]>
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> |
58687acb DZ |
22 | |
23 | #include <asm/irq_regs.h> | |
5d1c0f4a | 24 | #include <linux/kvm_para.h> |
58687acb DZ |
25 | #include <linux/perf_event.h> |
26 | ||
3c00ea82 | 27 | int watchdog_user_enabled = 1; |
4eec42f3 | 28 | int __read_mostly watchdog_thresh = 10; |
ed235875 AT |
29 | #ifdef CONFIG_SMP |
30 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
31 | #else | |
32 | #define sysctl_softlockup_all_cpu_backtrace 0 | |
33 | #endif | |
34 | ||
3c00ea82 | 35 | static int __read_mostly watchdog_running; |
0f34c400 | 36 | static u64 __read_mostly sample_period; |
58687acb DZ |
37 | |
38 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
39 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
40 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
41 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 42 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
43 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
44 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
23637d47 | 45 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
cafcd80d DZ |
46 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
47 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); | |
58687acb DZ |
48 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
49 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); | |
50 | #endif | |
ed235875 | 51 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 52 | |
58687acb DZ |
53 | /* boot commands */ |
54 | /* | |
55 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
56 | */ | |
23637d47 | 57 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
fef2c9bc DZ |
58 | static int hardlockup_panic = |
59 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
58687acb DZ |
60 | |
61 | static int __init hardlockup_panic_setup(char *str) | |
62 | { | |
63 | if (!strncmp(str, "panic", 5)) | |
64 | hardlockup_panic = 1; | |
fef2c9bc DZ |
65 | else if (!strncmp(str, "nopanic", 7)) |
66 | hardlockup_panic = 0; | |
5dc30558 | 67 | else if (!strncmp(str, "0", 1)) |
3c00ea82 | 68 | watchdog_user_enabled = 0; |
58687acb DZ |
69 | return 1; |
70 | } | |
71 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
72 | #endif | |
73 | ||
74 | unsigned int __read_mostly softlockup_panic = | |
75 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
76 | ||
77 | static int __init softlockup_panic_setup(char *str) | |
78 | { | |
79 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
80 | ||
81 | return 1; | |
82 | } | |
83 | __setup("softlockup_panic=", softlockup_panic_setup); | |
84 | ||
85 | static int __init nowatchdog_setup(char *str) | |
86 | { | |
3c00ea82 | 87 | watchdog_user_enabled = 0; |
58687acb DZ |
88 | return 1; |
89 | } | |
90 | __setup("nowatchdog", nowatchdog_setup); | |
91 | ||
92 | /* deprecated */ | |
93 | static int __init nosoftlockup_setup(char *str) | |
94 | { | |
3c00ea82 | 95 | watchdog_user_enabled = 0; |
58687acb DZ |
96 | return 1; |
97 | } | |
98 | __setup("nosoftlockup", nosoftlockup_setup); | |
99 | /* */ | |
ed235875 AT |
100 | #ifdef CONFIG_SMP |
101 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
102 | { | |
103 | sysctl_softlockup_all_cpu_backtrace = | |
104 | !!simple_strtol(str, NULL, 0); | |
105 | return 1; | |
106 | } | |
107 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
108 | #endif | |
58687acb | 109 | |
4eec42f3 MSB |
110 | /* |
111 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
112 | * lockups can have false positives under extreme conditions. So we generally | |
113 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
114 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
115 | * time the hard threshold is. | |
116 | */ | |
6e9101ae | 117 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
118 | { |
119 | return watchdog_thresh * 2; | |
120 | } | |
58687acb DZ |
121 | |
122 | /* | |
123 | * Returns seconds, approximately. We don't need nanosecond | |
