]>
Commit | Line | Data |
---|---|---|
58687acb DZ |
1 | /* |
2 | * Detect hard and soft lockups on a system | |
3 | * | |
4 | * started by Don Zickus, Copyright (C) 2010 Red Hat, Inc. | |
5 | * | |
86f5e6a7 FLVC |
6 | * Note: Most of this code is borrowed heavily from the original softlockup |
7 | * detector, so thanks to Ingo for the initial implementation. | |
8 | * Some chunks also taken from the old x86-specific nmi watchdog code, thanks | |
58687acb DZ |
9 | * to those contributors as well. |
10 | */ | |
11 | ||
5f92a7b0 | 12 | #define pr_fmt(fmt) "watchdog: " fmt |
4501980a | 13 | |
58687acb DZ |
14 | #include <linux/mm.h> |
15 | #include <linux/cpu.h> | |
16 | #include <linux/nmi.h> | |
17 | #include <linux/init.h> | |
58687acb DZ |
18 | #include <linux/module.h> |
19 | #include <linux/sysctl.h> | |
bcd951cf | 20 | #include <linux/smpboot.h> |
8bd75c77 | 21 | #include <linux/sched/rt.h> |
ae7e81c0 | 22 | #include <uapi/linux/sched/types.h> |
fe4ba3c3 | 23 | #include <linux/tick.h> |
82607adc | 24 | #include <linux/workqueue.h> |
e6017571 | 25 | #include <linux/sched/clock.h> |
b17b0153 | 26 | #include <linux/sched/debug.h> |
58687acb DZ |
27 | |
28 | #include <asm/irq_regs.h> | |
5d1c0f4a | 29 | #include <linux/kvm_para.h> |
81a4beef | 30 | #include <linux/kthread.h> |
58687acb | 31 | |
946d1977 | 32 | static DEFINE_MUTEX(watchdog_mutex); |
ab992dc3 | 33 | |
05a4a952 NP |
34 | int __read_mostly nmi_watchdog_enabled; |
35 | ||
36 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) | |
37 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
38 | NMI_WATCHDOG_ENABLED; | |
84d56e66 | 39 | #else |
249e52e3 | 40 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
84d56e66 | 41 | #endif |
05a4a952 NP |
42 | |
43 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
05a4a952 NP |
44 | /* |
45 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
46 | */ | |
47 | unsigned int __read_mostly hardlockup_panic = | |
48 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
49 | /* | |
50 | * We may not want to enable hard lockup detection by default in all cases, | |
51 | * for example when running the kernel as a guest on a hypervisor. In these | |
52 | * cases this function can be called to disable hard lockup detection. This | |
53 | * function should only be executed once by the boot processor before the | |
54 | * kernel command line parameters are parsed, because otherwise it is not | |
55 | * possible to override this in hardlockup_panic_setup(). | |
56 | */ | |
7a355820 | 57 | void __init hardlockup_detector_disable(void) |
05a4a952 NP |
58 | { |
59 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
60 | } | |
61 | ||
62 | static int __init hardlockup_panic_setup(char *str) | |
63 | { | |
64 | if (!strncmp(str, "panic", 5)) | |
65 | hardlockup_panic = 1; | |
66 | else if (!strncmp(str, "nopanic", 7)) | |
67 | hardlockup_panic = 0; | |
68 | else if (!strncmp(str, "0", 1)) | |
69 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
70 | else if (!strncmp(str, "1", 1)) | |
71 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
72 | return 1; | |
73 | } | |
74 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
75 | ||
368a7e2c TG |
76 | # ifdef CONFIG_SMP |
77 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; | |
05a4a952 | 78 | |
368a7e2c TG |
79 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
80 | { | |
81 | sysctl_hardlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); | |
82 | return 1; | |
83 | } | |
84 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
85 | # endif /* CONFIG_SMP */ | |
86 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ | |
05a4a952 | 87 | |
84d56e66 | 88 | int __read_mostly watchdog_user_enabled; |
4eec42f3 | 89 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 90 | |
05a4a952 | 91 | struct cpumask watchdog_cpumask __read_mostly; |
fe4ba3c3 CM |
92 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); |
93 | ||
ec6a9066 UO |
94 | /* |
95 | * The 'watchdog_running' variable is set to 1 when the watchdog threads | |
