]>
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 | ||
4501980a AM |
12 | #define pr_fmt(fmt) "NMI watchdog: " fmt |
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 | |
05a4a952 | 32 | /* Watchdog configuration */ |
ab992dc3 PZ |
33 | static DEFINE_MUTEX(watchdog_proc_mutex); |
34 | ||
05a4a952 NP |
35 | int __read_mostly nmi_watchdog_enabled; |
36 | ||
37 | #if defined(CONFIG_HARDLOCKUP_DETECTOR) || defined(CONFIG_HAVE_NMI_WATCHDOG) | |
38 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED | | |
39 | NMI_WATCHDOG_ENABLED; | |
84d56e66 | 40 | #else |
249e52e3 | 41 | unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; |
84d56e66 | 42 | #endif |
05a4a952 NP |
43 | |
44 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
45 | /* boot commands */ | |
46 | /* | |
47 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
48 | */ | |
49 | unsigned int __read_mostly hardlockup_panic = | |
50 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
51 | /* | |
52 | * We may not want to enable hard lockup detection by default in all cases, | |
53 | * for example when running the kernel as a guest on a hypervisor. In these | |
54 | * cases this function can be called to disable hard lockup detection. This | |
55 | * function should only be executed once by the boot processor before the | |
56 | * kernel command line parameters are parsed, because otherwise it is not | |
57 | * possible to override this in hardlockup_panic_setup(). | |
58 | */ | |
59 | void hardlockup_detector_disable(void) | |
60 | { | |
61 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
62 | } | |
63 | ||
64 | static int __init hardlockup_panic_setup(char *str) | |
65 | { | |
66 | if (!strncmp(str, "panic", 5)) | |
67 | hardlockup_panic = 1; | |
68 | else if (!strncmp(str, "nopanic", 7)) | |
69 | hardlockup_panic = 0; | |
70 | else if (!strncmp(str, "0", 1)) | |
71 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; | |
72 | else if (!strncmp(str, "1", 1)) | |
73 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
74 | return 1; | |
75 | } | |
76 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
77 | ||
78 | #endif | |
79 | ||
80 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR | |
84d56e66 | 81 | int __read_mostly soft_watchdog_enabled; |
05a4a952 NP |
82 | #endif |
83 | ||
84d56e66 | 84 | int __read_mostly watchdog_user_enabled; |
4eec42f3 | 85 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 86 | |
ed235875 AT |
87 | #ifdef CONFIG_SMP |
88 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
55537871 | 89 | int __read_mostly sysctl_hardlockup_all_cpu_backtrace; |
ed235875 | 90 | #endif |
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; |
ec6a9066 UO |
100 | /* |
101 | * If a subsystem has a need to deactivate the watchdog temporarily, it | |
102 | * can use the suspend/resume interface to achieve this. The content of | |
103 | * the 'watchdog_suspended' variable reflects this state. Existing threads | |
104 | * are parked/unparked by the lockup_detector_{suspend|resume} functions | |
105 | * (see comment blocks pertaining to those functions for further details). | |
106 | * | |
107 | * 'watchdog_suspended' also prevents threads from being registered/started | |
108 | * or unregistered/stopped via parameters in /proc/sys/kernel, so the state | |
109 | * of 'watchdog_running' cannot change while the watchdog is deactivated | |
110 | * temporarily (see related code in 'proc' handlers). | |
111 | */ | |
05a4a952 NP |
112 | int __read_mostly watchdog_suspended; |
113 | ||
114 | /* | |
115 | * These functions can be overridden if an architecture implements its | |
116 | * own hardlockup detector. | |
117 | */ | |
118 | int __weak watchdog_nmi_enable(unsigned int cpu) | |
119 | { | |
120 | return 0; | |
121 | } | |
122 | void __weak watchdog_nmi_disable(unsigned int cpu) | |
123 | { | |
124 | } | |
125 | ||
126 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR | |
127 | ||
128 | /* Helper for online, unparked cpus. */ | |
129 | #define for_each_watchdog_cpu(cpu) \ | |
130 | for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask) | |
131 | ||
