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> |
fe4ba3c3 | 22 | #include <linux/tick.h> |
58687acb DZ |
23 | |
24 | #include <asm/irq_regs.h> | |
5d1c0f4a | 25 | #include <linux/kvm_para.h> |
58687acb | 26 | #include <linux/perf_event.h> |
81a4beef | 27 | #include <linux/kthread.h> |
58687acb | 28 | |
84d56e66 UO |
29 | /* |
30 | * The run state of the lockup detectors is controlled by the content of the | |
31 | * 'watchdog_enabled' variable. Each lockup detector has its dedicated bit - | |
32 | * bit 0 for the hard lockup detector and bit 1 for the soft lockup detector. | |
33 | * | |
34 | * 'watchdog_user_enabled', 'nmi_watchdog_enabled' and 'soft_watchdog_enabled' | |
35 | * are variables that are only used as an 'interface' between the parameters | |
36 | * in /proc/sys/kernel and the internal state bits in 'watchdog_enabled'. The | |
37 | * 'watchdog_thresh' variable is handled differently because its value is not | |
38 | * boolean, and the lockup detectors are 'suspended' while 'watchdog_thresh' | |
39 | * is equal zero. | |
40 | */ | |
41 | #define NMI_WATCHDOG_ENABLED_BIT 0 | |
42 | #define SOFT_WATCHDOG_ENABLED_BIT 1 | |
43 | #define NMI_WATCHDOG_ENABLED (1 << NMI_WATCHDOG_ENABLED_BIT) | |
44 | #define SOFT_WATCHDOG_ENABLED (1 << SOFT_WATCHDOG_ENABLED_BIT) | |
45 | ||
ab992dc3 PZ |
46 | static DEFINE_MUTEX(watchdog_proc_mutex); |
47 | ||
84d56e66 UO |
48 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
49 | static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED|NMI_WATCHDOG_ENABLED; | |
50 | #else | |
51 | static unsigned long __read_mostly watchdog_enabled = SOFT_WATCHDOG_ENABLED; | |
52 | #endif | |
53 | int __read_mostly nmi_watchdog_enabled; | |
54 | int __read_mostly soft_watchdog_enabled; | |
55 | int __read_mostly watchdog_user_enabled; | |
4eec42f3 | 56 | int __read_mostly watchdog_thresh = 10; |
84d56e66 | 57 | |
ed235875 AT |
58 | #ifdef CONFIG_SMP |
59 | int __read_mostly sysctl_softlockup_all_cpu_backtrace; | |
60 | #else | |
61 | #define sysctl_softlockup_all_cpu_backtrace 0 | |
62 | #endif | |
fe4ba3c3 CM |
63 | static struct cpumask watchdog_cpumask __read_mostly; |
64 | unsigned long *watchdog_cpumask_bits = cpumask_bits(&watchdog_cpumask); | |
65 | ||
66 | /* Helper for online, unparked cpus. */ | |
67 | #define for_each_watchdog_cpu(cpu) \ | |
68 | for_each_cpu_and((cpu), cpu_online_mask, &watchdog_cpumask) | |
ed235875 | 69 | |
3c00ea82 | 70 | static int __read_mostly watchdog_running; |
0f34c400 | 71 | static u64 __read_mostly sample_period; |
58687acb DZ |
72 | |
73 | static DEFINE_PER_CPU(unsigned long, watchdog_touch_ts); | |
74 | static DEFINE_PER_CPU(struct task_struct *, softlockup_watchdog); | |
75 | static DEFINE_PER_CPU(struct hrtimer, watchdog_hrtimer); | |
76 | static DEFINE_PER_CPU(bool, softlockup_touch_sync); | |
58687acb | 77 | static DEFINE_PER_CPU(bool, soft_watchdog_warn); |
bcd951cf TG |
78 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts); |
79 | static DEFINE_PER_CPU(unsigned long, soft_lockup_hrtimer_cnt); | |
b1a8de1f | 80 | static DEFINE_PER_CPU(struct task_struct *, softlockup_task_ptr_saved); |
23637d47 | 81 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
cafcd80d DZ |
82 | static DEFINE_PER_CPU(bool, hard_watchdog_warn); |
83 | static DEFINE_PER_CPU(bool, watchdog_nmi_touch); | |
58687acb DZ |
84 | static DEFINE_PER_CPU(unsigned long, hrtimer_interrupts_saved); |
85 | static DEFINE_PER_CPU(struct perf_event *, watchdog_ev); | |
86 | #endif | |
ed235875 | 87 | static unsigned long soft_lockup_nmi_warn; |
58687acb | 88 | |
58687acb DZ |
89 | /* boot commands */ |
90 | /* | |
91 | * Should we panic when a soft-lockup or hard-lockup occurs: | |
92 | */ | |
23637d47 | 93 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
fef2c9bc DZ |
94 | static int hardlockup_panic = |
95 | CONFIG_BOOTPARAM_HARDLOCKUP_PANIC_VALUE; | |
6e7458a6 UO |
96 | /* |
97 | * We may not want to enable hard lockup detection by default in all cases, | |
98 | * for example when running the kernel as a guest on a hypervisor. In these | |
99 | * cases this function can be called to disable hard lockup detection. This | |
100 | * function should only be executed once by the boot processor before the | |
101 | * kernel command line parameters are parsed, because otherwise it is not | |
102 | * possible to override this in hardlockup_panic_setup(). | |
103 | */ | |
692297d8 | 104 | void hardlockup_detector_disable(void) |
6e7458a6 | 105 | { |
