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 | ||
84d56e66 UO |
27 | /* |
28 | * The run state of the lockup detectors is controlled by the content of the | |
29 | * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit - | |
30 | * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector. | |
31 | * | |
32 | * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled' | |
33 | * are variables that are only used as an 'interface' between the parameters | |
34 | * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The | |
35 | * 'watchdog_thresh' variable is handled differently because its value is not | |
36 | * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh' | |
37 | * is equal zero. | |
38 | */ | |
39 | #define NMI_WATCHDOG_ENABLED_BIT 0 | |
40 | #define SOFT_WATCHDOG_ENABLED_BIT 1 | |
41 | #define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT) | |
42 | #define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT) | |
43 | ||
44 | #ifdef CONFIG_HARDLOCKUP_DETECTOR | |
45 | static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED; | |
46 | #else | |
47 | static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; | |
48 | #endif | |
49 | int __read_mostly nmi_watchdog_enabled; | |
50 | int __read_mostly soft_watchdog_enabled; | |
51 | int __read_mostly watchdog_user_enabled; | |
4eec42f3 | 52 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 53 | |
ed235875 AT |
54 | #ifdef CONFIG_SMP |
55 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
56 | #else | |
57 | #define sysctl_softlockup_all_cpu_backtrace 0 | |
58 | #endif | |
59 | ||
3c00ea82 | 60 | static int __read_mostly watchdog_running; |
0f34c400 | 61 | static u64 __read_mostly sample_period; |
58687acb DZ |
62 | |
63 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
64 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
65 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
66 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 67 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
68 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
69 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 70 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
23637d47 | 71 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
cafcd80d DZ |
72 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
73 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); | |
58687acb DZ |
74 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
75 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); | |
76 | #endif | |
ed235875 | 77 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 78 | |
58687acb DZ |
79 | /* boot commands */ |
80 | /* | |
81 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
82 | */ | |
23637d47 | 83 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
fef2c9bc DZ |
84 | static int hardlockup_panic = |
85 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
6e7458a6 UO |
86 | /* |
87 | * We may not want to enable hard lockup detection by default in all cases, | |
88 | * for example when running the kernel as a guest on a hypervisor. In these | |
89 | * cases this function can be called to disable hard lockup detection. This | |
90 | * function should only be executed once by the boot processor before the | |
91 | * kernel command line parameters are parsed, because otherwise it is not | |
92 | * possible to override this in hardlockup_panic_setup(). | |
93 | */ | |
692297d8 | 94 | void hardlockup_detector_disable(void) |
6e7458a6 | 95 | { |
692297d8 | 96 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
6e7458a6 UO |
97 | } |
98 | ||
58687acb DZ |
99 | static int __init hardlockup_panic_setup(char *str) |
100 | { | |
101 | if (!strncmp(str, "panic", 5)) | |
102 | hardlockup_panic = 1; | |
fef2c9bc DZ |
103 | else if (!strncmp(str, "nopanic", 7)) |
104 | hardlockup_panic = 0; | |
5dc30558 | 105 | else if (!strncmp(str, "0", 1)) |
195daf66 UO |
106 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
107 | else if (!strncmp(str, "1", 1)) | |
108 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
58687acb DZ |