96 | * are registered/started and is set to 0 when the watchdog threads are | |
97 | * unregistered/stopped, so it is an indicator whether the threads exist. | |
98 | */ | |
3c00ea82 | 99 | static int __read_mostly watchdog_running; |
05a4a952 NP |
100 | |
101 | /* | |
102 | * These functions can be overridden if an architecture implements its | |
103 | * own hardlockup detector. | |
a10a842f NP |
104 | * |
105 | * watchdog_nmi_enable/disable can be implemented to start and stop when | |
106 | * softlockup watchdog threads start and stop. The arch must select the | |
107 | * SOFTLOCKUP_DETECTOR Kconfig. | |
05a4a952 NP |
108 | */ |
109 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
110 | { | |
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 | ||
a10a842f NP |
119 | /* |
120 | * watchdog_nmi_reconfigure can be implemented to be notified after any | |
121 | * watchdog configuration change. The arch hardlockup watchdog should | |
122 | * respond to the following variables: | |
123 | * - nmi_watchdog_enabled | |
124 | * - watchdog_thresh | |
125 | * - watchdog_cpumask | |
126 | * - sysctl_hardlockup_all_cpu_backtrace | |
127 | * - hardlockup_panic | |
a10a842f | 128 | */ |
2b9d7f23 | 129 | void __weak watchdog_nmi_reconfigure(void) { } |
a10a842f | 130 | |
05a4a952 NP |
131 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
132 | ||
133 | /* Helper for online, unparked cpus. */ | |
134 | #define for_each_watchdog_cpu(cpu) \ | |
135 | for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask) | |
136 | ||
2b9d7f23 TG |
137 | /* Global variables, exported for sysctl */ |
138 | unsigned int __read_mostly softlockup_panic = | |
139 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
140 | int __read_mostly soft_watchdog_enabled; | |
141 | ||
0f34c400 | 142 | static u64 __read_mostly sample_period; |
58687acb DZ |
143 | |
144 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
145 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
146 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
147 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 148 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
149 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
150 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 151 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 152 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 153 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 154 | |
58687acb DZ |
155 | static int __init softlockup_panic_setup(char *str) |
156 | { | |
157 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
58687acb DZ |
158 | return 1; |
159 | } | |
160 | __setup("softlockup_panic=", softlockup_panic_setup); | |
161 | ||
162 | static int __init nowatchdog_setup(char *str) | |
163 | { | |
195daf66 | 164 | watchdog_enabled = 0; |
58687acb DZ |
165 | return 1; |
166 | } | |
167 | __setup("nowatchdog", nowatchdog_setup); | |
168 | ||
58687acb DZ |
169 | static int __init nosoftlockup_setup(char *str) |
170 | { | |
195daf66 | 171 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
172 | return 1; |
173 | } | |
174 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 175 | |
ed235875 | 176 | #ifdef CONFIG_SMP |
368a7e2c TG |
177 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; |
178 | ||
ed235875 AT |
179 | static int __init softlockup_all_cpu_backtrace_setup(char *str) |
180 | { | |
368a7e2c | 181 | sysctl_softlockup_all_cpu_backtrace = !!simple_strtol(str, NULL, 0); |
ed235875 AT |
182 | return 1; |
183 | } | |
184 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 185 | #endif |
58687acb | 186 | |
941154bd TG |
187 | static void __lockup_detector_cleanup(void); |
188 | ||
4eec42f3 MSB |
189 | /* |
190 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
191 | * lockups can have false positives under extreme conditions. So we generally | |
192 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
193 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
194 | * time the hard threshold is. | |
195 | */ | |
6e9101ae | 196 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
197 | { |
