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 TG |
46 | int __read_mostly watchdog_thresh = 10; |
47 | ||
48 | struct cpumask watchdog_allowed_mask __read_mostly; | |
49 | static bool softlockup_threads_initialized __read_mostly; | |
50 | ||
51 | struct cpumask watchdog_cpumask __read_mostly; | |
52 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
53 | ||
05a4a952 | 54 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
05a4a952 NP |
55 | /* |
56 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
57 | */ | |
58 | unsigned int __read_mostly hardlockup_panic = | |
59 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
60 | /* | |
61 | * We may not want to enable hard lockup detection by default in all cases, | |
62 | * for example when running the kernel as a guest on a hypervisor. In these | |
63 | * cases this function can be called to disable hard lockup detection. This | |
64 | * function should only be executed once by the boot processor before the | |
65 | * kernel command line parameters are parsed, because otherwise it is not | |
66 | * possible to override this in hardlockup_panic_setup(). | |
67 | */ | |
7a355820 | 68 | void __init hardlockup_detector_disable(void) |
05a4a952 | 69 | { |
09154985 | 70 | nmi_watchdog_user_enabled = 0; |
05a4a952 NP |
71 | } |
72 | ||
73 | static int __init hardlockup_panic_setup(char *str) | |
74 | { | |
75 | if (!strncmp(str, "panic", 5)) | |
76 | hardlockup_panic = 1; | |
77 | else if (!strncmp(str, "nopanic", 7)) | |
78 | hardlockup_panic = 0; | |
79 | else if (!strncmp(str, "0", 1)) | |
09154985 | 80 | nmi_watchdog_user_enabled = 0; |
05a4a952 | 81 | else if (!strncmp(str, "1", 1)) |
09154985 | 82 | nmi_watchdog_user_enabled = 1; |
05a4a952 NP |
83 | return 1; |
84 | } | |
85 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
86 | ||
368a7e2c TG |
87 | # ifdef CONFIG_SMP |
88 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 89 | |
368a7e2c TG |
90 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
91 | { | |
92 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
93 | return 1; | |
94 | } | |
95 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
96 | # endif /* CONFIG_SMP */ | |
97 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 98 | |
05a4a952 NP |
99 | /* |
100 | * These functions can be overridden if an architecture implements its | |
101 | * own hardlockup detector. | |
a10a842f NP |
102 | * |
103 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
104 | * softlockup watchdog threads start and stop. The arch must select the | |
105 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
106 | */ |
107 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
108 | { | |
109 | return 0; | |
110 | } | |
941154bd | 111 | |
05a4a952 NP |
112 | void __weak watchdog_nmi_disable(unsigned int cpu) |
113 | { | |
941154bd | 114 | hardlockup_detector_perf_disable(); |
05a4a952 NP |
115 | } |
116 | ||
6592ad2f TG |
117 | /** |
118 | * watchdog_nmi_reconfigure - Optional function to reconfigure NMI watchdogs | |
119 | * @run: If false stop the watchdogs on all enabled CPUs | |
120 | * If true start the watchdogs on all enabled CPUs | |
121 | * | |
122 | * The core call order is: | |
123 | * watchdog_nmi_reconfigure(false); | |
124 | * update_variables(); | |
125 | * watchdog_nmi_reconfigure(true); | |
126 | * | |
127 | * The second call which starts the watchdogs again guarantees that the | |
128 | * following variables are stable across the call. | |
7feeb9cd | 129 | * - watchdog_enabled |
a10a842f NP |
130 | * - watchdog_thresh |
131 | * - watchdog_cpumask | |
6592ad2f TG |
132 | * |
133 | * After the call the variables can be changed again. | |
a10a842f | 134 | */ |
6592ad2f | 135 | void __weak watchdog_nmi_reconfigure(bool run) { } |
a10a842f | 136 | |
09154985 TG |
137 | /** |
138 | * lockup_detector_update_enable - Update the sysctl enable bit | |
139 | * | |
140 | * Caller needs to make sure that the NMI/perf watchdogs are off, so this | |
141 | * can't race with watchdog_nmi_disable(). | |