109 | return 1; |
110 | } | |
111 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
112 | #endif | |
113 | ||
114 | unsigned int __read_mostly softlockup_panic = | |
115 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
116 | ||
117 | static int __init softlockup_panic_setup(char *str) | |
118 | { | |
119 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
120 | ||
121 | return 1; | |
122 | } | |
123 | __setup("softlockup_panic=", softlockup_panic_setup); | |
124 | ||
125 | static int __init nowatchdog_setup(char *str) | |
126 | { | |
195daf66 | 127 | watchdog_enabled = 0; |
58687acb DZ |
128 | return 1; |
129 | } | |
130 | __setup("nowatchdog", nowatchdog_setup); | |
131 | ||
58687acb DZ |
132 | static int __init nosoftlockup_setup(char *str) |
133 | { | |
195daf66 | 134 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
135 | return 1; |
136 | } | |
137 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 138 | |
ed235875 AT |
139 | #ifdef CONFIG_SMP |
140 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
141 | { | |
142 | sysctl_softlockup_all_cpu_backtrace = | |
143 | !!simple_strtol(str, NULL, 0); | |
144 | return 1; | |
145 | } | |
146 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
147 | #endif | |
58687acb | 148 | |
4eec42f3 MSB |
149 | /* |
150 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
151 | * lockups can have false positives under extreme conditions. So we generally | |
152 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
153 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
154 | * time the hard threshold is. | |
155 | */ | |
6e9101ae | 156 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
157 | { |
158 | return watchdog_thresh * 2; | |
159 | } | |
58687acb DZ |
160 | |
161 | /* | |
162 | * Returns seconds, approximately. We don't need nanosecond | |
163 | * resolution, and we don't need to waste time with a big divide when | |
164 | * 2^30ns == 1.074s. | |
165 | */ | |
c06b4f19 | 166 | static unsigned long get_timestamp(void) |
58687acb | 167 | { |
545a2bf7 | 168 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
169 | } |
170 | ||
0f34c400 | 171 | static void set_sample_period(void) |
58687acb DZ |
172 | { |
173 | /* | |
586692a5 | 174 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
175 | * the divide by 5 is to give hrtimer several chances (two |
176 | * or three with the current relation between the soft | |
177 | * and hard thresholds) to increment before the | |
178 | * hardlockup detector generates a warning | |
58687acb | 179 | */ |
0f34c400 | 180 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
58687acb DZ |
181 | } |
182 | ||
183 | /* Commands for resetting the watchdog */ | |
184 | static void __touch_watchdog(void) | |
185 | { | |
c06b4f19 | 186 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
187 | } |
188 | ||
332fbdbc | 189 | void touch_softlockup_watchdog(void) |
58687acb | 190 | { |
7861144b AM |
191 | /* |
192 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
193 | * gets zeroed here, so use the raw_ operation. | |
194 | */ | |
195 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 196 | } |
0167c781 | 197 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 198 | |
332fbdbc | 199 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
200 | { |
201 | int cpu; | |
202 | ||
203 | /* | |
204 | * this is done lockless | |
205 | * do we care if a 0 races with a timestamp? | |
206 | * all it means is the softlock check starts one cycle later | |
207 | */ | |
208 | for_each_online_cpu(cpu) | |
209 | per_cpu(watchdog_touch_ts, cpu) = 0; | |
210 | } | |
211 | ||
cafcd80d | 212 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb DZ |
213 | void touch_nmi_watchdog(void) |
214 | { | |
62572e29 BZ |
215 | /* |
216 | * Using __raw here because some code paths have | |
217 | * preemption enabled. If preemption is enabled | |
218 | * then interrupts should be enabled too, in which | |
219 | * case we shouldn't have to worry about the watchdog | |
220 | * going off. | |
221 | */ | |
f7f66b05 | 222 | raw_cpu_write(watchdog_nmi_touch, true); |