198 | return watchdog_thresh * 2; | |
199 | } | |
58687acb DZ |
200 | |
201 | /* | |
202 | * Returns seconds, approximately. We don't need nanosecond | |
203 | * resolution, and we don't need to waste time with a big divide when | |
204 | * 2^30ns == 1.074s. | |
205 | */ | |
c06b4f19 | 206 | static unsigned long get_timestamp(void) |
58687acb | 207 | { |
545a2bf7 | 208 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
209 | } |
210 | ||
0f34c400 | 211 | static void set_sample_period(void) |
58687acb DZ |
212 | { |
213 | /* | |
586692a5 | 214 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
215 | * the divide by 5 is to give hrtimer several chances (two |
216 | * or three with the current relation between the soft | |
217 | * and hard thresholds) to increment before the | |
218 | * hardlockup detector generates a warning | |
58687acb | 219 | */ |
0f34c400 | 220 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
7edaeb68 | 221 | watchdog_update_hrtimer_threshold(sample_period); |
58687acb DZ |
222 | } |
223 | ||
224 | /* Commands for resetting the watchdog */ | |
225 | static void __touch_watchdog(void) | |
226 | { | |
c06b4f19 | 227 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
228 | } |
229 | ||
03e0d461 TH |
230 | /** |
231 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
232 | * | |
233 | * Call when the scheduler may have stalled for legitimate reasons | |
234 | * preventing the watchdog task from executing - e.g. the scheduler | |
235 | * entering idle state. This should only be used for scheduler events. | |
236 | * Use touch_softlockup_watchdog() for everything else. | |
237 | */ | |
238 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 239 | { |
7861144b AM |
240 | /* |
241 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
242 | * gets zeroed here, so use the raw_ operation. | |
243 | */ | |
244 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 245 | } |
03e0d461 TH |
246 | |
247 | void touch_softlockup_watchdog(void) | |
248 | { | |
249 | touch_softlockup_watchdog_sched(); | |
82607adc | 250 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 251 | } |
0167c781 | 252 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 253 | |
332fbdbc | 254 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
255 | { |
256 | int cpu; | |
257 | ||
258 | /* | |
259 | * this is done lockless | |
260 | * do we care if a 0 races with a timestamp? | |
261 | * all it means is the softlock check starts one cycle later | |
262 | */ | |
fe4ba3c3 | 263 | for_each_watchdog_cpu(cpu) |
58687acb | 264 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 265 | wq_watchdog_touch(-1); |
58687acb DZ |
266 | } |
267 | ||
58687acb DZ |
268 | void touch_softlockup_watchdog_sync(void) |
269 | { | |
f7f66b05 CL |
270 | __this_cpu_write(softlockup_touch_sync, true); |
271 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
272 | } |
273 | ||
26e09c6e | 274 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 275 | { |
c06b4f19 | 276 | unsigned long now = get_timestamp(); |
58687acb | 277 | |
39d2da21 | 278 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
279 | /* Warn about unreasonable delays. */ |
280 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
281 | return now - touch_ts; | |
282 | } | |
58687acb DZ |
283 | return 0; |
284 | } | |
285 | ||
05a4a952 NP |
286 | /* watchdog detector functions */ |
287 | bool is_hardlockup(void) | |
58687acb | 288 | { |
05a4a952 | 289 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 290 | |
05a4a952 NP |
291 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
292 | return true; | |
293 | ||
294 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
295 | return false; | |
73ce0511 | 296 | } |
05a4a952 NP |
297 | |
298 | static void watchdog_interrupt_count(void) | |
73ce0511 | 299 | { |
05a4a952 | 300 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 301 | } |
58687acb | 302 | |
58cf690a UO |
303 | static int watchdog_enable_all_cpus(void); |
304 | static void watchdog_disable_all_cpus(void); | |
305 | ||
58687acb DZ |
306 | /* watchdog kicker functions */ |