132 | atomic_t watchdog_park_in_progress = ATOMIC_INIT(0); | |
ec6a9066 | 133 | |
0f34c400 | 134 | static u64 __read_mostly sample_period; |
58687acb DZ |
135 | |
136 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
137 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
138 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
139 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 140 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
141 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
142 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 143 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
58687acb | 144 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
ed235875 | 145 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 146 | |
58687acb DZ |
147 | unsigned int __read_mostly softlockup_panic = |
148 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
149 | ||
150 | static int __init softlockup_panic_setup(char *str) | |
151 | { | |
152 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
153 | ||
154 | return 1; | |
155 | } | |
156 | __setup("softlockup_panic=", softlockup_panic_setup); | |
157 | ||
158 | static int __init nowatchdog_setup(char *str) | |
159 | { | |
195daf66 | 160 | watchdog_enabled = 0; |
58687acb DZ |
161 | return 1; |
162 | } | |
163 | __setup("nowatchdog", nowatchdog_setup); | |
164 | ||
58687acb DZ |
165 | static int __init nosoftlockup_setup(char *str) |
166 | { | |
195daf66 | 167 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
168 | return 1; |
169 | } | |
170 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 171 | |
ed235875 AT |
172 | #ifdef CONFIG_SMP |
173 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
174 | { | |
175 | sysctl_softlockup_all_cpu_backtrace = | |
176 | !!simple_strtol(str, NULL, 0); | |
177 | return 1; | |
178 | } | |
179 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
05a4a952 | 180 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
55537871 JK |
181 | static int __init hardlockup_all_cpu_backtrace_setup(char *str) |
182 | { | |
183 | sysctl_hardlockup_all_cpu_backtrace = | |
184 | !!simple_strtol(str, NULL, 0); | |
185 | return 1; | |
186 | } | |
187 | __setup("hardlockup_all_cpu_backtrace=", hardlockup_all_cpu_backtrace_setup); | |
ed235875 | 188 | #endif |
05a4a952 | 189 | #endif |
58687acb | 190 | |
4eec42f3 MSB |
191 | /* |
192 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
193 | * lockups can have false positives under extreme conditions. So we generally | |
194 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
195 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
196 | * time the hard threshold is. | |
197 | */ | |
6e9101ae | 198 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
199 | { |
200 | return watchdog_thresh * 2; | |
201 | } | |
58687acb DZ |
202 | |
203 | /* | |
204 | * Returns seconds, approximately. We don't need nanosecond | |
205 | * resolution, and we don't need to waste time with a big divide when | |
206 | * 2^30ns == 1.074s. | |
207 | */ | |
c06b4f19 | 208 | static unsigned long get_timestamp(void) |
58687acb | 209 | { |
545a2bf7 | 210 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
211 | } |
212 | ||
0f34c400 | 213 | static void set_sample_period(void) |
58687acb DZ |
214 | { |
215 | /* | |
586692a5 | 216 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
217 | * the divide by 5 is to give hrtimer several chances (two |
218 | * or three with the current relation between the soft | |
219 | * and hard thresholds) to increment before the | |
220 | * hardlockup detector generates a warning | |
58687acb | 221 | */ |
0f34c400 | 222 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
58687acb DZ |
223 | } |
224 | ||
225 | /* Commands for resetting the watchdog */ | |
226 | static void __touch_watchdog(void) | |
227 | { | |
c06b4f19 | 228 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
229 | } |
230 | ||
03e0d461 TH |
231 | /** |
232 | * touch_softlockup_watchdog_sched - touch watchdog on scheduler stalls | |
233 | * | |
234 | * Call when the scheduler may have stalled for legitimate reasons | |