332fbdbc | 223 | touch_softlockup_watchdog(); |
58687acb DZ |
224 | } |
225 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
226 | ||
cafcd80d DZ |
227 | #endif |
228 | ||
58687acb DZ |
229 | void touch_softlockup_watchdog_sync(void) |
230 | { | |
f7f66b05 CL |
231 | __this_cpu_write(softlockup_touch_sync, true); |
232 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
233 | } |
234 | ||
23637d47 | 235 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb | 236 | /* watchdog detector functions */ |
26e09c6e | 237 | static int is_hardlockup(void) |
58687acb | 238 | { |
909ea964 | 239 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
58687acb | 240 | |
909ea964 | 241 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
58687acb DZ |
242 | return 1; |
243 | ||
909ea964 | 244 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
58687acb DZ |
245 | return 0; |
246 | } | |
247 | #endif | |
248 | ||
26e09c6e | 249 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 250 | { |
c06b4f19 | 251 | unsigned long now = get_timestamp(); |
58687acb | 252 | |
195daf66 UO |
253 | if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) { |
254 | /* Warn about unreasonable delays. */ | |
255 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
256 | return now - touch_ts; | |
257 | } | |
58687acb DZ |
258 | return 0; |
259 | } | |
260 | ||
23637d47 | 261 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
1880c4ae | 262 | |
58687acb DZ |
263 | static struct perf_event_attr wd_hw_attr = { |
264 | .type = PERF_TYPE_HARDWARE, | |
265 | .config = PERF_COUNT_HW_CPU_CYCLES, | |
266 | .size = sizeof(struct perf_event_attr), | |
267 | .pinned = 1, | |
268 | .disabled = 1, | |
269 | }; | |
270 | ||
271 | /* Callback function for perf event subsystem */ | |
a8b0ca17 | 272 | static void watchdog_overflow_callback(struct perf_event *event, |
58687acb DZ |
273 | struct perf_sample_data *data, |
274 | struct pt_regs *regs) | |
275 | { | |
c6db67cd PZ |
276 | /* Ensure the watchdog never gets throttled */ |
277 | event->hw.interrupts = 0; | |
278 | ||
909ea964 CL |
279 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
280 | __this_cpu_write(watchdog_nmi_touch, false); | |
58687acb DZ |
281 | return; |
282 | } | |
283 | ||
284 | /* check for a hardlockup | |
285 | * This is done by making sure our timer interrupt | |
286 | * is incrementing. The timer interrupt should have | |
287 | * fired multiple times before we overflow'd. If it hasn't | |
288 | * then this is a good indication the cpu is stuck | |
289 | */ | |
26e09c6e DZ |
290 | if (is_hardlockup()) { |
291 | int this_cpu = smp_processor_id(); | |
292 | ||
58687acb | 293 | /* only print hardlockups once */ |
909ea964 | 294 | if (__this_cpu_read(hard_watchdog_warn) == true) |
58687acb DZ |
295 | return; |
296 | ||
297 | if (hardlockup_panic) | |
656c3b79 FF |
298 | panic("Watchdog detected hard LOCKUP on cpu %d", |
299 | this_cpu); | |
58687acb | 300 | else |
656c3b79 FF |
301 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", |
302 | this_cpu); | |
58687acb | 303 | |
909ea964 | 304 | __this_cpu_write(hard_watchdog_warn, true); |
58687acb DZ |
305 | return; |
306 | } | |
307 | ||
909ea964 | 308 | __this_cpu_write(hard_watchdog_warn, false); |
58687acb DZ |
309 | return; |
310 | } | |
bcd951cf TG |
311 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
312 | ||
58687acb DZ |
313 | static void watchdog_interrupt_count(void) |
314 | { | |
909ea964 | 315 | __this_cpu_inc(hrtimer_interrupts); |
58687acb | 316 | } |
bcd951cf TG |
317 | |
318 | static int watchdog_nmi_enable(unsigned int cpu); | |
319 | static void watchdog_nmi_disable(unsigned int cpu); | |
58687acb DZ |
320 | |
321 | /* watchdog kicker functions */ | |
322 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
323 | { | |
909ea964 | 324 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
325 | struct pt_regs *regs = get_irq_regs(); |
326 | int duration; | |
ed235875 | 327 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb DZ |
328 | |
329 | /* kick the hardlockup detector */ | |
330 | watchdog_interrupt_count(); | |