307 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
308 | { | |
909ea964 | 309 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
310 | struct pt_regs *regs = get_irq_regs(); |
311 | int duration; | |
ed235875 | 312 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 313 | |
01f0a027 | 314 | if (!watchdog_enabled) |
b94f5118 DZ |
315 | return HRTIMER_NORESTART; |
316 | ||
58687acb DZ |
317 | /* kick the hardlockup detector */ |
318 | watchdog_interrupt_count(); | |
319 | ||
320 | /* kick the softlockup detector */ | |
909ea964 | 321 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
322 | |
323 | /* .. and repeat */ | |
0f34c400 | 324 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
325 | |
326 | if (touch_ts == 0) { | |
909ea964 | 327 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
328 | /* |
329 | * If the time stamp was touched atomically | |
330 | * make sure the scheduler tick is up to date. | |
331 | */ | |
909ea964 | 332 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
333 | sched_clock_tick(); |
334 | } | |
5d1c0f4a EM |
335 | |
336 | /* Clear the guest paused flag on watchdog reset */ | |
337 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
338 | __touch_watchdog(); |
339 | return HRTIMER_RESTART; | |
340 | } | |
341 | ||
342 | /* check for a softlockup | |
343 | * This is done by making sure a high priority task is | |
344 | * being scheduled. The task touches the watchdog to | |
345 | * indicate it is getting cpu time. If it hasn't then | |
346 | * this is a good indication some task is hogging the cpu | |
347 | */ | |
26e09c6e | 348 | duration = is_softlockup(touch_ts); |
58687acb | 349 | if (unlikely(duration)) { |
5d1c0f4a EM |
350 | /* |
351 | * If a virtual machine is stopped by the host it can look to | |
352 | * the watchdog like a soft lockup, check to see if the host | |
353 | * stopped the vm before we issue the warning | |
354 | */ | |
355 | if (kvm_check_and_clear_guest_paused()) | |
356 | return HRTIMER_RESTART; | |
357 | ||
58687acb | 358 | /* only warn once */ |
b1a8de1f | 359 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
360 | /* | |
361 | * When multiple processes are causing softlockups the | |
362 | * softlockup detector only warns on the first one | |
363 | * because the code relies on a full quiet cycle to | |
364 | * re-arm. The second process prevents the quiet cycle | |
365 | * and never gets reported. Use task pointers to detect | |
366 | * this. | |
367 | */ | |
368 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
369 | current) { | |
370 | __this_cpu_write(soft_watchdog_warn, false); | |
371 | __touch_watchdog(); | |
372 | } | |
58687acb | 373 | return HRTIMER_RESTART; |
b1a8de1f | 374 | } |
58687acb | 375 | |
ed235875 AT |
376 | if (softlockup_all_cpu_backtrace) { |
377 | /* Prevent multiple soft-lockup reports if one cpu is already | |
378 | * engaged in dumping cpu back traces | |
379 | */ | |
380 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
381 | /* Someone else will report us. Let's give up */ | |
382 | __this_cpu_write(soft_watchdog_warn, true); | |
383 | return HRTIMER_RESTART; | |
384 | } | |
385 | } | |
386 | ||
656c3b79 | 387 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 388 | smp_processor_id(), duration, |
58687acb | 389 | current->comm, task_pid_nr(current)); |
b1a8de1f | 390 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
391 | print_modules(); |
392 | print_irqtrace_events(current); | |
393 | if (regs) | |
394 | show_regs(regs); | |
395 | else | |
396 | dump_stack(); | |
397 | ||
ed235875 AT |
398 | if (softlockup_all_cpu_backtrace) { |
399 | /* Avoid generating two back traces for current | |
400 | * given that one is already made above | |
401 | */ | |
402 | trigger_allbutself_cpu_backtrace(); | |
403 | ||
404 | clear_bit(0, &soft_lockup_nmi_warn); | |
405 | /* Barrier to sync with other cpus */ | |
406 | smp_mb__after_atomic(); | |
407 | } | |
408 | ||
69361eef | 409 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
410 | if (softlockup_panic) |