235 | * preventing the watchdog task from executing - e.g. the scheduler | |
236 | * entering idle state. This should only be used for scheduler events. | |
237 | * Use touch_softlockup_watchdog() for everything else. | |
238 | */ | |
239 | void touch_softlockup_watchdog_sched(void) | |
58687acb | 240 | { |
7861144b AM |
241 | /* |
242 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
243 | * gets zeroed here, so use the raw_ operation. | |
244 | */ | |
245 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 246 | } |
03e0d461 TH |
247 | |
248 | void touch_softlockup_watchdog(void) | |
249 | { | |
250 | touch_softlockup_watchdog_sched(); | |
82607adc | 251 | wq_watchdog_touch(raw_smp_processor_id()); |
03e0d461 | 252 | } |
0167c781 | 253 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 254 | |
332fbdbc | 255 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
256 | { |
257 | int cpu; | |
258 | ||
259 | /* | |
260 | * this is done lockless | |
261 | * do we care if a 0 races with a timestamp? | |
262 | * all it means is the softlock check starts one cycle later | |
263 | */ | |
fe4ba3c3 | 264 | for_each_watchdog_cpu(cpu) |
58687acb | 265 | per_cpu(watchdog_touch_ts, cpu) = 0; |
82607adc | 266 | wq_watchdog_touch(-1); |
58687acb DZ |
267 | } |
268 | ||
58687acb DZ |
269 | void touch_softlockup_watchdog_sync(void) |
270 | { | |
f7f66b05 CL |
271 | __this_cpu_write(softlockup_touch_sync, true); |
272 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
273 | } |
274 | ||
26e09c6e | 275 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 276 | { |
c06b4f19 | 277 | unsigned long now = get_timestamp(); |
58687acb | 278 | |
39d2da21 | 279 | if ((watchdog_enabled & SOFT_WATCHDOG_ENABLED) && watchdog_thresh){ |
195daf66 UO |
280 | /* Warn about unreasonable delays. */ |
281 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
282 | return now - touch_ts; | |
283 | } | |
58687acb DZ |
284 | return 0; |
285 | } | |
286 | ||
05a4a952 NP |
287 | /* watchdog detector functions */ |
288 | bool is_hardlockup(void) | |
58687acb | 289 | { |
05a4a952 | 290 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
bcd951cf | 291 | |
05a4a952 NP |
292 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
293 | return true; | |
294 | ||
295 | __this_cpu_write(hrtimer_interrupts_saved, hrint); | |
296 | return false; | |
73ce0511 | 297 | } |
05a4a952 NP |
298 | |
299 | static void watchdog_interrupt_count(void) | |
73ce0511 | 300 | { |
05a4a952 | 301 | __this_cpu_inc(hrtimer_interrupts); |
73ce0511 | 302 | } |
58687acb | 303 | |
58cf690a UO |
304 | static int watchdog_enable_all_cpus(void); |
305 | static void watchdog_disable_all_cpus(void); | |
306 | ||
58687acb DZ |
307 | /* watchdog kicker functions */ |
308 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
309 | { | |
909ea964 | 310 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
311 | struct pt_regs *regs = get_irq_regs(); |
312 | int duration; | |
ed235875 | 313 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb | 314 | |
b94f5118 DZ |
315 | if (atomic_read(&watchdog_park_in_progress) != 0) |
316 | return HRTIMER_NORESTART; | |
317 | ||
58687acb DZ |
318 | /* kick the hardlockup detector */ |
319 | watchdog_interrupt_count(); | |
320 | ||
321 | /* kick the softlockup detector */ | |
909ea964 | 322 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
323 | |
324 | /* .. and repeat */ | |
0f34c400 | 325 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
326 | |
327 | if (touch_ts == 0) { | |
909ea964 | 328 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
329 | /* |
330 | * If the time stamp was touched atomically | |
331 | * make sure the scheduler tick is up to date. | |
332 | */ | |
909ea964 | 333 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
334 | sched_clock_tick(); |
335 | } | |
5d1c0f4a EM |
336 | |
337 | /* Clear the guest paused flag on watchdog reset */ | |
338 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
339 | __touch_watchdog(); |
340 | return HRTIMER_RESTART; | |
341 | } | |
342 | ||
343 | /* check for a softlockup | |
344 | * This is done by making sure a high priority task is | |