142 | */ | |
143 | static void lockup_detector_update_enable(void) | |
144 | { | |
145 | watchdog_enabled = 0; | |
146 | if (!watchdog_user_enabled) | |
147 | return; | |
148 | if (nmi_watchdog_user_enabled) | |
149 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
150 | if (soft_watchdog_user_enabled) | |
151 | watchdog_enabled |= SOFT_WATCHDOG_ENABLED; | |
152 | } | |
153 | ||
05a4a952 NP |
154 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
155 | ||
2b9d7f23 TG |
156 | /* Global variables, exported for sysctl */ |
157 | unsigned int __read_mostly softlockup_panic = | |
158 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
2eb2527f | 159 | |
0f34c400 | 160 | static u64 __read_mostly sample_period; |
58687acb DZ |
161 | |
162 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
163 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
164 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
165 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 166 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
167 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
168 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 169 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 170 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 171 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 172 | |
58687acb DZ |
173 | static int __init softlockup_panic_setup(char *str) |
174 | { | |
175 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
176 | return 1; |
177 | } | |
178 | __setup("softlockup_panic=", softlockup_panic_setup); | |
179 | ||
180 | static int __init nowatchdog_setup(char *str) | |
181 | { | |
09154985 | 182 | watchdog_user_enabled = 0; |
58687acb DZ |
183 | return 1; |
184 | } | |
185 | __setup("nowatchdog", nowatchdog_setup); | |
186 | ||
58687acb DZ |
187 | static int __init nosoftlockup_setup(char *str) |
188 | { | |
09154985 | 189 | soft_watchdog_user_enabled = 0; |
58687acb DZ |
190 | return 1; |
191 | } | |
192 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 193 | |
ed235875 | 194 | #ifdef CONFIG_SMP |
368a7e2c TG |
195 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
196 | ||
ed235875 AT |
197 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
198 | { | |
368a7e2c | 199 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
200 | return 1; |
201 | } | |
202 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 203 | #endif |
58687acb | 204 | |
941154bd TG |
205 | static void __lockup_detector_cleanup(void); |
206 | ||
4eec42f3 MSB |
207 | /* |
208 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
209 | * lockups can have false positives under extreme conditions. So we generally | |
210 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
211 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
212 | * time the hard threshold is. | |
213 | */ | |
6e9101ae | 214 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
215 | { |
216 | return watchdog_thresh * 2; | |
217 | } | |
58687acb DZ |
218 | |
219 | /* | |
220 | * Returns seconds, approximately. We don't need nanosecond | |
221 | * resolution, and we don't need to waste time with a big divide when | |
222 | * 2^30ns == 1.074s. | |
223 | */ | |
c06b4f19 | 224 | static unsigned long get_timestamp(void) |
58687acb | 225 | { |
545a2bf7 | 226 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
227 | } |
228 | ||
0f34c400 | 229 | static void set_sample_period(void) |
58687acb DZ |
230 | { |
231 | /* | |
586692a5 | 232 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
233 | * the divide by 5 is to give hrtimer several chances (two |
234 | * or three with the current relation between the soft | |
235 | * and hard thresholds) to increment before the | |
236 | * hardlockup detector generates a warning | |
58687acb | 237 | */ |
0f34c400 | 238 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 239 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
240 | } |
241 | ||
242 | /* Commands for resetting the watchdog */ | |
243 | static void __touch_watchdog(void) | |