124 | * resolution, and we don't need to waste time with a big divide when | |
125 | * 2^30ns == 1.074s. | |
126 | */ | |
c06b4f19 | 127 | static unsigned long get_timestamp(void) |
58687acb | 128 | { |
c06b4f19 | 129 | return local_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
130 | } |
131 | ||
0f34c400 | 132 | static void set_sample_period(void) |
58687acb DZ |
133 | { |
134 | /* | |
586692a5 | 135 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
136 | * the divide by 5 is to give hrtimer several chances (two |
137 | * or three with the current relation between the soft | |
138 | * and hard thresholds) to increment before the | |
139 | * hardlockup detector generates a warning | |
58687acb | 140 | */ |
0f34c400 | 141 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
58687acb DZ |
142 | } |
143 | ||
144 | /* Commands for resetting the watchdog */ | |
145 | static void __touch_watchdog(void) | |
146 | { | |
c06b4f19 | 147 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
148 | } |
149 | ||
332fbdbc | 150 | void touch_softlockup_watchdog(void) |
58687acb | 151 | { |
7861144b AM |
152 | /* |
153 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
154 | * gets zeroed here, so use the raw_ operation. | |
155 | */ | |
156 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 157 | } |
0167c781 | 158 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 159 | |
332fbdbc | 160 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
161 | { |
162 | int cpu; | |
163 | ||
164 | /* | |
165 | * this is done lockless | |
166 | * do we care if a 0 races with a timestamp? | |
167 | * all it means is the softlock check starts one cycle later | |
168 | */ | |
169 | for_each_online_cpu(cpu) | |
170 | per_cpu(watchdog_touch_ts, cpu) = 0; | |
171 | } | |
172 | ||
cafcd80d | 173 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb DZ |
174 | void touch_nmi_watchdog(void) |
175 | { | |
62572e29 BZ |
176 | /* |
177 | * Using __raw here because some code paths have | |
178 | * preemption enabled. If preemption is enabled | |
179 | * then interrupts should be enabled too, in which | |
180 | * case we shouldn't have to worry about the watchdog | |
181 | * going off. | |
182 | */ | |
183 | __raw_get_cpu_var(watchdog_nmi_touch) = true; | |
332fbdbc | 184 | touch_softlockup_watchdog(); |
58687acb DZ |
185 | } |
186 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
187 | ||
cafcd80d DZ |
188 | #endif |
189 | ||
58687acb DZ |
190 | void touch_softlockup_watchdog_sync(void) |
191 | { | |
192 | __raw_get_cpu_var(softlockup_touch_sync) = true; | |
193 | __raw_get_cpu_var(watchdog_touch_ts) = 0; | |
194 | } | |
195 | ||
23637d47 | 196 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb | 197 | /* watchdog detector functions */ |
26e09c6e | 198 | static int is_hardlockup(void) |
58687acb | 199 | { |
909ea964 | 200 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
58687acb | 201 | |
909ea964 | 202 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
58687acb DZ |
203 | return 1; |
204 | ||
909ea964 | 205 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
58687acb DZ |
206 | return 0; |
207 | } | |
208 | #endif | |
209 | ||
26e09c6e | 210 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 211 | { |
c06b4f19 | 212 | unsigned long now = get_timestamp(); |
58687acb DZ |
213 | |
214 | /* Warn about unreasonable delays: */ | |
4eec42f3 | 215 | if (time_after(now, touch_ts + get_softlockup_thresh())) |
58687acb DZ |
216 | return now - touch_ts; |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
23637d47 | 221 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
1880c4ae | 222 | |
58687acb DZ |
223 | static struct perf_event_attr wd_hw_attr = { |
224 | .type = PERF_TYPE_HARDWARE, | |
225 | .config = PERF_COUNT_HW_CPU_CYCLES, | |
226 | .size = sizeof(struct perf_event_attr), | |
227 | .pinned = 1, | |
228 | .disabled = 1, | |
229 | }; | |
230 | ||
231 | /* Callback function for perf event subsystem */ | |