331 | ||
332 | /* kick the softlockup detector */ | |
909ea964 | 333 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
334 | |
335 | /* .. and repeat */ | |
0f34c400 | 336 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
337 | |
338 | if (touch_ts == 0) { | |
909ea964 | 339 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
340 | /* |
341 | * If the time stamp was touched atomically | |
342 | * make sure the scheduler tick is up to date. | |
343 | */ | |
909ea964 | 344 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
345 | sched_clock_tick(); |
346 | } | |
5d1c0f4a EM |
347 | |
348 | /* Clear the guest paused flag on watchdog reset */ | |
349 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
350 | __touch_watchdog(); |
351 | return HRTIMER_RESTART; | |
352 | } | |
353 | ||
354 | /* check for a softlockup | |
355 | * This is done by making sure a high priority task is | |
356 | * being scheduled. The task touches the watchdog to | |
357 | * indicate it is getting cpu time. If it hasn't then | |
358 | * this is a good indication some task is hogging the cpu | |
359 | */ | |
26e09c6e | 360 | duration = is_softlockup(touch_ts); |
58687acb | 361 | if (unlikely(duration)) { |
5d1c0f4a EM |
362 | /* |
363 | * If a virtual machine is stopped by the host it can look to | |
364 | * the watchdog like a soft lockup, check to see if the host | |
365 | * stopped the vm before we issue the warning | |
366 | */ | |
367 | if (kvm_check_and_clear_guest_paused()) | |
368 | return HRTIMER_RESTART; | |
369 | ||
58687acb | 370 | /* only warn once */ |
b1a8de1f | 371 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
372 | /* | |
373 | * When multiple processes are causing softlockups the | |
374 | * softlockup detector only warns on the first one | |
375 | * because the code relies on a full quiet cycle to | |
376 | * re-arm. The second process prevents the quiet cycle | |
377 | * and never gets reported. Use task pointers to detect | |
378 | * this. | |
379 | */ | |
380 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
381 | current) { | |
382 | __this_cpu_write(soft_watchdog_warn, false); | |
383 | __touch_watchdog(); | |
384 | } | |
58687acb | 385 | return HRTIMER_RESTART; |
b1a8de1f | 386 | } |
58687acb | 387 | |
ed235875 AT |
388 | if (softlockup_all_cpu_backtrace) { |
389 | /* Prevent multiple soft-lockup reports if one cpu is already | |
390 | * engaged in dumping cpu back traces | |
391 | */ | |
392 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
393 | /* Someone else will report us. Let's give up */ | |
394 | __this_cpu_write(soft_watchdog_warn, true); | |
395 | return HRTIMER_RESTART; | |
396 | } | |
397 | } | |
398 | ||
656c3b79 | 399 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 400 | smp_processor_id(), duration, |
58687acb | 401 | current->comm, task_pid_nr(current)); |
b1a8de1f | 402 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
403 | print_modules(); |
404 | print_irqtrace_events(current); | |
405 | if (regs) | |
406 | show_regs(regs); | |
407 | else | |
408 | dump_stack(); | |
409 | ||
ed235875 AT |
410 | if (softlockup_all_cpu_backtrace) { |
411 | /* Avoid generating two back traces for current | |
412 | * given that one is already made above | |
413 | */ | |
414 | trigger_allbutself_cpu_backtrace(); | |
415 | ||
416 | clear_bit(0, &soft_lockup_nmi_warn); | |
417 | /* Barrier to sync with other cpus */ | |
418 | smp_mb__after_atomic(); | |
419 | } | |
420 | ||
69361eef | 421 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
422 | if (softlockup_panic) |
423 | panic("softlockup: hung tasks"); | |
909ea964 | 424 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 425 | } else |
909ea964 | 426 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
427 | |
428 | return HRTIMER_RESTART; | |
429 | } | |
430 | ||
bcd951cf TG |
431 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
432 | { | |
433 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 434 | |
bcd951cf TG |
435 | sched_setscheduler(current, policy, ¶m); |
436 | } | |
437 | ||