411 | panic("softlockup: hung tasks"); | |
909ea964 | 412 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 413 | } else |
909ea964 | 414 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
415 | |
416 | return HRTIMER_RESTART; | |
417 | } | |
418 | ||
bcd951cf TG |
419 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
420 | { | |
421 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 422 | |
bcd951cf TG |
423 | sched_setscheduler(current, policy, ¶m); |
424 | } | |
425 | ||
426 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 427 | { |
01f0a027 | 428 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 429 | |
01f0a027 TG |
430 | /* |
431 | * Start the timer first to prevent the NMI watchdog triggering | |
432 | * before the timer has a chance to fire. | |
433 | */ | |
3935e895 BM |
434 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); |
435 | hrtimer->function = watchdog_timer_fn; | |
01f0a027 TG |
436 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
437 | HRTIMER_MODE_REL_PINNED); | |
3935e895 | 438 | |
01f0a027 TG |
439 | /* Initialize timestamp */ |
440 | __touch_watchdog(); | |
bcd951cf TG |
441 | /* Enable the perf event */ |
442 | watchdog_nmi_enable(cpu); | |
58687acb | 443 | |
bcd951cf | 444 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); |
bcd951cf | 445 | } |
58687acb | 446 | |
bcd951cf TG |
447 | static void watchdog_disable(unsigned int cpu) |
448 | { | |
01f0a027 | 449 | struct hrtimer *hrtimer = this_cpu_ptr(&watchdog_hrtimer); |
58687acb | 450 | |
bcd951cf | 451 | watchdog_set_prio(SCHED_NORMAL, 0); |
01f0a027 TG |
452 | /* |
453 | * Disable the perf event first. That prevents that a large delay | |
454 | * between disabling the timer and disabling the perf event causes | |
455 | * the perf NMI to detect a false positive. | |
456 | */ | |
bcd951cf | 457 | watchdog_nmi_disable(cpu); |
01f0a027 | 458 | hrtimer_cancel(hrtimer); |
58687acb DZ |
459 | } |
460 | ||
b8900bc0 FW |
461 | static void watchdog_cleanup(unsigned int cpu, bool online) |
462 | { | |
463 | watchdog_disable(cpu); | |
464 | } | |
465 | ||
bcd951cf TG |
466 | static int watchdog_should_run(unsigned int cpu) |
467 | { | |
468 | return __this_cpu_read(hrtimer_interrupts) != | |
469 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
470 | } | |
471 | ||
472 | /* | |
473 | * The watchdog thread function - touches the timestamp. | |
474 | * | |
0f34c400 | 475 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
476 | * default) to reset the softlockup timestamp. If this gets delayed |
477 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
478 | * triggers in watchdog_timer_fn(). | |
479 | */ | |
480 | static void watchdog(unsigned int cpu) | |
481 | { | |
482 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
483 | __this_cpu_read(hrtimer_interrupts)); | |
484 | __touch_watchdog(); | |
485 | } | |
58687acb | 486 | |
b8900bc0 FW |
487 | static struct smp_hotplug_thread watchdog_threads = { |
488 | .store = &softlockup_watchdog, | |
489 | .thread_should_run = watchdog_should_run, | |
490 | .thread_fn = watchdog, | |
491 | .thread_comm = "watchdog/%u", | |
492 | .setup = watchdog_enable, | |
493 | .cleanup = watchdog_cleanup, | |
494 | .park = watchdog_disable, | |
495 | .unpark = watchdog_enable, | |
496 | }; | |
497 | ||
81a4beef UO |
498 | /* |
499 | * park all watchdog threads that are specified in 'watchdog_cpumask' | |
ee7fed54 UO |
500 | * |
501 | * This function returns an error if kthread_park() of a watchdog thread | |
502 | * fails. In this situation, the watchdog threads of some CPUs can already | |
503 | * be parked and the watchdog threads of other CPUs can still be runnable. | |
504 | * Callers are expected to handle this special condition as appropriate in | |
505 | * their context. | |
a2a45b85 UO |
506 | * |
507 | * This function may only be called in a context that is protected against | |
508 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
509 | */ |
510 | static int watchdog_park_threads(void) | |
511 | { | |
512 | int cpu, ret = 0; | |
513 | ||
81a4beef UO |
514 | for_each_watchdog_cpu(cpu) { |