244 | { | |
c06b4f19 | 245 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
246 | } |
247 | ||
03e0d461 TH |
248 | /** |
249 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
250 | * | |
251 | * Call when the scheduler may have stalled for legitimate reasons | |
252 | * preventing the watchdog task from executing - e.g. the scheduler | |
253 | * entering idle state. This should only be used for scheduler events. | |
254 | * Use touch_softlockup_watchdog() for everything else. | |
255 | */ | |
256 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 257 | { |
7861144b AM |
258 | /* |
259 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
260 | * gets zeroed here, so use the raw_ operation. | |
261 | */ | |
262 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 263 | } |
03e0d461 TH |
264 | |
265 | void touch_softlockup_watchdog(void) | |
266 | { | |
267 | touch_softlockup_watchdog_sched(); | |
82607adc | 268 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 269 | } |
0167c781 | 270 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 271 | |
332fbdbc | 272 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
273 | { |
274 | int cpu; | |
275 | ||
276 | /* | |
d57108d4 TG |
277 | * watchdog_mutex cannpt be taken here, as this might be called |
278 | * from (soft)interrupt context, so the access to | |
279 | * watchdog_allowed_cpumask might race with a concurrent update. | |
280 | * | |
281 | * The watchdog time stamp can race against a concurrent real | |
282 | * update as well, the only side effect might be a cycle delay for | |
283 | * the softlockup check. | |
58687acb | 284 | */ |
d57108d4 | 285 | for_each_cpu(cpu, &watchdog_allowed_mask) |
58687acb | 286 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 287 | wq_watchdog_touch(-1); |
58687acb DZ |
288 | } |
289 | ||
58687acb DZ |
290 | void touch_softlockup_watchdog_sync(void) |
291 | { | |
f7f66b05 CL |
292 | __this_cpu_write(softlockup_touch_sync, true); |
293 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
294 | } |
295 | ||
26e09c6e | 296 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 297 | { |
c06b4f19 | 298 | unsigned long now = get_timestamp(); |
58687acb | 299 | |
39d2da21 | 300 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
301 | /* Warn about unreasonable delays. */ |
302 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
303 | return now - touch_ts; | |
304 | } | |
58687acb DZ |
305 | return 0; |
306 | } | |
307 | ||
05a4a952 NP |
308 | /* watchdog detector functions */ |
309 | bool is_hardlockup(void) | |
58687acb | 310 | { |
05a4a952 | 311 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 312 | |
05a4a952 NP |
313 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
314 | return true; | |
315 | ||
316 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
317 | return false; | |
73ce0511 | 318 | } |
05a4a952 NP |
319 | |
320 | static void watchdog_interrupt_count(void) | |
73ce0511 | 321 | { |
05a4a952 | 322 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 323 | } |
58687acb DZ |
324 | |
325 | /* watchdog kicker functions */ | |
326 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
327 | { | |
909ea964 | 328 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
329 | struct pt_regs *regs = get_irq_regs(); |
330 | int duration; | |
ed235875 | 331 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 332 | |
01f0a027 | 333 | if (!watchdog_enabled) |
b94f5118 DZ |
334 | return HRTIMER_NORESTART; |
335 | ||
58687acb DZ |
336 | /* kick the hardlockup detector */ |
337 | watchdog_interrupt_count(); | |
338 | ||
339 | /* kick the softlockup detector */ | |
909ea964 | 340 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
341 | |
342 | /* .. and repeat */ | |
0f34c400 | 343 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
344 | |
345 | if (touch_ts == 0) { | |
909ea964 | 346 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
347 | /* |
348 | * If the time stamp was touched atomically | |