692297d8 | 106 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
6e7458a6 UO |
107 | } |
108 | ||
58687acb DZ |
109 | static int __init hardlockup_panic_setup(char *str) |
110 | { | |
111 | if (!strncmp(str, "panic", 5)) | |
112 | hardlockup_panic = 1; | |
fef2c9bc DZ |
113 | else if (!strncmp(str, "nopanic", 7)) |
114 | hardlockup_panic = 0; | |
5dc30558 | 115 | else if (!strncmp(str, "0", 1)) |
195daf66 UO |
116 | watchdog_enabled &= ~NMI_WATCHDOG_ENABLED; |
117 | else if (!strncmp(str, "1", 1)) | |
118 | watchdog_enabled |= NMI_WATCHDOG_ENABLED; | |
58687acb DZ |
119 | return 1; |
120 | } | |
121 | __setup("nmi_watchdog=", hardlockup_panic_setup); | |
122 | #endif | |
123 | ||
124 | unsigned int __read_mostly softlockup_panic = | |
125 | CONFIG_BOOTPARAM_SOFTLOCKUP_PANIC_VALUE; | |
126 | ||
127 | static int __init softlockup_panic_setup(char *str) | |
128 | { | |
129 | softlockup_panic = simple_strtoul(str, NULL, 0); | |
130 | ||
131 | return 1; | |
132 | } | |
133 | __setup("softlockup_panic=", softlockup_panic_setup); | |
134 | ||
135 | static int __init nowatchdog_setup(char *str) | |
136 | { | |
195daf66 | 137 | watchdog_enabled = 0; |
58687acb DZ |
138 | return 1; |
139 | } | |
140 | __setup("nowatchdog", nowatchdog_setup); | |
141 | ||
58687acb DZ |
142 | static int __init nosoftlockup_setup(char *str) |
143 | { | |
195daf66 | 144 | watchdog_enabled &= ~SOFT_WATCHDOG_ENABLED; |
58687acb DZ |
145 | return 1; |
146 | } | |
147 | __setup("nosoftlockup", nosoftlockup_setup); | |
195daf66 | 148 | |
ed235875 AT |
149 | #ifdef CONFIG_SMP |
150 | static int __init softlockup_all_cpu_backtrace_setup(char *str) | |
151 | { | |
152 | sysctl_softlockup_all_cpu_backtrace = | |
153 | !!simple_strtol(str, NULL, 0); | |
154 | return 1; | |
155 | } | |
156 | __setup("softlockup_all_cpu_backtrace=", softlockup_all_cpu_backtrace_setup); | |
157 | #endif | |
58687acb | 158 | |
4eec42f3 MSB |
159 | /* |
160 | * Hard-lockup warnings should be triggered after just a few seconds. Soft- | |
161 | * lockups can have false positives under extreme conditions. So we generally | |
162 | * want a higher threshold for soft lockups than for hard lockups. So we couple | |
163 | * the thresholds with a factor: we make the soft threshold twice the amount of | |
164 | * time the hard threshold is. | |
165 | */ | |
6e9101ae | 166 | static int get_softlockup_thresh(void) |
4eec42f3 MSB |
167 | { |
168 | return watchdog_thresh * 2; | |
169 | } | |
58687acb DZ |
170 | |
171 | /* | |
172 | * Returns seconds, approximately. We don't need nanosecond | |
173 | * resolution, and we don't need to waste time with a big divide when | |
174 | * 2^30ns == 1.074s. | |
175 | */ | |
c06b4f19 | 176 | static unsigned long get_timestamp(void) |
58687acb | 177 | { |
545a2bf7 | 178 | return running_clock() >> 30LL; /* 2^30 ~= 10^9 */ |
58687acb DZ |
179 | } |
180 | ||
0f34c400 | 181 | static void set_sample_period(void) |
58687acb DZ |
182 | { |
183 | /* | |
586692a5 | 184 | * convert watchdog_thresh from seconds to ns |
86f5e6a7 FLVC |
185 | * the divide by 5 is to give hrtimer several chances (two |
186 | * or three with the current relation between the soft | |
187 | * and hard thresholds) to increment before the | |
188 | * hardlockup detector generates a warning | |
58687acb | 189 | */ |
0f34c400 | 190 | sample_period = get_softlockup_thresh() * ((u64)NSEC_PER_SEC / 5); |
58687acb DZ |
191 | } |
192 | ||
193 | /* Commands for resetting the watchdog */ | |
194 | static void __touch_watchdog(void) | |
195 | { | |
c06b4f19 | 196 | __this_cpu_write(watchdog_touch_ts, get_timestamp()); |
58687acb DZ |
197 | } |
198 | ||
332fbdbc | 199 | void touch_softlockup_watchdog(void) |
58687acb | 200 | { |
7861144b AM |
201 | /* |
202 | * Preemption can be enabled. It doesn't matter which CPU's timestamp | |
203 | * gets zeroed here, so use the raw_ operation. | |
204 | */ | |
205 | raw_cpu_write(watchdog_touch_ts, 0); | |
58687acb | 206 | } |
0167c781 | 207 | EXPORT_SYMBOL(touch_softlockup_watchdog); |
58687acb | 208 | |
332fbdbc | 209 | void touch_all_softlockup_watchdogs(void) |
58687acb DZ |
210 | { |
211 | int cpu; | |
212 | ||
213 | /* | |
214 | * this is done lockless | |
215 | * do we care if a 0 races with a timestamp? | |
216 | * all it means is the softlock check starts one cycle later | |
217 | */ | |
fe4ba3c3 | 218 | for_each_watchdog_cpu(cpu) |
58687acb DZ |
219 | per_cpu(watchdog_touch_ts, cpu) = 0; |
220 | } | |
221 | ||
cafcd80d | 222 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb DZ |