515 | ret = kthread_park(per_cpu(softlockup_watchdog, cpu)); | |
516 | if (ret) | |
517 | break; | |
518 | } | |
81a4beef UO |
519 | return ret; |
520 | } | |
521 | ||
522 | /* | |
523 | * unpark all watchdog threads that are specified in 'watchdog_cpumask' | |
a2a45b85 UO |
524 | * |
525 | * This function may only be called in a context that is protected against | |
526 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
527 | */ |
528 | static void watchdog_unpark_threads(void) | |
529 | { | |
530 | int cpu; | |
531 | ||
81a4beef UO |
532 | for_each_watchdog_cpu(cpu) |
533 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); | |
81a4beef UO |
534 | } |
535 | ||
b43cb43c | 536 | static int update_watchdog_all_cpus(void) |
9809b18f | 537 | { |
b43cb43c UO |
538 | int ret; |
539 | ||
540 | ret = watchdog_park_threads(); | |
541 | if (ret) | |
542 | return ret; | |
543 | ||
d4bdd0b2 | 544 | watchdog_unpark_threads(); |
b43cb43c UO |
545 | |
546 | return 0; | |
9809b18f MH |
547 | } |
548 | ||
b2f57c3a | 549 | static int watchdog_enable_all_cpus(void) |
58687acb | 550 | { |
b8900bc0 | 551 | int err = 0; |
58687acb | 552 | |
3c00ea82 | 553 | if (!watchdog_running) { |
230ec939 FW |
554 | err = smpboot_register_percpu_thread_cpumask(&watchdog_threads, |
555 | &watchdog_cpumask); | |
b8900bc0 FW |
556 | if (err) |
557 | pr_err("Failed to create watchdog threads, disabled\n"); | |
230ec939 | 558 | else |
3c00ea82 | 559 | watchdog_running = 1; |
b2f57c3a UO |
560 | } else { |
561 | /* | |
562 | * Enable/disable the lockup detectors or | |
563 | * change the sample period 'on the fly'. | |
564 | */ | |
b43cb43c UO |
565 | err = update_watchdog_all_cpus(); |
566 | ||
567 | if (err) { | |
568 | watchdog_disable_all_cpus(); | |
569 | pr_err("Failed to update lockup detectors, disabled\n"); | |
570 | } | |
bcd951cf | 571 | } |
b8900bc0 | 572 | |
b43cb43c UO |
573 | if (err) |
574 | watchdog_enabled = 0; | |
575 | ||
b8900bc0 | 576 | return err; |
58687acb DZ |
577 | } |
578 | ||
579 | static void watchdog_disable_all_cpus(void) | |
580 | { | |
3c00ea82 FW |
581 | if (watchdog_running) { |
582 | watchdog_running = 0; | |
b8900bc0 | 583 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 584 | } |
58687acb DZ |
585 | } |
586 | ||
2b9d7f23 TG |
587 | #else /* CONFIG_SOFTLOCKUP_DETECTOR */ |
588 | static inline int watchdog_park_threads(void) { return 0; } | |
589 | static inline void watchdog_unpark_threads(void) { } | |
590 | static inline int watchdog_enable_all_cpus(void) { return 0; } | |
591 | static inline void watchdog_disable_all_cpus(void) { } | |
592 | static inline void set_sample_period(void) { } | |
593 | #endif /* !CONFIG_SOFTLOCKUP_DETECTOR */ | |
05a4a952 | 594 | |
941154bd TG |
595 | static void __lockup_detector_cleanup(void) |
596 | { | |
597 | lockdep_assert_held(&watchdog_mutex); | |
598 | hardlockup_detector_perf_cleanup(); | |
599 | } | |
600 | ||
601 | /** | |
602 | * lockup_detector_cleanup - Cleanup after cpu hotplug or sysctl changes | |
603 | * | |
604 | * Caller must not hold the cpu hotplug rwsem. | |
605 | */ | |
606 | void lockup_detector_cleanup(void) | |
607 | { | |
608 | mutex_lock(&watchdog_mutex); | |
609 | __lockup_detector_cleanup(); | |
610 | mutex_unlock(&watchdog_mutex); | |
611 | } | |
612 | ||
6554fd8c TG |
613 | /** |
614 | * lockup_detector_soft_poweroff - Interface to stop lockup detector(s) | |
615 | * | |
616 | * Special interface for parisc. It prevents lockup detector warnings from | |
617 | * the default pm_poweroff() function which busy loops forever. | |
618 | */ | |
619 | void lockup_detector_soft_poweroff(void) | |
620 | { | |
621 | watchdog_enabled = 0; | |
622 | } | |
623 | ||
58cf690a UO |
624 | #ifdef CONFIG_SYSCTL |
625 | ||
58687acb | 626 | /* |
a0c9cbb9 UO |
627 | * Update the run state of the lockup detectors. |
628 | */ | |
629 | static int proc_watchdog_update(void) | |
630 | { | |
631 | int err = 0; | |
632 | ||
633 | /* | |
634 | * Watchdog threads won't be started if they are already active. | |
635 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes | |
636 | * care of this. If those threads are already active, the sample | |
637 | * period will be updated and the lockup detectors will be enabled | |
638 | * or disabled 'on the fly'. | |
639 | */ | |
640 | if (watchdog_enabled && watchdog_thresh) | |
b2f57c3a | 641 | err = watchdog_enable_all_cpus(); |
a0c9cbb9 UO |
642 | else |
643 | watchdog_disable_all_cpus(); | |
644 | ||
a10a842f NP |
645 | watchdog_nmi_reconfigure(); |
646 | ||
941154bd TG |
647 | __lockup_detector_cleanup(); |
648 | ||
a0c9cbb9 UO |
649 | return err; |
650 | ||
651 | } | |
652 | ||
ef246a21 UO |
653 | /* |
654 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
655 | * | |
656 | * caller | table->data points to | 'which' contains the flag(s) | |
657 | * -------------------|-----------------------|----------------------------- | |
658 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
659 | * | | with SOFT_WATCHDOG_ENABLED | |
660 | * -------------------|-----------------------|----------------------------- | |
661 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
662 | * -------------------|-----------------------|----------------------------- | |
663 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
664 | */ | |
665 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
666 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
667 | { | |
668 | int err, old, new; | |
669 | int *watchdog_param = (int *)table->data; | |
670 | ||
b7a34981 | 671 | cpu_hotplug_disable(); |
946d1977 | 672 | mutex_lock(&watchdog_mutex); |
ef246a21 UO |
673 | |
674 | /* | |
675 | * If the parameter is being read return the state of the corresponding | |
676 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
677 | * run state of the lockup detectors. | |
678 | */ | |
679 | if (!write) { | |
680 | *watchdog_param = (watchdog_enabled & which) != 0; | |
681 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
682 | } else { | |
683 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
684 | if (err) | |
685 | goto out; | |
686 | ||
687 | /* | |
688 | * There is a race window between fetching the current value | |
689 | * from 'watchdog_enabled' and storing the new value. During | |
690 | * this race window, watchdog_nmi_enable() can sneak in and | |
691 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
692 | * The 'cmpxchg' detects this race and the loop retries. | |
693 | */ | |
694 | do { | |
695 | old = watchdog_enabled; | |
696 | /* | |
697 | * If the parameter value is not zero set the | |
698 | * corresponding bit(s), else clear it(them). | |
699 | */ | |
700 | if (*watchdog_param) | |
701 | new = old | which; | |
702 | else | |
703 | new = old & ~which; | |
704 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
705 | ||
706 | /* | |
b43cb43c UO |
707 | * Update the run state of the lockup detectors. There is _no_ |
708 | * need to check the value returned by proc_watchdog_update() | |
709 | * and to restore the previous value of 'watchdog_enabled' as | |
710 | * both lockup detectors are disabled if proc_watchdog_update() | |
711 | * returns an error. | |
ef246a21 | 712 | */ |
a1ee1932 JH |
713 | if (old == new) |
714 | goto out; | |
715 | ||
ef246a21 | 716 | err = proc_watchdog_update(); |
ef246a21 UO |
717 | } |
718 | out: | |
946d1977 | 719 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 720 | cpu_hotplug_enable(); |
ef246a21 UO |
721 | return err; |
722 | } | |
723 | ||
83a80a39 UO |
724 | /* |
725 | * /proc/sys/kernel/watchdog | |
726 | */ | |
727 | int proc_watchdog(struct ctl_table *table, int write, | |
728 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
729 | { | |
730 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
731 | table, write, buffer, lenp, ppos); | |
732 | } | |
733 | ||
734 | /* | |
735 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 736 | */ |
83a80a39 UO |
737 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
738 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
739 | { | |
740 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
741 | table, write, buffer, lenp, ppos); | |