349 | * make sure the scheduler tick is up to date. | |
350 | */ | |
909ea964 | 351 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
352 | sched_clock_tick(); |
353 | } | |
5d1c0f4a EM |
354 | |
355 | /* Clear the guest paused flag on watchdog reset */ | |
356 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
357 | __touch_watchdog(); |
358 | return HRTIMER_RESTART; | |
359 | } | |
360 | ||
361 | /* check for a softlockup | |
362 | * This is done by making sure a high priority task is | |
363 | * being scheduled. The task touches the watchdog to | |
364 | * indicate it is getting cpu time. If it hasn't then | |
365 | * this is a good indication some task is hogging the cpu | |
366 | */ | |
26e09c6e | 367 | duration = is_softlockup(touch_ts); |
58687acb | 368 | if (unlikely(duration)) { |
5d1c0f4a EM |
369 | /* |
370 | * If a virtual machine is stopped by the host it can look to | |
371 | * the watchdog like a soft lockup, check to see if the host | |
372 | * stopped the vm before we issue the warning | |
373 | */ | |
374 | if (kvm_check_and_clear_guest_paused()) | |
375 | return HRTIMER_RESTART; | |
376 | ||
58687acb | 377 | /* only warn once */ |
b1a8de1f | 378 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
379 | /* | |
380 | * When multiple processes are causing softlockups the | |
381 | * softlockup detector only warns on the first one | |
382 | * because the code relies on a full quiet cycle to | |
383 | * re-arm. The second process prevents the quiet cycle | |
384 | * and never gets reported. Use task pointers to detect | |
385 | * this. | |
386 | */ | |
387 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
388 | current) { | |
389 | __this_cpu_write(soft_watchdog_warn, false); | |
390 | __touch_watchdog(); | |
391 | } | |
58687acb | 392 | return HRTIMER_RESTART; |
b1a8de1f | 393 | } |
58687acb | 394 | |
ed235875 AT |
395 | if (softlockup_all_cpu_backtrace) { |
396 | /* Prevent multiple soft-lockup reports if one cpu is already | |
397 | * engaged in dumping cpu back traces | |
398 | */ | |
399 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
400 | /* Someone else will report us. Let's give up */ | |
401 | __this_cpu_write(soft_watchdog_warn, true); | |
402 | return HRTIMER_RESTART; | |
403 | } | |
404 | } | |
405 | ||
656c3b79 | 406 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 407 | smp_processor_id(), duration, |
58687acb | 408 | current->comm, task_pid_nr(current)); |
b1a8de1f | 409 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
410 | print_modules(); |
411 | print_irqtrace_events(current); | |
412 | if (regs) | |
413 | show_regs(regs); | |
414 | else | |
415 | dump_stack(); | |
416 | ||
ed235875 AT |
417 | if (softlockup_all_cpu_backtrace) { |
418 | /* Avoid generating two back traces for current | |
419 | * given that one is already made above | |
420 | */ | |
421 | trigger_allbutself_cpu_backtrace(); | |
422 | ||
423 | clear_bit(0, &soft_lockup_nmi_warn); | |
424 | /* Barrier to sync with other cpus */ | |
425 | smp_mb__after_atomic(); | |
426 | } | |
427 | ||
69361eef | 428 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
429 | if (softlockup_panic) |
430 | panic("softlockup: hung tasks"); | |
909ea964 | 431 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 432 | } else |
909ea964 | 433 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
434 | |
435 | return HRTIMER_RESTART; | |
436 | } | |
437 | ||
bcd951cf TG |
438 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
439 | { | |
440 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 441 | |
bcd951cf TG |
442 | sched_setscheduler(current, policy, ¶m); |
443 | } | |
444 | ||
445 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 446 | { |
01f0a027 | 447 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 448 | |
01f0a027 TG |
449 | /* |
450 | * Start the timer first to prevent the NMI watchdog triggering | |
451 | * before the timer has a chance to fire. | |
452 | */ | |
3935e895 BM |
453 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