345 | * being scheduled. The task touches the watchdog to | |
346 | * indicate it is getting cpu time. If it hasn't then | |
347 | * this is a good indication some task is hogging the cpu | |
348 | */ | |
26e09c6e | 349 | duration = is_softlockup(touch_ts); |
58687acb | 350 | if (unlikely(duration)) { |
5d1c0f4a EM |
351 | /* |
352 | * If a virtual machine is stopped by the host it can look to | |
353 | * the watchdog like a soft lockup, check to see if the host | |
354 | * stopped the vm before we issue the warning | |
355 | */ | |
356 | if (kvm_check_and_clear_guest_paused()) | |
357 | return HRTIMER_RESTART; | |
358 | ||
58687acb | 359 | /* only warn once */ |
b1a8de1f | 360 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
361 | /* | |
362 | * When multiple processes are causing softlockups the | |
363 | * softlockup detector only warns on the first one | |
364 | * because the code relies on a full quiet cycle to | |
365 | * re-arm. The second process prevents the quiet cycle | |
366 | * and never gets reported. Use task pointers to detect | |
367 | * this. | |
368 | */ | |
369 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
370 | current) { | |
371 | __this_cpu_write(soft_watchdog_warn, false); | |
372 | __touch_watchdog(); | |
373 | } | |
58687acb | 374 | return HRTIMER_RESTART; |
b1a8de1f | 375 | } |
58687acb | 376 | |
ed235875 AT |
377 | if (softlockup_all_cpu_backtrace) { |
378 | /* Prevent multiple soft-lockup reports if one cpu is already | |
379 | * engaged in dumping cpu back traces | |
380 | */ | |
381 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
382 | /* Someone else will report us. Let's give up */ | |
383 | __this_cpu_write(soft_watchdog_warn, true); | |
384 | return HRTIMER_RESTART; | |
385 | } | |
386 | } | |
387 | ||
656c3b79 | 388 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 389 | smp_processor_id(), duration, |
58687acb | 390 | current->comm, task_pid_nr(current)); |
b1a8de1f | 391 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
392 | print_modules(); |
393 | print_irqtrace_events(current); | |
394 | if (regs) | |
395 | show_regs(regs); | |
396 | else | |
397 | dump_stack(); | |
398 | ||
ed235875 AT |
399 | if (softlockup_all_cpu_backtrace) { |
400 | /* Avoid generating two back traces for current | |
401 | * given that one is already made above | |
402 | */ | |
403 | trigger_allbutself_cpu_backtrace(); | |
404 | ||
405 | clear_bit(0, &soft_lockup_nmi_warn); | |
406 | /* Barrier to sync with other cpus */ | |
407 | smp_mb__after_atomic(); | |
408 | } | |
409 | ||
69361eef | 410 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
411 | if (softlockup_panic) |
412 | panic("softlockup: hung tasks"); | |
909ea964 | 413 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 414 | } else |
909ea964 | 415 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
416 | |
417 | return HRTIMER_RESTART; | |
418 | } | |
419 | ||
bcd951cf TG |
420 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
421 | { | |
422 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 423 | |
bcd951cf TG |
424 | sched_setscheduler(current, policy, ¶m); |
425 | } | |
426 | ||
427 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 428 | { |
f7f66b05 | 429 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 430 | |
3935e895 BM |
431 | /* kick off the timer for the hardlockup detector */ |
432 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
433 | hrtimer->function = watchdog_timer_fn; | |
434 | ||
bcd951cf TG |
435 | /* Enable the perf event */ |
436 | watchdog_nmi_enable(cpu); | |
58687acb | 437 | |
58687acb | 438 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
0f34c400 | 439 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
58687acb DZ |
440 | HRTIMER_MODE_REL_PINNED); |
441 | ||
bcd951cf TG |
442 | /* initialize timestamp */ |
443 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
444 | __touch_watchdog(); | |
445 | } | |
58687acb | 446 | |
bcd951cf TG |
447 | static void watchdog_disable(unsigned int cpu) |
448 | { | |
f7f66b05 | 449 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 450 | |