223 | void touch_nmi_watchdog(void) |
224 | { | |
62572e29 BZ |
225 | /* |
226 | * Using __raw here because some code paths have | |
227 | * preemption enabled. If preemption is enabled | |
228 | * then interrupts should be enabled too, in which | |
229 | * case we shouldn't have to worry about the watchdog | |
230 | * going off. | |
231 | */ | |
f7f66b05 | 232 | raw_cpu_write(watchdog_nmi_touch, true); |
332fbdbc | 233 | touch_softlockup_watchdog(); |
58687acb DZ |
234 | } |
235 | EXPORT_SYMBOL(touch_nmi_watchdog); | |
236 | ||
cafcd80d DZ |
237 | #endif |
238 | ||
58687acb DZ |
239 | void touch_softlockup_watchdog_sync(void) |
240 | { | |
f7f66b05 CL |
241 | __this_cpu_write(softlockup_touch_sync, true); |
242 | __this_cpu_write(watchdog_touch_ts, 0); | |
58687acb DZ |
243 | } |
244 | ||
23637d47 | 245 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
58687acb | 246 | /* watchdog detector functions */ |
26e09c6e | 247 | static int is_hardlockup(void) |
58687acb | 248 | { |
909ea964 | 249 | unsigned long hrint = __this_cpu_read(hrtimer_interrupts); |
58687acb | 250 | |
909ea964 | 251 | if (__this_cpu_read(hrtimer_interrupts_saved) == hrint) |
58687acb DZ |
252 | return 1; |
253 | ||
909ea964 | 254 | __this_cpu_write(hrtimer_interrupts_saved, hrint); |
58687acb DZ |
255 | return 0; |
256 | } | |
257 | #endif | |
258 | ||
26e09c6e | 259 | static int is_softlockup(unsigned long touch_ts) |
58687acb | 260 | { |
c06b4f19 | 261 | unsigned long now = get_timestamp(); |
58687acb | 262 | |
195daf66 UO |
263 | if (watchdog_enabled & SOFT_WATCHDOG_ENABLED) { |
264 | /* Warn about unreasonable delays. */ | |
265 | if (time_after(now, touch_ts + get_softlockup_thresh())) | |
266 | return now - touch_ts; | |
267 | } | |
58687acb DZ |
268 | return 0; |
269 | } | |
270 | ||
23637d47 | 271 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
1880c4ae | 272 | |
58687acb DZ |
273 | static struct perf_event_attr wd_hw_attr = { |
274 | .type = PERF_TYPE_HARDWARE, | |
275 | .config = PERF_COUNT_HW_CPU_CYCLES, | |
276 | .size = sizeof(struct perf_event_attr), | |
277 | .pinned = 1, | |
278 | .disabled = 1, | |
279 | }; | |
280 | ||
281 | /* Callback function for perf event subsystem */ | |
a8b0ca17 | 282 | static void watchdog_overflow_callback(struct perf_event *event, |
58687acb DZ |
283 | struct perf_sample_data *data, |
284 | struct pt_regs *regs) | |
285 | { | |
c6db67cd PZ |
286 | /* Ensure the watchdog never gets throttled */ |
287 | event->hw.interrupts = 0; | |
288 | ||
909ea964 CL |
289 | if (__this_cpu_read(watchdog_nmi_touch) == true) { |
290 | __this_cpu_write(watchdog_nmi_touch, false); | |
58687acb DZ |
291 | return; |
292 | } | |
293 | ||
294 | /* check for a hardlockup | |
295 | * This is done by making sure our timer interrupt | |
296 | * is incrementing. The timer interrupt should have | |
297 | * fired multiple times before we overflow'd. If it hasn't | |
298 | * then this is a good indication the cpu is stuck | |
299 | */ | |
26e09c6e DZ |
300 | if (is_hardlockup()) { |
301 | int this_cpu = smp_processor_id(); | |
302 | ||
58687acb | 303 | /* only print hardlockups once */ |
909ea964 | 304 | if (__this_cpu_read(hard_watchdog_warn) == true) |
58687acb DZ |
305 | return; |
306 | ||
307 | if (hardlockup_panic) | |
656c3b79 FF |
308 | panic("Watchdog detected hard LOCKUP on cpu %d", |
309 | this_cpu); | |
58687acb | 310 | else |
656c3b79 FF |
311 | WARN(1, "Watchdog detected hard LOCKUP on cpu %d", |
312 | this_cpu); | |
58687acb | 313 | |
909ea964 | 314 | __this_cpu_write(hard_watchdog_warn, true); |
58687acb DZ |
315 | return; |
316 | } | |
317 | ||
909ea964 | 318 | __this_cpu_write(hard_watchdog_warn, false); |
58687acb DZ |
319 | return; |
320 | } | |
bcd951cf TG |
321 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
322 | ||
58687acb DZ |
323 | static void watchdog_interrupt_count(void) |
324 | { | |
909ea964 | 325 | __this_cpu_inc(hrtimer_interrupts); |
58687acb | 326 | } |
bcd951cf TG |
327 | |
328 | static int watchdog_nmi_enable(unsigned int cpu); | |
329 | static void watchdog_nmi_disable(unsigned int cpu); | |
58687acb DZ |
330 | |
331 | /* watchdog kicker functions */ | |
332 | static enum hrtimer_restart watchdog_timer_fn(struct hrtimer *hrtimer) | |
333 | { | |
909ea964 | 334 | unsigned long touch_ts = __this_cpu_read(watchdog_touch_ts); |
58687acb DZ |
335 | struct pt_regs *regs = get_irq_regs(); |
336 | int duration; | |
ed235875 | 337 | int softlockup_all_cpu_backtrace = sysctl_softlockup_all_cpu_backtrace; |