438 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 439 | { |
f7f66b05 | 440 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 441 | |
3935e895 BM |
442 | /* kick off the timer for the hardlockup detector */ |
443 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
444 | hrtimer->function = watchdog_timer_fn; | |
445 | ||
bcd951cf TG |
446 | /* Enable the perf event */ |
447 | watchdog_nmi_enable(cpu); | |
58687acb | 448 | |
58687acb | 449 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
0f34c400 | 450 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
58687acb DZ |
451 | HRTIMER_MODE_REL_PINNED); |
452 | ||
bcd951cf TG |
453 | /* initialize timestamp */ |
454 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
455 | __touch_watchdog(); | |
456 | } | |
58687acb | 457 | |
bcd951cf TG |
458 | static void watchdog_disable(unsigned int cpu) |
459 | { | |
f7f66b05 | 460 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 461 | |
bcd951cf TG |
462 | watchdog_set_prio(SCHED_NORMAL, 0); |
463 | hrtimer_cancel(hrtimer); | |
464 | /* disable the perf event */ | |
465 | watchdog_nmi_disable(cpu); | |
58687acb DZ |
466 | } |
467 | ||
b8900bc0 FW |
468 | static void watchdog_cleanup(unsigned int cpu, bool online) |
469 | { | |
470 | watchdog_disable(cpu); | |
471 | } | |
472 | ||
bcd951cf TG |
473 | static int watchdog_should_run(unsigned int cpu) |
474 | { | |
475 | return __this_cpu_read(hrtimer_interrupts) != | |
476 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
477 | } | |
478 | ||
479 | /* | |
480 | * The watchdog thread function - touches the timestamp. | |
481 | * | |
0f34c400 | 482 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
483 | * default) to reset the softlockup timestamp. If this gets delayed |
484 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
485 | * triggers in watchdog_timer_fn(). | |
486 | */ | |
487 | static void watchdog(unsigned int cpu) | |
488 | { | |
489 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
490 | __this_cpu_read(hrtimer_interrupts)); | |
491 | __touch_watchdog(); | |
bcfba4f4 UO |
492 | |
493 | /* | |
494 | * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the | |
495 | * failure path. Check for failures that can occur asynchronously - | |
496 | * for example, when CPUs are on-lined - and shut down the hardware | |
497 | * perf event on each CPU accordingly. | |
498 | * | |
499 | * The only non-obvious place this bit can be cleared is through | |
500 | * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a | |
501 | * pr_info here would be too noisy as it would result in a message | |
502 | * every few seconds if the hardlockup was disabled but the softlockup | |
503 | * enabled. | |
504 | */ | |
505 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
506 | watchdog_nmi_disable(cpu); | |
bcd951cf | 507 | } |
58687acb | 508 | |
23637d47 | 509 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
a7027046 DZ |
510 | /* |
511 | * People like the simple clean cpu node info on boot. | |
512 | * Reduce the watchdog noise by only printing messages | |
513 | * that are different from what cpu0 displayed. | |
514 | */ | |
515 | static unsigned long cpu0_err; | |
516 | ||
bcd951cf | 517 | static int watchdog_nmi_enable(unsigned int cpu) |
58687acb DZ |
518 | { |
519 | struct perf_event_attr *wd_attr; | |
520 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
521 | ||
195daf66 UO |
522 | /* nothing to do if the hard lockup detector is disabled */ |
523 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
524 | goto out; | |
6e7458a6 | 525 | |
58687acb DZ |
526 | /* is it already setup and enabled? */ |
527 | if (event && event->state > PERF_EVENT_STATE_OFF) | |
528 | goto out; | |
529 | ||
530 | /* it is setup but not enabled */ | |
531 | if (event != NULL) | |
532 | goto out_enable; | |
533 | ||
58687acb | 534 | wd_attr = &wd_hw_attr; |
4eec42f3 | 535 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); |
1880c4ae CG |
536 | |
537 | /* Try to register using hardware perf events */ | |
4dc0da86 | 538 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); |
a7027046 DZ |