bcd951cf TG |
451 | watchdog_set_prio(SCHED_NORMAL, 0); |
452 | hrtimer_cancel(hrtimer); | |
453 | /* disable the perf event */ | |
454 | watchdog_nmi_disable(cpu); | |
58687acb DZ |
455 | } |
456 | ||
b8900bc0 FW |
457 | static void watchdog_cleanup(unsigned int cpu, bool online) |
458 | { | |
459 | watchdog_disable(cpu); | |
460 | } | |
461 | ||
bcd951cf TG |
462 | static int watchdog_should_run(unsigned int cpu) |
463 | { | |
464 | return __this_cpu_read(hrtimer_interrupts) != | |
465 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
466 | } | |
467 | ||
468 | /* | |
469 | * The watchdog thread function - touches the timestamp. | |
470 | * | |
0f34c400 | 471 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
472 | * default) to reset the softlockup timestamp. If this gets delayed |
473 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
474 | * triggers in watchdog_timer_fn(). | |
475 | */ | |
476 | static void watchdog(unsigned int cpu) | |
477 | { | |
478 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
479 | __this_cpu_read(hrtimer_interrupts)); | |
480 | __touch_watchdog(); | |
bcfba4f4 UO |
481 | |
482 | /* | |
483 | * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the | |
484 | * failure path. Check for failures that can occur asynchronously - | |
485 | * for example, when CPUs are on-lined - and shut down the hardware | |
486 | * perf event on each CPU accordingly. | |
487 | * | |
488 | * The only non-obvious place this bit can be cleared is through | |
489 | * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a | |
490 | * pr_info here would be too noisy as it would result in a message | |
491 | * every few seconds if the hardlockup was disabled but the softlockup | |
492 | * enabled. | |
493 | */ | |
494 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
495 | watchdog_nmi_disable(cpu); | |
bcd951cf | 496 | } |
58687acb | 497 | |
b8900bc0 FW |
498 | static struct smp_hotplug_thread watchdog_threads = { |
499 | .store = &softlockup_watchdog, | |
500 | .thread_should_run = watchdog_should_run, | |
501 | .thread_fn = watchdog, | |
502 | .thread_comm = "watchdog/%u", | |
503 | .setup = watchdog_enable, | |
504 | .cleanup = watchdog_cleanup, | |
505 | .park = watchdog_disable, | |
506 | .unpark = watchdog_enable, | |
507 | }; | |
508 | ||
81a4beef UO |
509 | /* |
510 | * park all watchdog threads that are specified in 'watchdog_cpumask' | |
ee7fed54 UO |
511 | * |
512 | * This function returns an error if kthread_park() of a watchdog thread | |
513 | * fails. In this situation, the watchdog threads of some CPUs can already | |
514 | * be parked and the watchdog threads of other CPUs can still be runnable. | |
515 | * Callers are expected to handle this special condition as appropriate in | |
516 | * their context. | |
a2a45b85 UO |
517 | * |
518 | * This function may only be called in a context that is protected against | |
519 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
520 | */ |
521 | static int watchdog_park_threads(void) | |
522 | { | |
523 | int cpu, ret = 0; | |
524 | ||
b94f5118 DZ |
525 | atomic_set(&watchdog_park_in_progress, 1); |
526 | ||
81a4beef UO |
527 | for_each_watchdog_cpu(cpu) { |
528 | ret = kthread_park(per_cpu(softlockup_watchdog, cpu)); | |
529 | if (ret) | |
530 | break; | |
531 | } | |
81a4beef | 532 | |
b94f5118 DZ |
533 | atomic_set(&watchdog_park_in_progress, 0); |
534 | ||
81a4beef UO |
535 | return ret; |
536 | } | |
537 | ||
538 | /* | |
539 | * unpark all watchdog threads that are specified in 'watchdog_cpumask' | |
a2a45b85 UO |
540 | * |
541 | * This function may only be called in a context that is protected against | |
542 | * races with CPU hotplug - for example, via get_online_cpus(). | |
81a4beef UO |
543 | */ |
544 | static void watchdog_unpark_threads(void) | |
545 | { | |
546 | int cpu; | |
547 | ||
81a4beef UO |
548 | for_each_watchdog_cpu(cpu) |
549 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); | |
81a4beef UO |
550 | } |
551 | ||
b43cb43c | 552 | static int update_watchdog_all_cpus(void) |
9809b18f | 553 | { |
b43cb43c UO |
554 | int ret; |
555 | ||