a8b0ca17 | 232 | static void watchdog_overflow_callback(struct perf_event *event, |
58687acb DZ |
233 | struct perf_sample_data *data, |
234 | struct pt_regs *regs) | |
235 | { | |
c6db67cd PZ |
236 | /* Ensure the watchdog never gets throttled */ |
237 | event->hw.interrupts = 0; | |
238 | ||
909ea964 CL |
239 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
240 | __this_cpu_write(watchdog_nmi_touch, false); | |
58687acb DZ |
241 | return; |
242 | } | |
243 | ||
244 | /* check for a hardlockup | |
245 | * This is done by making sure our timer interrupt | |
246 | * is incrementing. The timer interrupt should have | |
247 | * fired multiple times before we overflow'd. If it hasn't | |
248 | * then this is a good indication the cpu is stuck | |
249 | */ | |
26e09c6e DZ |
250 | if (is_hardlockup()) { |
251 | int this_cpu = smp_processor_id(); | |
252 | ||
58687acb | 253 | /* only print hardlockups once */ |
909ea964 | 254 | if (__this_cpu_read(hard_watchdog_warn) == true) |
58687acb DZ |
255 | return; |
256 | ||
257 | if (hardlockup_panic) | |
258 | panic("Watchdog detected hard LOCKUP on cpu %d", this_cpu); | |
259 | else | |
260 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", this_cpu); | |
261 | ||
909ea964 | 262 | __this_cpu_write(hard_watchdog_warn, true); |
58687acb DZ |
263 | return; |
264 | } | |
265 | ||
909ea964 | 266 | __this_cpu_write(hard_watchdog_warn, false); |
58687acb DZ |
267 | return; |
268 | } | |
bcd951cf TG |
269 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
270 | ||
58687acb DZ |
271 | static void watchdog_interrupt_count(void) |
272 | { | |
909ea964 | 273 | __this_cpu_inc(hrtimer_interrupts); |
58687acb | 274 | } |
bcd951cf TG |
275 | |
276 | static int watchdog_nmi_enable(unsigned int cpu); | |
277 | static void watchdog_nmi_disable(unsigned int cpu); | |
58687acb DZ |
278 | |
279 | /* watchdog kicker functions */ | |
280 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
281 | { | |
909ea964 | 282 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
283 | struct pt_regs *regs = get_irq_regs(); |
284 | int duration; | |
ed235875 | 285 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb DZ |
286 | |
287 | /* kick the hardlockup detector */ | |
288 | watchdog_interrupt_count(); | |
289 | ||
290 | /* kick the softlockup detector */ | |
909ea964 | 291 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
292 | |
293 | /* .. and repeat */ | |
0f34c400 | 294 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
295 | |
296 | if (touch_ts == 0) { | |
909ea964 | 297 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
298 | /* |
299 | * If the time stamp was touched atomically | |
300 | * make sure the scheduler tick is up to date. | |
301 | */ | |
909ea964 | 302 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
303 | sched_clock_tick(); |
304 | } | |
5d1c0f4a EM |
305 | |
306 | /* Clear the guest paused flag on watchdog reset */ | |
307 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
308 | __touch_watchdog(); |
309 | return HRTIMER_RESTART; | |
310 | } | |
311 | ||
312 | /* check for a softlockup | |
313 | * This is done by making sure a high priority task is | |
314 | * being scheduled. The task touches the watchdog to | |
315 | * indicate it is getting cpu time. If it hasn't then | |
316 | * this is a good indication some task is hogging the cpu | |
317 | */ | |
26e09c6e | 318 | duration = is_softlockup(touch_ts); |
58687acb | 319 | if (unlikely(duration)) { |
5d1c0f4a EM |
320 | /* |
321 | * If a virtual machine is stopped by the host it can look to | |
322 | * the watchdog like a soft lockup, check to see if the host | |
323 | * stopped the vm before we issue the warning | |
324 | */ | |
325 | if (kvm_check_and_clear_guest_paused()) | |
326 | return HRTIMER_RESTART; | |
327 | ||
58687acb | 328 | /* only warn once */ |
909ea964 | 329 | if (__this_cpu_read(soft_watchdog_warn) == true) |
58687acb DZ |