58687acb DZ |
338 | |
339 | /* kick the hardlockup detector */ | |
340 | watchdog_interrupt_count(); | |
341 | ||
342 | /* kick the softlockup detector */ | |
909ea964 | 343 | wake_up_process(__this_cpu_read(softlockup_watchdog)); |
58687acb DZ |
344 | |
345 | /* .. and repeat */ | |
0f34c400 | 346 | hrtimer_forward_now(hrtimer, ns_to_ktime(sample_period)); |
58687acb DZ |
347 | |
348 | if (touch_ts == 0) { | |
909ea964 | 349 | if (unlikely(__this_cpu_read(softlockup_touch_sync))) { |
58687acb DZ |
350 | /* |
351 | * If the time stamp was touched atomically | |
352 | * make sure the scheduler tick is up to date. | |
353 | */ | |
909ea964 | 354 | __this_cpu_write(softlockup_touch_sync, false); |
58687acb DZ |
355 | sched_clock_tick(); |
356 | } | |
5d1c0f4a EM |
357 | |
358 | /* Clear the guest paused flag on watchdog reset */ | |
359 | kvm_check_and_clear_guest_paused(); | |
58687acb DZ |
360 | __touch_watchdog(); |
361 | return HRTIMER_RESTART; | |
362 | } | |
363 | ||
364 | /* check for a softlockup | |
365 | * This is done by making sure a high priority task is | |
366 | * being scheduled. The task touches the watchdog to | |
367 | * indicate it is getting cpu time. If it hasn't then | |
368 | * this is a good indication some task is hogging the cpu | |
369 | */ | |
26e09c6e | 370 | duration = is_softlockup(touch_ts); |
58687acb | 371 | if (unlikely(duration)) { |
5d1c0f4a EM |
372 | /* |
373 | * If a virtual machine is stopped by the host it can look to | |
374 | * the watchdog like a soft lockup, check to see if the host | |
375 | * stopped the vm before we issue the warning | |
376 | */ | |
377 | if (kvm_check_and_clear_guest_paused()) | |
378 | return HRTIMER_RESTART; | |
379 | ||
58687acb | 380 | /* only warn once */ |
b1a8de1f | 381 | if (__this_cpu_read(soft_watchdog_warn) == true) { |
382 | /* | |
383 | * When multiple processes are causing softlockups the | |
384 | * softlockup detector only warns on the first one | |
385 | * because the code relies on a full quiet cycle to | |
386 | * re-arm. The second process prevents the quiet cycle | |
387 | * and never gets reported. Use task pointers to detect | |
388 | * this. | |
389 | */ | |
390 | if (__this_cpu_read(softlockup_task_ptr_saved) != | |
391 | current) { | |
392 | __this_cpu_write(soft_watchdog_warn, false); | |
393 | __touch_watchdog(); | |
394 | } | |
58687acb | 395 | return HRTIMER_RESTART; |
b1a8de1f | 396 | } |
58687acb | 397 | |
ed235875 AT |
398 | if (softlockup_all_cpu_backtrace) { |
399 | /* Prevent multiple soft-lockup reports if one cpu is already | |
400 | * engaged in dumping cpu back traces | |
401 | */ | |
402 | if (test_and_set_bit(0, &soft_lockup_nmi_warn)) { | |
403 | /* Someone else will report us. Let's give up */ | |
404 | __this_cpu_write(soft_watchdog_warn, true); | |
405 | return HRTIMER_RESTART; | |
406 | } | |
407 | } | |
408 | ||
656c3b79 | 409 | pr_emerg("BUG: soft lockup - CPU#%d stuck for %us! [%s:%d]\n", |
26e09c6e | 410 | smp_processor_id(), duration, |
58687acb | 411 | current->comm, task_pid_nr(current)); |
b1a8de1f | 412 | __this_cpu_write(softlockup_task_ptr_saved, current); |
58687acb DZ |
413 | print_modules(); |
414 | print_irqtrace_events(current); | |
415 | if (regs) | |
416 | show_regs(regs); | |
417 | else | |
418 | dump_stack(); | |
419 | ||
ed235875 AT |
420 | if (softlockup_all_cpu_backtrace) { |
421 | /* Avoid generating two back traces for current | |
422 | * given that one is already made above | |
423 | */ | |
424 | trigger_allbutself_cpu_backtrace(); | |
425 | ||
426 | clear_bit(0, &soft_lockup_nmi_warn); | |
427 | /* Barrier to sync with other cpus */ | |
428 | smp_mb__after_atomic(); | |
429 | } | |
430 | ||
69361eef | 431 | add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK); |
58687acb DZ |
432 | if (softlockup_panic) |
433 | panic("softlockup: hung tasks"); | |
909ea964 | 434 | __this_cpu_write(soft_watchdog_warn, true); |
58687acb | 435 | } else |
909ea964 | 436 | __this_cpu_write(soft_watchdog_warn, false); |
58687acb DZ |
437 | |
438 | return HRTIMER_RESTART; | |
439 | } | |
440 | ||
bcd951cf TG |
441 | static void watchdog_set_prio(unsigned int policy, unsigned int prio) |
442 | { | |
443 | struct sched_param param = { .sched_priority = prio }; | |
58687acb | 444 | |
bcd951cf TG |
445 | sched_setscheduler(current, policy, ¶m); |
446 | } | |
447 | ||
448 | static void watchdog_enable(unsigned int cpu) | |
58687acb | 449 | { |
f7f66b05 | 450 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 451 | |