539 | |
540 | /* save cpu0 error for future comparision */ | |
541 | if (cpu == 0 && IS_ERR(event)) | |
542 | cpu0_err = PTR_ERR(event); | |
543 | ||
58687acb | 544 | if (!IS_ERR(event)) { |
a7027046 DZ |
545 | /* only print for cpu0 or different than cpu0 */ |
546 | if (cpu == 0 || cpu0_err) | |
547 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); | |
58687acb DZ |
548 | goto out_save; |
549 | } | |
550 | ||
bcfba4f4 UO |
551 | /* |
552 | * Disable the hard lockup detector if _any_ CPU fails to set up | |
553 | * set up the hardware perf event. The watchdog() function checks | |
554 | * the NMI_WATCHDOG_ENABLED bit periodically. | |
555 | * | |
556 | * The barriers are for syncing up watchdog_enabled across all the | |
557 | * cpus, as clear_bit() does not use barriers. | |
558 | */ | |
559 | smp_mb__before_atomic(); | |
560 | clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled); | |
561 | smp_mb__after_atomic(); | |
562 | ||
a7027046 DZ |
563 | /* skip displaying the same error again */ |
564 | if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) | |
565 | return PTR_ERR(event); | |
5651f7f4 DZ |
566 | |
567 | /* vary the KERN level based on the returned errno */ | |
568 | if (PTR_ERR(event) == -EOPNOTSUPP) | |
4501980a | 569 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); |
5651f7f4 | 570 | else if (PTR_ERR(event) == -ENOENT) |
656c3b79 | 571 | pr_warn("disabled (cpu%i): hardware events not enabled\n", |
4501980a | 572 | cpu); |
5651f7f4 | 573 | else |
4501980a AM |
574 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", |
575 | cpu, PTR_ERR(event)); | |
bcfba4f4 UO |
576 | |
577 | pr_info("Shutting down hard lockup detector on all cpus\n"); | |
578 | ||
eac24335 | 579 | return PTR_ERR(event); |
58687acb DZ |
580 | |
581 | /* success path */ | |
582 | out_save: | |
583 | per_cpu(watchdog_ev, cpu) = event; | |
584 | out_enable: | |
585 | perf_event_enable(per_cpu(watchdog_ev, cpu)); | |
586 | out: | |
587 | return 0; | |
588 | } | |
589 | ||
bcd951cf | 590 | static void watchdog_nmi_disable(unsigned int cpu) |
58687acb DZ |
591 | { |
592 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
593 | ||
594 | if (event) { | |
595 | perf_event_disable(event); | |
596 | per_cpu(watchdog_ev, cpu) = NULL; | |
597 | ||
598 | /* should be in cleanup, but blocks oprofile */ | |
599 | perf_event_release_kernel(event); | |
600 | } | |
df577149 UO |
601 | if (cpu == 0) { |
602 | /* watchdog_nmi_enable() expects this to be zero initially. */ | |
603 | cpu0_err = 0; | |
604 | } | |
58687acb | 605 | } |
b3738d29 SE |
606 | |
607 | void watchdog_nmi_enable_all(void) | |
608 | { | |
609 | int cpu; | |
610 | ||
611 | if (!watchdog_user_enabled) | |
612 | return; | |
613 | ||
614 | get_online_cpus(); | |
615 | for_each_online_cpu(cpu) | |
616 | watchdog_nmi_enable(cpu); | |
617 | put_online_cpus(); | |
618 | } | |
619 | ||
620 | void watchdog_nmi_disable_all(void) | |
621 | { | |
622 | int cpu; | |
623 | ||
624 | if (!watchdog_running) | |
625 | return; | |
626 | ||
627 | get_online_cpus(); | |
628 | for_each_online_cpu(cpu) | |
629 | watchdog_nmi_disable(cpu); | |
630 | put_online_cpus(); | |
631 | } | |
58687acb | 632 | #else |
bcd951cf TG |
633 | static int watchdog_nmi_enable(unsigned int cpu) { return 0; } |
634 | static void watchdog_nmi_disable(unsigned int cpu) { return; } | |
b3738d29 SE |
635 | void watchdog_nmi_enable_all(void) {} |
636 | void watchdog_nmi_disable_all(void) {} | |
23637d47 | 637 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
58687acb | 638 | |
b8900bc0 FW |
639 | static struct smp_hotplug_thread watchdog_threads = { |
640 | .store = &softlockup_watchdog, | |
641 | .thread_should_run = watchdog_should_run, | |
642 | .thread_fn = watchdog, | |
643 | .thread_comm = "watchdog/%u", | |
644 | .setup = watchdog_enable, | |
645 | .cleanup = watchdog_cleanup, | |
646 | .park = watchdog_disable, | |
647 | .unpark = watchdog_enable, | |
648 | }; | |
649 | ||
9809b18f MH |
650 | static void restart_watchdog_hrtimer(void *info) |
651 | { | |
f7f66b05 | 652 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