330 | return HRTIMER_RESTART; |
331 | ||
ed235875 AT |
332 | if (softlockup_all_cpu_backtrace) { |
333 | /* Prevent multiple soft-lockup reports if one cpu is already | |
334 | * engaged in dumping cpu back traces | |
335 | */ | |
336 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
337 | /* Someone else will report us. Let's give up */ | |
338 | __this_cpu_write(soft_watchdog_warn, true); | |
339 | return HRTIMER_RESTART; | |
340 | } | |
341 | } | |
342 | ||
b0f4c4b3 | 343 | printk(KERN_EMERG "BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 344 | smp_processor_id(), duration, |
58687acb DZ |
345 | current->comm, task_pid_nr(current)); |
346 | print_modules(); | |
347 | print_irqtrace_events(current); | |
348 | if (regs) | |
349 | show_regs(regs); | |
350 | else | |
351 | dump_stack(); | |
352 | ||
ed235875 AT |
353 | if (softlockup_all_cpu_backtrace) { |
354 | /* Avoid generating two back traces for current | |
355 | * given that one is already made above | |
356 | */ | |
357 | trigger_allbutself_cpu_backtrace(); | |
358 | ||
359 | clear_bit(0, &soft_lockup_nmi_warn); | |
360 | /* Barrier to sync with other cpus */ | |
361 | smp_mb__after_atomic(); | |
362 | } | |
363 | ||
58687acb DZ |
364 | if (softlockup_panic) |
365 | panic("softlockup: hung tasks"); | |
909ea964 | 366 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 367 | } else |
909ea964 | 368 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
369 | |
370 | return HRTIMER_RESTART; | |
371 | } | |
372 | ||
bcd951cf TG |
373 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
374 | { | |
375 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 376 | |
bcd951cf TG |
377 | sched_setscheduler(current, policy, ¶m); |
378 | } | |
379 | ||
380 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 381 | { |
26e09c6e | 382 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); |
58687acb | 383 | |
3935e895 BM |
384 | /* kick off the timer for the hardlockup detector */ |
385 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
386 | hrtimer->function = watchdog_timer_fn; | |
387 | ||
bcd951cf TG |
388 | /* Enable the perf event */ |
389 | watchdog_nmi_enable(cpu); | |
58687acb | 390 | |
58687acb | 391 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
0f34c400 | 392 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
58687acb DZ |
393 | HRTIMER_MODE_REL_PINNED); |
394 | ||
bcd951cf TG |
395 | /* initialize timestamp */ |
396 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
397 | __touch_watchdog(); | |
398 | } | |
58687acb | 399 | |
bcd951cf TG |
400 | static void watchdog_disable(unsigned int cpu) |
401 | { | |
402 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); | |
58687acb | 403 | |
bcd951cf TG |
404 | watchdog_set_prio(SCHED_NORMAL, 0); |
405 | hrtimer_cancel(hrtimer); | |
406 | /* disable the perf event */ | |
407 | watchdog_nmi_disable(cpu); | |
58687acb DZ |
408 | } |
409 | ||
b8900bc0 FW |
410 | static void watchdog_cleanup(unsigned int cpu, bool online) |
411 | { | |
412 | watchdog_disable(cpu); | |
413 | } | |
414 | ||
bcd951cf TG |
415 | static int watchdog_should_run(unsigned int cpu) |
416 | { | |
417 | return __this_cpu_read(hrtimer_interrupts) != | |
418 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
419 | } | |
420 | ||
421 | /* | |
422 | * The watchdog thread function - touches the timestamp. | |
423 | * | |
0f34c400 | 424 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
425 | * default) to reset the softlockup timestamp. If this gets delayed |
426 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
427 | * triggers in watchdog_timer_fn(). | |
428 | */ | |
429 | static void watchdog(unsigned int cpu) | |
430 | { | |
431 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
432 | __this_cpu_read(hrtimer_interrupts)); | |
433 | __touch_watchdog(); | |
434 | } | |
58687acb | 435 | |
23637d47 | 436 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
a7027046 DZ |
437 | /* |