3935e895 BM |
452 | /* kick off the timer for the hardlockup detector */ |
453 | hrtimer_init(hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
454 | hrtimer->function = watchdog_timer_fn; | |
455 | ||
bcd951cf TG |
456 | /* Enable the perf event */ |
457 | watchdog_nmi_enable(cpu); | |
58687acb | 458 | |
58687acb | 459 | /* done here because hrtimer_start can only pin to smp_processor_id() */ |
0f34c400 | 460 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), |
58687acb DZ |
461 | HRTIMER_MODE_REL_PINNED); |
462 | ||
bcd951cf TG |
463 | /* initialize timestamp */ |
464 | watchdog_set_prio(SCHED_FIFO, MAX_RT_PRIO - 1); | |
465 | __touch_watchdog(); | |
466 | } | |
58687acb | 467 | |
bcd951cf TG |
468 | static void watchdog_disable(unsigned int cpu) |
469 | { | |
f7f66b05 | 470 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
58687acb | 471 | |
bcd951cf TG |
472 | watchdog_set_prio(SCHED_NORMAL, 0); |
473 | hrtimer_cancel(hrtimer); | |
474 | /* disable the perf event */ | |
475 | watchdog_nmi_disable(cpu); | |
58687acb DZ |
476 | } |
477 | ||
b8900bc0 FW |
478 | static void watchdog_cleanup(unsigned int cpu, bool online) |
479 | { | |
480 | watchdog_disable(cpu); | |
481 | } | |
482 | ||
bcd951cf TG |
483 | static int watchdog_should_run(unsigned int cpu) |
484 | { | |
485 | return __this_cpu_read(hrtimer_interrupts) != | |
486 | __this_cpu_read(soft_lockup_hrtimer_cnt); | |
487 | } | |
488 | ||
489 | /* | |
490 | * The watchdog thread function - touches the timestamp. | |
491 | * | |
0f34c400 | 492 | * It only runs once every sample_period seconds (4 seconds by |
bcd951cf TG |
493 | * default) to reset the softlockup timestamp. If this gets delayed |
494 | * for more than 2*watchdog_thresh seconds then the debug-printout | |
495 | * triggers in watchdog_timer_fn(). | |
496 | */ | |
497 | static void watchdog(unsigned int cpu) | |
498 | { | |
499 | __this_cpu_write(soft_lockup_hrtimer_cnt, | |
500 | __this_cpu_read(hrtimer_interrupts)); | |
501 | __touch_watchdog(); | |
bcfba4f4 UO |
502 | |
503 | /* | |
504 | * watchdog_nmi_enable() clears the NMI_WATCHDOG_ENABLED bit in the | |
505 | * failure path. Check for failures that can occur asynchronously - | |
506 | * for example, when CPUs are on-lined - and shut down the hardware | |
507 | * perf event on each CPU accordingly. | |
508 | * | |
509 | * The only non-obvious place this bit can be cleared is through | |
510 | * watchdog_nmi_enable(), so a pr_info() is placed there. Placing a | |
511 | * pr_info here would be too noisy as it would result in a message | |
512 | * every few seconds if the hardlockup was disabled but the softlockup | |
513 | * enabled. | |
514 | */ | |
515 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
516 | watchdog_nmi_disable(cpu); | |
bcd951cf | 517 | } |
58687acb | 518 | |
23637d47 | 519 | #ifdef CONFIG_HARDLOCKUP_DETECTOR |
a7027046 DZ |
520 | /* |
521 | * People like the simple clean cpu node info on boot. | |
522 | * Reduce the watchdog noise by only printing messages | |
523 | * that are different from what cpu0 displayed. | |
524 | */ | |
525 | static unsigned long cpu0_err; | |
526 | ||
bcd951cf | 527 | static int watchdog_nmi_enable(unsigned int cpu) |
58687acb DZ |
528 | { |
529 | struct perf_event_attr *wd_attr; | |
530 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
531 | ||
195daf66 UO |
532 | /* nothing to do if the hard lockup detector is disabled */ |
533 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
534 | goto out; | |
6e7458a6 | 535 | |
58687acb DZ |
536 | /* is it already setup and enabled? */ |
537 | if (event && event->state > PERF_EVENT_STATE_OFF) | |
538 | goto out; | |
539 | ||
540 | /* it is setup but not enabled */ | |
541 | if (event != NULL) | |
542 | goto out_enable; | |
543 | ||
58687acb | 544 | wd_attr = &wd_hw_attr; |
4eec42f3 | 545 | wd_attr->sample_period = hw_nmi_get_sample_period(watchdog_thresh); |
1880c4ae CG |
546 | |
547 | /* Try to register using hardware perf events */ | |
4dc0da86 | 548 | event = perf_event_create_kernel_counter(wd_attr, cpu, NULL, watchdog_overflow_callback, NULL); |
a7027046 DZ |
549 | |
550 | /* save cpu0 error for future comparision */ | |
551 | if (cpu == 0 && IS_ERR(event)) | |
552 | cpu0_err = PTR_ERR(event); | |
553 | ||
58687acb | 554 | if (!IS_ERR(event)) { |
a7027046 DZ |
555 | /* only print for cpu0 or different than cpu0 */ |
556 | if (cpu == 0 || cpu0_err) | |
557 | pr_info("enabled on all CPUs, permanently consumes one hw-PMU counter.\n"); | |