454 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
455 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
456 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 457 | |
01f0a027 TG |
458 | /* Initialize timestamp */ |
459 | __touch_watchdog(); | |
bcd951cf TG |
460 | /* Enable the perf event */ |
461 | watchdog_nmi_enable(cpu); | |
58687acb | 462 | |
bcd951cf | 463 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 464 | } |
58687acb | 465 | |
bcd951cf TG |
466 | static void watchdog_disable(unsigned int cpu) |
467 | { | |
01f0a027 | 468 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 469 | |
bcd951cf | 470 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
471 | /* |
472 | * Disable the perf event first. That prevents that a large delay | |
473 | * between disabling the timer and disabling the perf event causes | |
474 | * the perf NMI to detect a false positive. | |
475 | */ | |
bcd951cf | 476 | watchdog_nmi_disable(cpu); |
01f0a027 | 477 | hrtimer_cancel(hrtimer); |
58687acb DZ |
478 | } |
479 | ||
b8900bc0 FW |
480 | static void watchdog_cleanup(unsigned int cpu, bool online) |
481 | { | |
482 | watchdog_disable(cpu); | |
483 | } | |
484 | ||
bcd951cf TG |
485 | static int watchdog_should_run(unsigned int cpu) |
486 | { | |
487 | return __this_cpu_read(hrtimer_interrupts) != | |
488 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
489 | } | |
490 | ||
491 | /* | |
492 | * The watchdog thread function - touches the timestamp. | |
493 | * | |
0f34c400 | 494 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
495 | * default) to reset the softlockup timestamp. If this gets delayed |
496 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
497 | * triggers in watchdog_timer_fn(). | |
498 | */ | |
499 | static void watchdog(unsigned int cpu) | |
500 | { | |
501 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
502 | __this_cpu_read(hrtimer_interrupts)); | |
503 | __touch_watchdog(); | |
504 | } | |
58687acb | 505 | |
b8900bc0 FW |
506 | static struct smp_hotplug_thread watchdog_threads = { |
507 | .store = &softlockup_watchdog, | |
508 | .thread_should_run = watchdog_should_run, | |
509 | .thread_fn = watchdog, | |
510 | .thread_comm = "watchdog/%u", | |
511 | .setup = watchdog_enable, | |
512 | .cleanup = watchdog_cleanup, | |
513 | .park = watchdog_disable, | |
514 | .unpark = watchdog_enable, | |
515 | }; | |
516 | ||
2eb2527f TG |
517 | static void softlockup_update_smpboot_threads(void) |
518 | { | |
519 | lockdep_assert_held(&watchdog_mutex); | |
520 | ||
521 | if (!softlockup_threads_initialized) | |
522 | return; | |
523 | ||
524 | smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
525 | &watchdog_allowed_mask); | |
526 | __lockup_detector_cleanup(); | |
527 | } | |
528 | ||
529 | /* Temporarily park all watchdog threads */ | |
530 | static void softlockup_park_all_threads(void) | |
531 | { | |
532 | cpumask_clear(&watchdog_allowed_mask); | |
533 | softlockup_update_smpboot_threads(); | |
534 | } | |
535 | ||
e8b62b2d TG |
536 | /* Unpark enabled threads */ |
537 | static void softlockup_unpark_threads(void) | |
2eb2527f TG |
538 | { |
539 | cpumask_copy(&watchdog_allowed_mask, &watchdog_cpumask); | |
540 | softlockup_update_smpboot_threads(); | |
541 | } | |
542 | ||
09154985 | 543 | static void softlockup_reconfigure_threads(void) |
2eb2527f | 544 | { |
6592ad2f | 545 | watchdog_nmi_reconfigure(false); |
2eb2527f TG |
546 | softlockup_park_all_threads(); |
547 | set_sample_period(); | |
09154985 TG |
548 | lockup_detector_update_enable(); |
549 | if (watchdog_enabled && watchdog_thresh) | |
e8b62b2d | 550 | softlockup_unpark_threads(); |
6592ad2f | 551 | watchdog_nmi_reconfigure(true); |
2eb2527f TG |
552 | } |
553 | ||
554 | /* | |
555 | * Create the watchdog thread infrastructure. | |
556 | * | |
557 | * The threads are not unparked as watchdog_allowed_mask is empty. When | |
558 | * the threads are sucessfully initialized, take the proper locks and | |
559 | * unpark the threads in the watchdog_cpumask if the watchdog is enabled. | |