438 | * People like the simple clean cpu node info on boot. | |
439 | * Reduce the watchdog noise by only printing messages | |
440 | * that are different from what cpu0 displayed. | |
441 | */ | |
442 | static unsigned long cpu0_err; | |
443 | ||
bcd951cf | 444 | static int watchdog_nmi_enable(unsigned int cpu) |
58687acb DZ |
445 | { |
446 | struct perf_event_attr *wd_attr; | |
447 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
448 | ||
449 | /* is it already setup and enabled? */ | |
450 | if (event && event->state > PERF_EVENT_STATE_OFF) | |
451 | goto out; | |
452 | ||
453 | /* it is setup but not enabled */ | |
454 | if (event != NULL) | |
455 | goto out_enable; | |
456 | ||
58687acb | 457 | wd_attr = &wd_hw_attr; |
4eec42f3 | 458 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); |
1880c4ae CG |
459 | |
460 | /* Try to register using hardware perf events */ | |
4dc0da86 | 461 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); |
a7027046 DZ |
462 | |
463 | /* save cpu0 error for future comparision */ | |
464 | if (cpu == 0 && IS_ERR(event)) | |
465 | cpu0_err = PTR_ERR(event); | |
466 | ||
58687acb | 467 | if (!IS_ERR(event)) { |
a7027046 DZ |
468 | /* only print for cpu0 or different than cpu0 */ |
469 | if (cpu == 0 || cpu0_err) | |
470 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); | |
58687acb DZ |
471 | goto out_save; |
472 | } | |
473 | ||
a7027046 DZ |
474 | /* skip displaying the same error again */ |
475 | if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) | |
476 | return PTR_ERR(event); | |
5651f7f4 DZ |
477 | |
478 | /* vary the KERN level based on the returned errno */ | |
479 | if (PTR_ERR(event) == -EOPNOTSUPP) | |
4501980a | 480 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); |
5651f7f4 | 481 | else if (PTR_ERR(event) == -ENOENT) |
4501980a AM |
482 | pr_warning("disabled (cpu%i): hardware events not enabled\n", |
483 | cpu); | |
5651f7f4 | 484 | else |
4501980a AM |
485 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", |
486 | cpu, PTR_ERR(event)); | |
eac24335 | 487 | return PTR_ERR(event); |
58687acb DZ |
488 | |
489 | /* success path */ | |
490 | out_save: | |
491 | per_cpu(watchdog_ev, cpu) = event; | |
492 | out_enable: | |
493 | perf_event_enable(per_cpu(watchdog_ev, cpu)); | |
494 | out: | |
495 | return 0; | |
496 | } | |
497 | ||
bcd951cf | 498 | static void watchdog_nmi_disable(unsigned int cpu) |
58687acb DZ |
499 | { |
500 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
501 | ||
502 | if (event) { | |
503 | perf_event_disable(event); | |
504 | per_cpu(watchdog_ev, cpu) = NULL; | |
505 | ||
506 | /* should be in cleanup, but blocks oprofile */ | |
507 | perf_event_release_kernel(event); | |
508 | } | |
509 | return; | |
510 | } | |
511 | #else | |
bcd951cf TG |
512 | static int watchdog_nmi_enable(unsigned int cpu) { return 0; } |
513 | static void watchdog_nmi_disable(unsigned int cpu) { return; } | |
23637d47 | 514 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
58687acb | 515 | |
b8900bc0 FW |
516 | static struct smp_hotplug_thread watchdog_threads = { |
517 | .store = &softlockup_watchdog, | |
518 | .thread_should_run = watchdog_should_run, | |
519 | .thread_fn = watchdog, | |
520 | .thread_comm = "watchdog/%u", | |
521 | .setup = watchdog_enable, | |
522 | .cleanup = watchdog_cleanup, | |
523 | .park = watchdog_disable, | |
524 | .unpark = watchdog_enable, | |
525 | }; | |
526 | ||
9809b18f MH |
527 | static void restart_watchdog_hrtimer(void *info) |
528 | { | |
529 | struct hrtimer *hrtimer = &__raw_get_cpu_var(watchdog_hrtimer); | |
530 | int ret; | |
531 | ||
532 | /* | |
533 | * No need to cancel and restart hrtimer if it is currently executing | |
534 | * because it will reprogram itself with the new period now. | |
535 | * We should never see it unqueued here because we are running per-cpu | |
536 | * with interrupts disabled. | |
537 | */ | |
538 | ret = hrtimer_try_to_cancel(hrtimer); | |
539 | if (ret == 1) | |