9809b18f MH |
653 | int ret; |
654 | ||
655 | /* | |
656 | * No need to cancel and restart hrtimer if it is currently executing | |
657 | * because it will reprogram itself with the new period now. | |
658 | * We should never see it unqueued here because we are running per-cpu | |
659 | * with interrupts disabled. | |
660 | */ | |
661 | ret = hrtimer_try_to_cancel(hrtimer); | |
662 | if (ret == 1) | |
663 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), | |
664 | HRTIMER_MODE_REL_PINNED); | |
665 | } | |
666 | ||
b2f57c3a | 667 | static void update_watchdog(int cpu) |
9809b18f | 668 | { |
9809b18f MH |
669 | /* |
670 | * Make sure that perf event counter will adopt to a new | |
671 | * sampling period. Updating the sampling period directly would | |
672 | * be much nicer but we do not have an API for that now so | |
673 | * let's use a big hammer. | |
674 | * Hrtimer will adopt the new period on the next tick but this | |
675 | * might be late already so we have to restart the timer as well. | |
676 | */ | |
677 | watchdog_nmi_disable(cpu); | |
e0a23b06 | 678 | smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1); |
9809b18f MH |
679 | watchdog_nmi_enable(cpu); |
680 | } | |
681 | ||
b2f57c3a | 682 | static void update_watchdog_all_cpus(void) |
9809b18f MH |
683 | { |
684 | int cpu; | |
685 | ||
686 | get_online_cpus(); | |
9809b18f | 687 | for_each_online_cpu(cpu) |
b2f57c3a | 688 | update_watchdog(cpu); |
9809b18f MH |
689 | put_online_cpus(); |
690 | } | |
691 | ||
b2f57c3a | 692 | static int watchdog_enable_all_cpus(void) |
58687acb | 693 | { |
b8900bc0 | 694 | int err = 0; |
58687acb | 695 | |
3c00ea82 | 696 | if (!watchdog_running) { |
b8900bc0 FW |
697 | err = smpboot_register_percpu_thread(&watchdog_threads); |
698 | if (err) | |
699 | pr_err("Failed to create watchdog threads, disabled\n"); | |
700 | else | |
3c00ea82 | 701 | watchdog_running = 1; |
b2f57c3a UO |
702 | } else { |
703 | /* | |
704 | * Enable/disable the lockup detectors or | |
705 | * change the sample period 'on the fly'. | |
706 | */ | |
707 | update_watchdog_all_cpus(); | |
bcd951cf | 708 | } |
b8900bc0 FW |
709 | |
710 | return err; | |
58687acb DZ |
711 | } |
712 | ||
b8900bc0 FW |
713 | /* prepare/enable/disable routines */ |
714 | /* sysctl functions */ | |
715 | #ifdef CONFIG_SYSCTL | |
58687acb DZ |
716 | static void watchdog_disable_all_cpus(void) |
717 | { | |
3c00ea82 FW |
718 | if (watchdog_running) { |
719 | watchdog_running = 0; | |
b8900bc0 | 720 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 721 | } |
58687acb DZ |
722 | } |
723 | ||
58687acb | 724 | /* |
a0c9cbb9 UO |
725 | * Update the run state of the lockup detectors. |
726 | */ | |
727 | static int proc_watchdog_update(void) | |
728 | { | |
729 | int err = 0; | |
730 | ||
731 | /* | |
732 | * Watchdog threads won't be started if they are already active. | |
733 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes | |
734 | * care of this. If those threads are already active, the sample | |
735 | * period will be updated and the lockup detectors will be enabled | |
736 | * or disabled 'on the fly'. | |
737 | */ | |
738 | if (watchdog_enabled && watchdog_thresh) | |
b2f57c3a | 739 | err = watchdog_enable_all_cpus(); |
a0c9cbb9 UO |
740 | else |
741 | watchdog_disable_all_cpus(); | |
742 | ||
743 | return err; | |
744 | ||
745 | } | |
746 | ||
f54c2274 UO |
747 | static DEFINE_MUTEX(watchdog_proc_mutex); |
748 | ||
ef246a21 UO |
749 | /* |
750 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
751 | * | |
752 | * caller | table->data points to | 'which' contains the flag(s) | |
753 | * -------------------|-----------------------|----------------------------- | |
754 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
755 | * | | with SOFT_WATCHDOG_ENABLED | |
756 | * -------------------|-----------------------|----------------------------- | |
757 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
758 | * -------------------|-----------------------|----------------------------- | |
759 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