58687acb DZ |
558 | goto out_save; |
559 | } | |
560 | ||
bcfba4f4 UO |
561 | /* |
562 | * Disable the hard lockup detector if _any_ CPU fails to set up | |
563 | * set up the hardware perf event. The watchdog() function checks | |
564 | * the NMI_WATCHDOG_ENABLED bit periodically. | |
565 | * | |
566 | * The barriers are for syncing up watchdog_enabled across all the | |
567 | * cpus, as clear_bit() does not use barriers. | |
568 | */ | |
569 | smp_mb__before_atomic(); | |
570 | clear_bit(NMI_WATCHDOG_ENABLED_BIT, &watchdog_enabled); | |
571 | smp_mb__after_atomic(); | |
572 | ||
a7027046 DZ |
573 | /* skip displaying the same error again */ |
574 | if (cpu > 0 && (PTR_ERR(event) == cpu0_err)) | |
575 | return PTR_ERR(event); | |
5651f7f4 DZ |
576 | |
577 | /* vary the KERN level based on the returned errno */ | |
578 | if (PTR_ERR(event) == -EOPNOTSUPP) | |
4501980a | 579 | pr_info("disabled (cpu%i): not supported (no LAPIC?)\n", cpu); |
5651f7f4 | 580 | else if (PTR_ERR(event) == -ENOENT) |
656c3b79 | 581 | pr_warn("disabled (cpu%i): hardware events not enabled\n", |
4501980a | 582 | cpu); |
5651f7f4 | 583 | else |
4501980a AM |
584 | pr_err("disabled (cpu%i): unable to create perf event: %ld\n", |
585 | cpu, PTR_ERR(event)); | |
bcfba4f4 UO |
586 | |
587 | pr_info("Shutting down hard lockup detector on all cpus\n"); | |
588 | ||
eac24335 | 589 | return PTR_ERR(event); |
58687acb DZ |
590 | |
591 | /* success path */ | |
592 | out_save: | |
593 | per_cpu(watchdog_ev, cpu) = event; | |
594 | out_enable: | |
595 | perf_event_enable(per_cpu(watchdog_ev, cpu)); | |
596 | out: | |
597 | return 0; | |
598 | } | |
599 | ||
bcd951cf | 600 | static void watchdog_nmi_disable(unsigned int cpu) |
58687acb DZ |
601 | { |
602 | struct perf_event *event = per_cpu(watchdog_ev, cpu); | |
603 | ||
604 | if (event) { | |
605 | perf_event_disable(event); | |
606 | per_cpu(watchdog_ev, cpu) = NULL; | |
607 | ||
608 | /* should be in cleanup, but blocks oprofile */ | |
609 | perf_event_release_kernel(event); | |
610 | } | |
df577149 UO |
611 | if (cpu == 0) { |
612 | /* watchdog_nmi_enable() expects this to be zero initially. */ | |
613 | cpu0_err = 0; | |
614 | } | |
58687acb | 615 | } |
b3738d29 SE |
616 | |
617 | void watchdog_nmi_enable_all(void) | |
618 | { | |
619 | int cpu; | |
620 | ||
ab992dc3 PZ |
621 | mutex_lock(&watchdog_proc_mutex); |
622 | ||
623 | if (!(watchdog_enabled & NMI_WATCHDOG_ENABLED)) | |
624 | goto unlock; | |
b3738d29 SE |
625 | |
626 | get_online_cpus(); | |
fe4ba3c3 | 627 | for_each_watchdog_cpu(cpu) |
b3738d29 SE |
628 | watchdog_nmi_enable(cpu); |
629 | put_online_cpus(); | |
ab992dc3 PZ |
630 | |
631 | unlock: | |
1173ff09 | 632 | mutex_unlock(&watchdog_proc_mutex); |
b3738d29 SE |
633 | } |
634 | ||
635 | void watchdog_nmi_disable_all(void) | |
636 | { | |
637 | int cpu; | |
638 | ||
ab992dc3 PZ |
639 | mutex_lock(&watchdog_proc_mutex); |
640 | ||
b3738d29 | 641 | if (!watchdog_running) |
ab992dc3 | 642 | goto unlock; |
b3738d29 SE |
643 | |
644 | get_online_cpus(); | |
fe4ba3c3 | 645 | for_each_watchdog_cpu(cpu) |
b3738d29 SE |
646 | watchdog_nmi_disable(cpu); |
647 | put_online_cpus(); | |
ab992dc3 PZ |
648 | |
649 | unlock: | |
650 | mutex_unlock(&watchdog_proc_mutex); | |
b3738d29 | 651 | } |
58687acb | 652 | #else |
bcd951cf TG |
653 | static int watchdog_nmi_enable(unsigned int cpu) { return 0; } |
654 | static void watchdog_nmi_disable(unsigned int cpu) { return; } | |
23637d47 | 655 | #endif /* CONFIG_HARDLOCKUP_DETECTOR */ |
58687acb | 656 | |
b8900bc0 FW |
657 | static struct smp_hotplug_thread watchdog_threads = { |
658 | .store = &softlockup_watchdog, | |
659 | .thread_should_run = watchdog_should_run, | |
660 | .thread_fn = watchdog, | |
661 | .thread_comm = "watchdog/%u", | |
662 | .setup = watchdog_enable, | |
663 | .cleanup = watchdog_cleanup, | |
664 | .park = watchdog_disable, | |
665 | .unpark = watchdog_enable, | |
666 | }; | |
667 | ||
81a4beef UO |
668 | /* |
669 | * park all watchdog threads that are specified in 'watchdog_cpumask' | |
670 | */ | |
671 | static int watchdog_park_threads(void) | |
672 | { | |
673 | int cpu, ret = 0; | |
674 | ||
675 | get_online_cpus(); | |
676 | for_each_watchdog_cpu(cpu) { | |
677 | ret = kthread_park(per_cpu(softlockup_watchdog, cpu)); | |
678 | if (ret) | |
679 | break; | |
680 | } | |
681 | if (ret) { | |
682 | for_each_watchdog_cpu(cpu) | |
683 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); | |
684 | } | |