560 | */ | |
561 | static __init void softlockup_init_threads(void) | |
562 | { | |
563 | int ret; | |
564 | ||
565 | /* | |
566 | * If sysctl is off and watchdog got disabled on the command line, | |
567 | * nothing to do here. | |
568 | */ | |
09154985 TG |
569 | lockup_detector_update_enable(); |
570 | ||
2eb2527f TG |
571 | if (!IS_ENABLED(CONFIG_SYSCTL) && |
572 | !(watchdog_enabled && watchdog_thresh)) | |
573 | return; | |
574 | ||
575 | ret = smpboot_register_percpu_thread_cpumask(&watchdog_threads, | |
576 | &watchdog_allowed_mask); | |
577 | if (ret) { | |
578 | pr_err("Failed to initialize soft lockup detector threads\n"); | |
579 | return; | |
580 | } | |
581 | ||
582 | mutex_lock(&watchdog_mutex); | |
583 | softlockup_threads_initialized = true; | |
09154985 | 584 | softlockup_reconfigure_threads(); |
2eb2527f TG |
585 | mutex_unlock(&watchdog_mutex); |
586 | } | |
587 | ||
2b9d7f23 TG |
588 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
589 | static inline int watchdog_park_threads(void) { return 0; } | |
590 | static inline void watchdog_unpark_threads(void) { } | |
591 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
592 | static inline void watchdog_disable_all_cpus(void) { } | |
2eb2527f | 593 | static inline void softlockup_init_threads(void) { } |
09154985 | 594 | static void softlockup_reconfigure_threads(void) |
6592ad2f TG |
595 | { |
596 | watchdog_nmi_reconfigure(false); | |
09154985 | 597 | lockup_detector_update_enable(); |
6592ad2f TG |
598 | watchdog_nmi_reconfigure(true); |
599 | } | |
2b9d7f23 | 600 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ |
05a4a952 | 601 | |
941154bd TG |
602 | static void __lockup_detector_cleanup(void) |
603 | { | |
604 | lockdep_assert_held(&watchdog_mutex); | |
605 | hardlockup_detector_perf_cleanup(); | |
606 | } | |
607 | ||
608 | /** | |
609 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
610 | * | |
611 | * Caller must not hold the cpu hotplug rwsem. | |
612 | */ | |
613 | void lockup_detector_cleanup(void) | |
614 | { | |
615 | mutex_lock(&watchdog_mutex); | |
616 | __lockup_detector_cleanup(); | |
617 | mutex_unlock(&watchdog_mutex); | |
618 | } | |
619 | ||
6554fd8c TG |
620 | /** |
621 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
622 | * | |
623 | * Special interface for parisc. It prevents lockup detector warnings from | |
624 | * the default pm_poweroff() function which busy loops forever. | |
625 | */ | |
626 | void lockup_detector_soft_poweroff(void) | |
627 | { | |
628 | watchdog_enabled = 0; | |
629 | } | |
630 | ||
58cf690a UO |
631 | #ifdef CONFIG_SYSCTL |
632 | ||
e8b62b2d | 633 | /* Propagate any changes to the watchdog threads */ |
d57108d4 | 634 | static void proc_watchdog_update(void) |
a0c9cbb9 | 635 | { |
e8b62b2d TG |
636 | /* Remove impossible cpus to keep sysctl output clean. */ |
637 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, cpu_possible_mask); | |
09154985 | 638 | softlockup_reconfigure_threads(); |
a0c9cbb9 UO |
639 | } |
640 | ||
ef246a21 UO |
641 | /* |
642 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
643 | * | |
7feeb9cd TG |
644 | * caller | table->data points to | 'which' |
645 | * -------------------|----------------------------|-------------------------- | |
646 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED | | |
647 | * | | SOFT_WATCHDOG_ENABLED | |
648 | * -------------------|----------------------------|-------------------------- | |
649 | * proc_nmi_watchdog | nmi_watchdog_user_enabled | NMI_WATCHDOG_ENABLED | |
650 | * -------------------|----------------------------|-------------------------- | |
651 | * proc_soft_watchdog | soft_watchdog_user_enabled | SOFT_WATCHDOG_ENABLED | |
ef246a21 UO |
652 | */ |
653 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
654 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
655 | { | |
09154985 | 656 | int err, old, *param = table->data; |
ef246a21 | 657 | |
b7a34981 | 658 | cpu_hotplug_disable(); |
946d1977 | 659 | mutex_lock(&watchdog_mutex); |