760 | */ | |
761 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
762 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
763 | { | |
764 | int err, old, new; | |
765 | int *watchdog_param = (int *)table->data; | |
766 | ||
767 | mutex_lock(&watchdog_proc_mutex); | |
768 | ||
769 | /* | |
770 | * If the parameter is being read return the state of the corresponding | |
771 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
772 | * run state of the lockup detectors. | |
773 | */ | |
774 | if (!write) { | |
775 | *watchdog_param = (watchdog_enabled & which) != 0; | |
776 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
777 | } else { | |
778 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
779 | if (err) | |
780 | goto out; | |
781 | ||
782 | /* | |
783 | * There is a race window between fetching the current value | |
784 | * from 'watchdog_enabled' and storing the new value. During | |
785 | * this race window, watchdog_nmi_enable() can sneak in and | |
786 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
787 | * The 'cmpxchg' detects this race and the loop retries. | |
788 | */ | |
789 | do { | |
790 | old = watchdog_enabled; | |
791 | /* | |
792 | * If the parameter value is not zero set the | |
793 | * corresponding bit(s), else clear it(them). | |
794 | */ | |
795 | if (*watchdog_param) | |
796 | new = old | which; | |
797 | else | |
798 | new = old & ~which; | |
799 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
800 | ||
801 | /* | |
802 | * Update the run state of the lockup detectors. | |
803 | * Restore 'watchdog_enabled' on failure. | |
804 | */ | |
805 | err = proc_watchdog_update(); | |
806 | if (err) | |
807 | watchdog_enabled = old; | |
808 | } | |
809 | out: | |
810 | mutex_unlock(&watchdog_proc_mutex); | |
811 | return err; | |
812 | } | |
813 | ||
83a80a39 UO |
814 | /* |
815 | * /proc/sys/kernel/watchdog | |
816 | */ | |
817 | int proc_watchdog(struct ctl_table *table, int write, | |
818 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
819 | { | |
820 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
821 | table, write, buffer, lenp, ppos); | |
822 | } | |
823 | ||
824 | /* | |
825 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 826 | */ |
83a80a39 UO |
827 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
828 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
829 | { | |
830 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
831 | table, write, buffer, lenp, ppos); | |
832 | } | |
833 | ||
834 | /* | |
835 | * /proc/sys/kernel/soft_watchdog | |
836 | */ | |
837 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
838 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
839 | { | |
840 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
841 | table, write, buffer, lenp, ppos); | |
842 | } | |
58687acb | 843 | |
83a80a39 UO |
844 | /* |
845 | * /proc/sys/kernel/watchdog_thresh | |
846 | */ | |
847 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
848 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 849 | { |
83a80a39 | 850 | int err, old; |
58687acb | 851 | |
359e6fab | 852 | mutex_lock(&watchdog_proc_mutex); |
bcd951cf | 853 | |
83a80a39 | 854 | old = ACCESS_ONCE(watchdog_thresh); |
b8900bc0 | 855 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 856 | |
b8900bc0 | 857 | if (err || !write) |
359e6fab | 858 | goto out; |
e04ab2bc | 859 | |
b66a2356 | 860 | /* |
83a80a39 UO |
861 | * Update the sample period. |
862 | * Restore 'watchdog_thresh' on failure. | |
b66a2356 | 863 | */ |
83a80a39 UO |
864 | set_sample_period(); |
865 | err = proc_watchdog_update(); | |
866 | if (err) | |
867 | watchdog_thresh = old; | |
359e6fab MH |
868 | out: |
869 | mutex_unlock(&watchdog_proc_mutex); | |
b8900bc0 | 870 | return err; |
58687acb | 871 | } |
58687acb DZ |
872 | #endif /* CONFIG_SYSCTL */ |
873 | ||
004417a6 | 874 | void __init lockup_detector_init(void) |
58687acb | 875 | { |
0f34c400 | 876 | set_sample_period(); |
b8900bc0 | 877 | |
195daf66 | 878 | if (watchdog_enabled) |
b2f57c3a | 879 | watchdog_enable_all_cpus(); |
58687acb | 880 | } |