556 | ret = watchdog_park_threads(); | |
557 | if (ret) | |
558 | return ret; | |
559 | ||
d4bdd0b2 | 560 | watchdog_unpark_threads(); |
b43cb43c UO |
561 | |
562 | return 0; | |
9809b18f MH |
563 | } |
564 | ||
b2f57c3a | 565 | static int watchdog_enable_all_cpus(void) |
58687acb | 566 | { |
b8900bc0 | 567 | int err = 0; |
58687acb | 568 | |
3c00ea82 | 569 | if (!watchdog_running) { |
230ec939 FW |
570 | err = smpboot_register_percpu_thread_cpumask(&watchdog_threads, |
571 | &watchdog_cpumask); | |
b8900bc0 FW |
572 | if (err) |
573 | pr_err("Failed to create watchdog threads, disabled\n"); | |
230ec939 | 574 | else |
3c00ea82 | 575 | watchdog_running = 1; |
b2f57c3a UO |
576 | } else { |
577 | /* | |
578 | * Enable/disable the lockup detectors or | |
579 | * change the sample period 'on the fly'. | |
580 | */ | |
b43cb43c UO |
581 | err = update_watchdog_all_cpus(); |
582 | ||
583 | if (err) { | |
584 | watchdog_disable_all_cpus(); | |
585 | pr_err("Failed to update lockup detectors, disabled\n"); | |
586 | } | |
bcd951cf | 587 | } |
b8900bc0 | 588 | |
b43cb43c UO |
589 | if (err) |
590 | watchdog_enabled = 0; | |
591 | ||
b8900bc0 | 592 | return err; |
58687acb DZ |
593 | } |
594 | ||
595 | static void watchdog_disable_all_cpus(void) | |
596 | { | |
3c00ea82 FW |
597 | if (watchdog_running) { |
598 | watchdog_running = 0; | |
b8900bc0 | 599 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 600 | } |
58687acb DZ |
601 | } |
602 | ||
05a4a952 NP |
603 | #else /* SOFTLOCKUP */ |
604 | static int watchdog_park_threads(void) | |
605 | { | |
606 | return 0; | |
607 | } | |
608 | ||
609 | static void watchdog_unpark_threads(void) | |
610 | { | |
611 | } | |
612 | ||
613 | static int watchdog_enable_all_cpus(void) | |
614 | { | |
615 | return 0; | |
616 | } | |
617 | ||
618 | static void watchdog_disable_all_cpus(void) | |
619 | { | |
620 | } | |
621 | ||
622 | static void set_sample_period(void) | |
623 | { | |
624 | } | |
625 | #endif /* SOFTLOCKUP */ | |
626 | ||
627 | /* | |
628 | * Suspend the hard and soft lockup detector by parking the watchdog threads. | |
629 | */ | |
630 | int lockup_detector_suspend(void) | |
631 | { | |
632 | int ret = 0; | |
633 | ||
634 | get_online_cpus(); | |
635 | mutex_lock(&watchdog_proc_mutex); | |
636 | /* | |
637 | * Multiple suspend requests can be active in parallel (counted by | |
638 | * the 'watchdog_suspended' variable). If the watchdog threads are | |
639 | * running, the first caller takes care that they will be parked. | |
640 | * The state of 'watchdog_running' cannot change while a suspend | |
641 | * request is active (see related code in 'proc' handlers). | |
642 | */ | |
643 | if (watchdog_running && !watchdog_suspended) | |
644 | ret = watchdog_park_threads(); | |
645 | ||
646 | if (ret == 0) | |
647 | watchdog_suspended++; | |
648 | else { | |
649 | watchdog_disable_all_cpus(); | |
650 | pr_err("Failed to suspend lockup detectors, disabled\n"); | |
651 | watchdog_enabled = 0; | |
652 | } | |
653 | ||
654 | mutex_unlock(&watchdog_proc_mutex); | |
655 | ||
656 | return ret; | |
657 | } | |
658 | ||
659 | /* | |
660 | * Resume the hard and soft lockup detector by unparking the watchdog threads. | |
661 | */ | |
662 | void lockup_detector_resume(void) | |
663 | { | |
664 | mutex_lock(&watchdog_proc_mutex); | |
665 | ||
666 | watchdog_suspended--; | |
667 | /* | |
668 | * The watchdog threads are unparked if they were previously running | |
669 | * and if there is no more active suspend request. | |
670 | */ | |
671 | if (watchdog_running && !watchdog_suspended) | |
672 | watchdog_unpark_threads(); | |
673 | ||
674 | mutex_unlock(&watchdog_proc_mutex); | |
675 | put_online_cpus(); | |
676 | } | |
677 | ||
58cf690a UO |
678 | #ifdef CONFIG_SYSCTL |
679 | ||
58687acb | 680 | /* |
a0c9cbb9 UO |
681 | * Update the run state of the lockup detectors. |
682 | */ | |
683 | static int proc_watchdog_update(void) | |
684 | { | |
685 | int err = 0; | |
686 | ||
687 | /* | |
688 | * Watchdog threads won't be started if they are already active. | |
689 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes | |
690 | * care of this. If those threads are already active, the sample | |
691 | * period will be updated and the lockup detectors will be enabled | |
692 | * or disabled 'on the fly'. | |
693 | */ | |
694 | if (watchdog_enabled && watchdog_thresh) | |
b2f57c3a | 695 | err = watchdog_enable_all_cpus(); |
a0c9cbb9 UO |
696 | else |
697 | watchdog_disable_all_cpus(); | |
698 | ||
699 | return err; | |
700 | ||
701 | } | |
702 | ||
ef246a21 UO |
703 | /* |
704 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
705 | * | |
706 | * caller | table->data points to | 'which' contains the flag(s) | |
707 | * -------------------|-----------------------|----------------------------- | |
708 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
709 | * | | with SOFT_WATCHDOG_ENABLED | |
710 | * -------------------|-----------------------|----------------------------- | |
711 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
712 | * -------------------|-----------------------|----------------------------- | |
713 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
714 | */ | |
715 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
716 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
717 | { | |
718 | int err, old, new; | |
719 | int *watchdog_param = (int *)table->data; | |
720 | ||
8614ddef | 721 | get_online_cpus(); |
ef246a21 UO |
722 | mutex_lock(&watchdog_proc_mutex); |
723 | ||
8c073d27 UO |
724 | if (watchdog_suspended) { |
725 | /* no parameter changes allowed while watchdog is suspended */ | |
726 | err = -EAGAIN; | |
727 | goto out; | |
728 | } | |
729 | ||
ef246a21 UO |
730 | /* |
731 | * If the parameter is being read return the state of the corresponding | |
732 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
733 | * run state of the lockup detectors. | |
734 | */ | |
735 | if (!write) { | |
736 | *watchdog_param = (watchdog_enabled & which) != 0; | |
737 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
738 | } else { | |
739 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
740 | if (err) | |
741 | goto out; | |
742 | ||
743 | /* | |
744 | * There is a race window between fetching the current value | |
745 | * from 'watchdog_enabled' and storing the new value. During | |
746 | * this race window, watchdog_nmi_enable() can sneak in and | |
747 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
748 | * The 'cmpxchg' detects this race and the loop retries. | |
749 | */ | |
750 | do { | |
751 | old = watchdog_enabled; | |
752 | /* | |
753 | * If the parameter value is not zero set the | |
754 | * corresponding bit(s), else clear it(them). | |
755 | */ | |
756 | if (*watchdog_param) | |
757 | new = old | which; | |
758 | else | |
759 | new = old & ~which; | |
760 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
761 | ||
762 | /* | |
b43cb43c UO |
763 | * Update the run state of the lockup detectors. There is _no_ |
764 | * need to check the value returned by proc_watchdog_update() | |
765 | * and to restore the previous value of 'watchdog_enabled' as | |
766 | * both lockup detectors are disabled if proc_watchdog_update() | |
767 | * returns an error. | |
ef246a21 | 768 | */ |
a1ee1932 JH |
769 | if (old == new) |
770 | goto out; | |
771 | ||
ef246a21 | 772 | err = proc_watchdog_update(); |
ef246a21 UO |
773 | } |
774 | out: | |
775 | mutex_unlock(&watchdog_proc_mutex); | |
8614ddef | 776 | put_online_cpus(); |
ef246a21 UO |
777 | return err; |
778 | } | |
779 | ||
83a80a39 UO |
780 | /* |
781 | * /proc/sys/kernel/watchdog | |
782 | */ | |
783 | int proc_watchdog(struct ctl_table *table, int write, | |
784 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
785 | { | |
786 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
787 | table, write, buffer, lenp, ppos); | |
788 | } | |
789 | ||
790 | /* | |
791 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 792 | */ |
83a80a39 UO |
793 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
794 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
795 | { | |
796 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
797 | table, write, buffer, lenp, ppos); | |
798 | } | |
799 | ||
800 | /* | |