ef246a21 | 660 | |
ef246a21 | 661 | if (!write) { |
09154985 TG |
662 | /* |
663 | * On read synchronize the userspace interface. This is a | |
664 | * racy snapshot. | |
665 | */ | |
666 | *param = (watchdog_enabled & which) != 0; | |
ef246a21 UO |
667 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
668 | } else { | |
09154985 | 669 | old = READ_ONCE(*param); |
ef246a21 | 670 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
09154985 | 671 | if (!err && old != READ_ONCE(*param)) |
d57108d4 | 672 | proc_watchdog_update(); |
ef246a21 | 673 | } |
946d1977 | 674 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 675 | cpu_hotplug_enable(); |
ef246a21 UO |
676 | return err; |
677 | } | |
678 | ||
83a80a39 UO |
679 | /* |
680 | * /proc/sys/kernel/watchdog | |
681 | */ | |
682 | int proc_watchdog(struct ctl_table *table, int write, | |
683 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
684 | { | |
685 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
686 | table, write, buffer, lenp, ppos); | |
687 | } | |
688 | ||
689 | /* | |
690 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 691 | */ |
83a80a39 UO |
692 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
693 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
694 | { | |
695 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
696 | table, write, buffer, lenp, ppos); | |
697 | } | |
698 | ||
699 | /* | |
700 | * /proc/sys/kernel/soft_watchdog | |
701 | */ | |
702 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
703 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
704 | { | |
705 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
706 | table, write, buffer, lenp, ppos); | |
707 | } | |
58687acb | 708 | |
83a80a39 UO |
709 | /* |
710 | * /proc/sys/kernel/watchdog_thresh | |
711 | */ | |
712 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
713 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 714 | { |
d57108d4 | 715 | int err, old; |
58687acb | 716 | |
b7a34981 | 717 | cpu_hotplug_disable(); |
946d1977 | 718 | mutex_lock(&watchdog_mutex); |
bcd951cf | 719 | |
d57108d4 | 720 | old = READ_ONCE(watchdog_thresh); |
b8900bc0 | 721 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 722 | |
d57108d4 TG |
723 | if (!err && write && old != READ_ONCE(watchdog_thresh)) |
724 | proc_watchdog_update(); | |
e04ab2bc | 725 | |
946d1977 | 726 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 727 | cpu_hotplug_enable(); |
b8900bc0 | 728 | return err; |
58687acb | 729 | } |
fe4ba3c3 CM |
730 | |
731 | /* | |
732 | * The cpumask is the mask of possible cpus that the watchdog can run | |
733 | * on, not the mask of cpus it is actually running on. This allows the | |
734 | * user to specify a mask that will include cpus that have not yet | |
735 | * been brought online, if desired. | |
736 | */ | |
737 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
738 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
739 | { | |
740 | int err; | |
741 | ||
b7a34981 | 742 | cpu_hotplug_disable(); |
946d1977 | 743 | mutex_lock(&watchdog_mutex); |
8c073d27 | 744 | |
fe4ba3c3 | 745 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
05ba3de7 | 746 | if (!err && write) |
e8b62b2d | 747 | proc_watchdog_update(); |
5490125d | 748 | |
946d1977 | 749 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 750 | cpu_hotplug_enable(); |
fe4ba3c3 CM |
751 | return err; |
752 | } | |
58687acb DZ |
753 | #endif /* CONFIG_SYSCTL */ |
754 | ||
004417a6 | 755 | void __init lockup_detector_init(void) |
58687acb | 756 | { |
fe4ba3c3 CM |
757 | #ifdef CONFIG_NO_HZ_FULL |
758 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
759 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
760 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
761 | } else |
762 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
763 | #else | |
764 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
765 | #endif | |
766 | ||
d57108d4 | 767 | softlockup_init_threads(); |
58687acb | 768 | } |