540 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), | |
541 | HRTIMER_MODE_REL_PINNED); | |
542 | } | |
543 | ||
544 | static void update_timers(int cpu) | |
545 | { | |
9809b18f MH |
546 | /* |
547 | * Make sure that perf event counter will adopt to a new | |
548 | * sampling period. Updating the sampling period directly would | |
549 | * be much nicer but we do not have an API for that now so | |
550 | * let's use a big hammer. | |
551 | * Hrtimer will adopt the new period on the next tick but this | |
552 | * might be late already so we have to restart the timer as well. | |
553 | */ | |
554 | watchdog_nmi_disable(cpu); | |
e0a23b06 | 555 | smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1); |
9809b18f MH |
556 | watchdog_nmi_enable(cpu); |
557 | } | |
558 | ||
559 | static void update_timers_all_cpus(void) | |
560 | { | |
561 | int cpu; | |
562 | ||
563 | get_online_cpus(); | |
9809b18f MH |
564 | for_each_online_cpu(cpu) |
565 | update_timers(cpu); | |
9809b18f MH |
566 | put_online_cpus(); |
567 | } | |
568 | ||
569 | static int watchdog_enable_all_cpus(bool sample_period_changed) | |
58687acb | 570 | { |
b8900bc0 | 571 | int err = 0; |
58687acb | 572 | |
3c00ea82 | 573 | if (!watchdog_running) { |
b8900bc0 FW |
574 | err = smpboot_register_percpu_thread(&watchdog_threads); |
575 | if (err) | |
576 | pr_err("Failed to create watchdog threads, disabled\n"); | |
577 | else | |
3c00ea82 | 578 | watchdog_running = 1; |
9809b18f MH |
579 | } else if (sample_period_changed) { |
580 | update_timers_all_cpus(); | |
bcd951cf | 581 | } |
b8900bc0 FW |
582 | |
583 | return err; | |
58687acb DZ |
584 | } |
585 | ||
b8900bc0 FW |
586 | /* prepare/enable/disable routines */ |
587 | /* sysctl functions */ | |
588 | #ifdef CONFIG_SYSCTL | |
58687acb DZ |
589 | static void watchdog_disable_all_cpus(void) |
590 | { | |
3c00ea82 FW |
591 | if (watchdog_running) { |
592 | watchdog_running = 0; | |
b8900bc0 | 593 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 594 | } |
58687acb DZ |
595 | } |
596 | ||
58687acb | 597 | /* |
586692a5 | 598 | * proc handler for /proc/sys/kernel/nmi_watchdog,watchdog_thresh |
58687acb DZ |
599 | */ |
600 | ||
586692a5 MSB |
601 | int proc_dowatchdog(struct ctl_table *table, int write, |
602 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 603 | { |
b8900bc0 | 604 | int err, old_thresh, old_enabled; |
359e6fab | 605 | static DEFINE_MUTEX(watchdog_proc_mutex); |
58687acb | 606 | |
359e6fab | 607 | mutex_lock(&watchdog_proc_mutex); |
b8900bc0 | 608 | old_thresh = ACCESS_ONCE(watchdog_thresh); |
3c00ea82 | 609 | old_enabled = ACCESS_ONCE(watchdog_user_enabled); |
bcd951cf | 610 | |
b8900bc0 FW |
611 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
612 | if (err || !write) | |
359e6fab | 613 | goto out; |
e04ab2bc | 614 | |
0f34c400 | 615 | set_sample_period(); |
b66a2356 | 616 | /* |
617 | * Watchdog threads shouldn't be enabled if they are | |
3c00ea82 | 618 | * disabled. The 'watchdog_running' variable check in |
b66a2356 | 619 | * watchdog_*_all_cpus() function takes care of this. |
620 | */ | |
3c00ea82 | 621 | if (watchdog_user_enabled && watchdog_thresh) |
9809b18f | 622 | err = watchdog_enable_all_cpus(old_thresh != watchdog_thresh); |
e04ab2bc MSB |
623 | else |
624 | watchdog_disable_all_cpus(); | |
625 | ||
b8900bc0 FW |
626 | /* Restore old values on failure */ |
627 | if (err) { | |
628 | watchdog_thresh = old_thresh; | |
3c00ea82 | 629 | watchdog_user_enabled = old_enabled; |
b8900bc0 | 630 | } |
359e6fab MH |
631 | out: |
632 | mutex_unlock(&watchdog_proc_mutex); | |
b8900bc0 | 633 | return err; |
58687acb | 634 | } |
58687acb DZ |
635 | #endif /* CONFIG_SYSCTL */ |
636 | ||
004417a6 | 637 | void __init lockup_detector_init(void) |
58687acb | 638 | { |
0f34c400 | 639 | set_sample_period(); |
b8900bc0 | 640 | |
3c00ea82 | 641 | if (watchdog_user_enabled) |
9809b18f | 642 | watchdog_enable_all_cpus(false); |
58687acb | 643 | } |