else |
750 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
751 | #else | |
752 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
753 | #endif | |
754 | ||
d57108d4 | 755 | softlockup_init_threads(); |
58687acb | 756 | } |