685 | put_online_cpus(); | |
686 | ||
687 | return ret; | |
688 | } | |
689 | ||
690 | /* | |
691 | * unpark all watchdog threads that are specified in 'watchdog_cpumask' | |
692 | */ | |
693 | static void watchdog_unpark_threads(void) | |
694 | { | |
695 | int cpu; | |
696 | ||
697 | get_online_cpus(); | |
698 | for_each_watchdog_cpu(cpu) | |
699 | kthread_unpark(per_cpu(softlockup_watchdog, cpu)); | |
700 | put_online_cpus(); | |
701 | } | |
702 | ||
9809b18f MH |
703 | static void restart_watchdog_hrtimer(void *info) |
704 | { | |
f7f66b05 | 705 | struct hrtimer *hrtimer = raw_cpu_ptr(&watchdog_hrtimer); |
9809b18f MH |
706 | int ret; |
707 | ||
708 | /* | |
709 | * No need to cancel and restart hrtimer if it is currently executing | |
710 | * because it will reprogram itself with the new period now. | |
711 | * We should never see it unqueued here because we are running per-cpu | |
712 | * with interrupts disabled. | |
713 | */ | |
714 | ret = hrtimer_try_to_cancel(hrtimer); | |
715 | if (ret == 1) | |
716 | hrtimer_start(hrtimer, ns_to_ktime(sample_period), | |
717 | HRTIMER_MODE_REL_PINNED); | |
718 | } | |
719 | ||
b2f57c3a | 720 | static void update_watchdog(int cpu) |
9809b18f | 721 | { |
9809b18f MH |
722 | /* |
723 | * Make sure that perf event counter will adopt to a new | |
724 | * sampling period. Updating the sampling period directly would | |
725 | * be much nicer but we do not have an API for that now so | |
726 | * let's use a big hammer. | |
727 | * Hrtimer will adopt the new period on the next tick but this | |
728 | * might be late already so we have to restart the timer as well. | |
729 | */ | |
730 | watchdog_nmi_disable(cpu); | |
e0a23b06 | 731 | smp_call_function_single(cpu, restart_watchdog_hrtimer, NULL, 1); |
9809b18f MH |
732 | watchdog_nmi_enable(cpu); |
733 | } | |
734 | ||
b2f57c3a | 735 | static void update_watchdog_all_cpus(void) |
9809b18f MH |
736 | { |
737 | int cpu; | |
738 | ||
739 | get_online_cpus(); | |
fe4ba3c3 | 740 | for_each_watchdog_cpu(cpu) |
b2f57c3a | 741 | update_watchdog(cpu); |
9809b18f MH |
742 | put_online_cpus(); |
743 | } | |
744 | ||
b2f57c3a | 745 | static int watchdog_enable_all_cpus(void) |
58687acb | 746 | { |
b8900bc0 | 747 | int err = 0; |
58687acb | 748 | |
3c00ea82 | 749 | if (!watchdog_running) { |
230ec939 FW |
750 | err = smpboot_register_percpu_thread_cpumask(&watchdog_threads, |
751 | &watchdog_cpumask); | |
b8900bc0 FW |
752 | if (err) |
753 | pr_err("Failed to create watchdog threads, disabled\n"); | |
230ec939 | 754 | else |
3c00ea82 | 755 | watchdog_running = 1; |
b2f57c3a UO |
756 | } else { |
757 | /* | |
758 | * Enable/disable the lockup detectors or | |
759 | * change the sample period 'on the fly'. | |
760 | */ | |
761 | update_watchdog_all_cpus(); | |
bcd951cf | 762 | } |
b8900bc0 FW |
763 | |
764 | return err; | |
58687acb DZ |
765 | } |
766 | ||
b8900bc0 FW |
767 | /* prepare/enable/disable routines */ |
768 | /* sysctl functions */ | |
769 | #ifdef CONFIG_SYSCTL | |
58687acb DZ |
770 | static void watchdog_disable_all_cpus(void) |
771 | { | |
3c00ea82 FW |
772 | if (watchdog_running) { |
773 | watchdog_running = 0; | |
b8900bc0 | 774 | smpboot_unregister_percpu_thread(&watchdog_threads); |
bcd951cf | 775 | } |
58687acb DZ |
776 | } |
777 | ||
58687acb | 778 | /* |
a0c9cbb9 UO |
779 | * Update the run state of the lockup detectors. |
780 | */ | |
781 | static int proc_watchdog_update(void) | |
782 | { | |
783 | int err = 0; | |
784 | ||
785 | /* | |
786 | * Watchdog threads won't be started if they are already active. | |
787 | * The 'watchdog_running' variable in watchdog_*_all_cpus() takes | |
788 | * care of this. If those threads are already active, the sample | |
789 | * period will be updated and the lockup detectors will be enabled | |
790 | * or disabled 'on the fly'. | |
791 | */ | |
792 | if (watchdog_enabled && watchdog_thresh) | |
b2f57c3a | 793 | err = watchdog_enable_all_cpus(); |
a0c9cbb9 UO |
794 | else |
795 | watchdog_disable_all_cpus(); | |
796 | ||
797 | return err; | |
798 | ||
799 | } | |
800 | ||
ef246a21 UO |
801 | /* |
802 | * common function for watchdog, nmi_watchdog and soft_watchdog parameter | |
803 | * | |
804 | * caller | table->data points to | 'which' contains the flag(s) | |
805 | * -------------------|-----------------------|----------------------------- | |
806 | * proc_watchdog | watchdog_user_enabled | NMI_WATCHDOG_ENABLED or'ed | |
807 | * | | with SOFT_WATCHDOG_ENABLED | |