742 | } | |
743 | ||
744 | /* | |
745 | * /proc/sys/kernel/soft_watchdog | |
746 | */ | |
747 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
748 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
749 | { | |
750 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
751 | table, write, buffer, lenp, ppos); | |
752 | } | |
58687acb | 753 | |
83a80a39 UO |
754 | /* |
755 | * /proc/sys/kernel/watchdog_thresh | |
756 | */ | |
757 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
758 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 759 | { |
a1ee1932 | 760 | int err, old, new; |
58687acb | 761 | |
b7a34981 | 762 | cpu_hotplug_disable(); |
946d1977 | 763 | mutex_lock(&watchdog_mutex); |
bcd951cf | 764 | |
83a80a39 | 765 | old = ACCESS_ONCE(watchdog_thresh); |
b8900bc0 | 766 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 767 | |
b8900bc0 | 768 | if (err || !write) |
359e6fab | 769 | goto out; |
e04ab2bc | 770 | |
b66a2356 | 771 | /* |
d283c640 | 772 | * Update the sample period. Restore on failure. |
b66a2356 | 773 | */ |
a1ee1932 JH |
774 | new = ACCESS_ONCE(watchdog_thresh); |
775 | if (old == new) | |
776 | goto out; | |
777 | ||
83a80a39 UO |
778 | set_sample_period(); |
779 | err = proc_watchdog_update(); | |
d283c640 | 780 | if (err) { |
83a80a39 | 781 | watchdog_thresh = old; |
d283c640 UO |
782 | set_sample_period(); |
783 | } | |
359e6fab | 784 | out: |
946d1977 | 785 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 786 | cpu_hotplug_enable(); |
b8900bc0 | 787 | return err; |
58687acb | 788 | } |
fe4ba3c3 | 789 | |
2b9d7f23 TG |
790 | static int watchdog_update_cpus(void) |
791 | { | |
792 | if (IS_ENABLED(CONFIG_SOFTLOCKUP_DETECTOR)) { | |
793 | return smpboot_update_cpumask_percpu_thread(&watchdog_threads, | |
794 | &watchdog_cpumask); | |
795 | } | |
796 | return 0; | |
797 | } | |
798 | ||
fe4ba3c3 CM |
799 | /* |
800 | * The cpumask is the mask of possible cpus that the watchdog can run | |
801 | * on, not the mask of cpus it is actually running on. This allows the | |
802 | * user to specify a mask that will include cpus that have not yet | |
803 | * been brought online, if desired. | |
804 | */ | |
805 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
806 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
807 | { | |
808 | int err; | |
809 | ||
b7a34981 | 810 | cpu_hotplug_disable(); |
946d1977 | 811 | mutex_lock(&watchdog_mutex); |
8c073d27 | 812 | |
fe4ba3c3 CM |
813 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
814 | if (!err && write) { | |
815 | /* Remove impossible cpus to keep sysctl output cleaner. */ | |
816 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, | |
817 | cpu_possible_mask); | |
818 | ||
819 | if (watchdog_running) { | |
820 | /* | |
821 | * Failure would be due to being unable to allocate | |
822 | * a temporary cpumask, so we are likely not in a | |
823 | * position to do much else to make things better. | |
824 | */ | |
a10a842f | 825 | if (watchdog_update_cpus() != 0) |
fe4ba3c3 CM |
826 | pr_err("cpumask update failed\n"); |
827 | } | |
a10a842f NP |
828 | |
829 | watchdog_nmi_reconfigure(); | |
941154bd | 830 | __lockup_detector_cleanup(); |
fe4ba3c3 | 831 | } |
5490125d | 832 | |
946d1977 | 833 | mutex_unlock(&watchdog_mutex); |
b7a34981 | 834 | cpu_hotplug_enable(); |
fe4ba3c3 CM |
835 | return err; |
836 | } | |
837 | ||
58687acb DZ |
838 | #endif /* CONFIG_SYSCTL */ |
839 | ||
004417a6 | 840 | void __init lockup_detector_init(void) |
58687acb | 841 | { |
0f34c400 | 842 | set_sample_period(); |
b8900bc0 | 843 | |
fe4ba3c3 CM |
844 | #ifdef CONFIG_NO_HZ_FULL |
845 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
846 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
847 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
848 | } else |
849 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
850 | #else | |
851 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
852 | #endif | |
853 | ||
195daf66 | 854 | if (watchdog_enabled) |
b2f57c3a | 855 | watchdog_enable_all_cpus(); |
58687acb | 856 | } |