801 | * /proc/sys/kernel/soft_watchdog | |
802 | */ | |
803 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
804 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
805 | { | |
806 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
807 | table, write, buffer, lenp, ppos); | |
808 | } | |
58687acb | 809 | |
83a80a39 UO |
810 | /* |
811 | * /proc/sys/kernel/watchdog_thresh | |
812 | */ | |
813 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
814 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 815 | { |
a1ee1932 | 816 | int err, old, new; |
58687acb | 817 | |
8614ddef | 818 | get_online_cpus(); |
359e6fab | 819 | mutex_lock(&watchdog_proc_mutex); |
bcd951cf | 820 | |
8c073d27 UO |
821 | if (watchdog_suspended) { |
822 | /* no parameter changes allowed while watchdog is suspended */ | |
823 | err = -EAGAIN; | |
824 | goto out; | |
825 | } | |
826 | ||
83a80a39 | 827 | old = ACCESS_ONCE(watchdog_thresh); |
b8900bc0 | 828 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 829 | |
b8900bc0 | 830 | if (err || !write) |
359e6fab | 831 | goto out; |
e04ab2bc | 832 | |
b66a2356 | 833 | /* |
d283c640 | 834 | * Update the sample period. Restore on failure. |
b66a2356 | 835 | */ |
a1ee1932 JH |
836 | new = ACCESS_ONCE(watchdog_thresh); |
837 | if (old == new) | |
838 | goto out; | |
839 | ||
83a80a39 UO |
840 | set_sample_period(); |
841 | err = proc_watchdog_update(); | |
d283c640 | 842 | if (err) { |
83a80a39 | 843 | watchdog_thresh = old; |
d283c640 UO |
844 | set_sample_period(); |
845 | } | |
359e6fab MH |
846 | out: |
847 | mutex_unlock(&watchdog_proc_mutex); | |
8614ddef | 848 | put_online_cpus(); |
b8900bc0 | 849 | return err; |
58687acb | 850 | } |
fe4ba3c3 CM |
851 | |
852 | /* | |
853 | * The cpumask is the mask of possible cpus that the watchdog can run | |
854 | * on, not the mask of cpus it is actually running on. This allows the | |
855 | * user to specify a mask that will include cpus that have not yet | |
856 | * been brought online, if desired. | |
857 | */ | |
858 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
859 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
860 | { | |
861 | int err; | |
862 | ||
8614ddef | 863 | get_online_cpus(); |
fe4ba3c3 | 864 | mutex_lock(&watchdog_proc_mutex); |
8c073d27 UO |
865 | |
866 | if (watchdog_suspended) { | |
867 | /* no parameter changes allowed while watchdog is suspended */ | |
868 | err = -EAGAIN; | |
869 | goto out; | |
870 | } | |
871 | ||
fe4ba3c3 CM |
872 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); |
873 | if (!err && write) { | |
874 | /* Remove impossible cpus to keep sysctl output cleaner. */ | |
875 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, | |
876 | cpu_possible_mask); | |
877 | ||
878 | if (watchdog_running) { | |
879 | /* | |
880 | * Failure would be due to being unable to allocate | |
881 | * a temporary cpumask, so we are likely not in a | |
882 | * position to do much else to make things better. | |
883 | */ | |
05a4a952 | 884 | #ifdef CONFIG_SOFTLOCKUP_DETECTOR |
fe4ba3c3 CM |
885 | if (smpboot_update_cpumask_percpu_thread( |
886 | &watchdog_threads, &watchdog_cpumask) != 0) | |
887 | pr_err("cpumask update failed\n"); | |
05a4a952 | 888 | #endif |
fe4ba3c3 CM |
889 | } |
890 | } | |
8c073d27 | 891 | out: |
fe4ba3c3 | 892 | mutex_unlock(&watchdog_proc_mutex); |
8614ddef | 893 | put_online_cpus(); |
fe4ba3c3 CM |
894 | return err; |
895 | } | |
896 | ||
58687acb DZ |
897 | #endif /* CONFIG_SYSCTL */ |
898 | ||
004417a6 | 899 | void __init lockup_detector_init(void) |
58687acb | 900 | { |
0f34c400 | 901 | set_sample_period(); |
b8900bc0 | 902 | |
fe4ba3c3 CM |
903 | #ifdef CONFIG_NO_HZ_FULL |
904 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
905 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
906 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
907 | } else |
908 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
909 | #else | |
910 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
911 | #endif | |
912 | ||
195daf66 | 913 | if (watchdog_enabled) |
b2f57c3a | 914 | watchdog_enable_all_cpus(); |
58687acb | 915 | } |