808 | * -------------------|-----------------------|----------------------------- | |
809 | * proc_nmi_watchdog | nmi_watchdog_enabled | NMI_WATCHDOG_ENABLED | |
810 | * -------------------|-----------------------|----------------------------- | |
811 | * proc_soft_watchdog | soft_watchdog_enabled | SOFT_WATCHDOG_ENABLED | |
812 | */ | |
813 | static int proc_watchdog_common(int which, struct ctl_table *table, int write, | |
814 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
815 | { | |
816 | int err, old, new; | |
817 | int *watchdog_param = (int *)table->data; | |
818 | ||
819 | mutex_lock(&watchdog_proc_mutex); | |
820 | ||
821 | /* | |
822 | * If the parameter is being read return the state of the corresponding | |
823 | * bit(s) in 'watchdog_enabled', else update 'watchdog_enabled' and the | |
824 | * run state of the lockup detectors. | |
825 | */ | |
826 | if (!write) { | |
827 | *watchdog_param = (watchdog_enabled & which) != 0; | |
828 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
829 | } else { | |
830 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); | |
831 | if (err) | |
832 | goto out; | |
833 | ||
834 | /* | |
835 | * There is a race window between fetching the current value | |
836 | * from 'watchdog_enabled' and storing the new value. During | |
837 | * this race window, watchdog_nmi_enable() can sneak in and | |
838 | * clear the NMI_WATCHDOG_ENABLED bit in 'watchdog_enabled'. | |
839 | * The 'cmpxchg' detects this race and the loop retries. | |
840 | */ | |
841 | do { | |
842 | old = watchdog_enabled; | |
843 | /* | |
844 | * If the parameter value is not zero set the | |
845 | * corresponding bit(s), else clear it(them). | |
846 | */ | |
847 | if (*watchdog_param) | |
848 | new = old | which; | |
849 | else | |
850 | new = old & ~which; | |
851 | } while (cmpxchg(&watchdog_enabled, old, new) != old); | |
852 | ||
853 | /* | |
854 | * Update the run state of the lockup detectors. | |
855 | * Restore 'watchdog_enabled' on failure. | |
856 | */ | |
857 | err = proc_watchdog_update(); | |
858 | if (err) | |
859 | watchdog_enabled = old; | |
860 | } | |
861 | out: | |
862 | mutex_unlock(&watchdog_proc_mutex); | |
863 | return err; | |
864 | } | |
865 | ||
83a80a39 UO |
866 | /* |
867 | * /proc/sys/kernel/watchdog | |
868 | */ | |
869 | int proc_watchdog(struct ctl_table *table, int write, | |
870 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
871 | { | |
872 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED|SOFT_WATCHDOG_ENABLED, | |
873 | table, write, buffer, lenp, ppos); | |
874 | } | |
875 | ||
876 | /* | |
877 | * /proc/sys/kernel/nmi_watchdog | |
58687acb | 878 | */ |
83a80a39 UO |
879 | int proc_nmi_watchdog(struct ctl_table *table, int write, |
880 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
881 | { | |
882 | return proc_watchdog_common(NMI_WATCHDOG_ENABLED, | |
883 | table, write, buffer, lenp, ppos); | |
884 | } | |
885 | ||
886 | /* | |
887 | * /proc/sys/kernel/soft_watchdog | |
888 | */ | |
889 | int proc_soft_watchdog(struct ctl_table *table, int write, | |
890 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
891 | { | |
892 | return proc_watchdog_common(SOFT_WATCHDOG_ENABLED, | |
893 | table, write, buffer, lenp, ppos); | |
894 | } | |
58687acb | 895 | |
83a80a39 UO |
896 | /* |
897 | * /proc/sys/kernel/watchdog_thresh | |
898 | */ | |
899 | int proc_watchdog_thresh(struct ctl_table *table, int write, | |
900 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
58687acb | 901 | { |
83a80a39 | 902 | int err, old; |
58687acb | 903 | |
359e6fab | 904 | mutex_lock(&watchdog_proc_mutex); |
bcd951cf | 905 | |
83a80a39 | 906 | old = ACCESS_ONCE(watchdog_thresh); |
b8900bc0 | 907 | err = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
83a80a39 | 908 | |
b8900bc0 | 909 | if (err || !write) |
359e6fab | 910 | goto out; |
e04ab2bc | 911 | |
b66a2356 | 912 | /* |
83a80a39 UO |
913 | * Update the sample period. |
914 | * Restore 'watchdog_thresh' on failure. | |
b66a2356 | 915 | */ |
83a80a39 UO |
916 | set_sample_period(); |
917 | err = proc_watchdog_update(); | |
918 | if (err) | |
919 | watchdog_thresh = old; | |
359e6fab MH |
920 | out: |
921 | mutex_unlock(&watchdog_proc_mutex); | |
b8900bc0 | 922 | return err; |
58687acb | 923 | } |
fe4ba3c3 CM |
924 | |
925 | /* | |
926 | * The cpumask is the mask of possible cpus that the watchdog can run | |
927 | * on, not the mask of cpus it is actually running on. This allows the | |
928 | * user to specify a mask that will include cpus that have not yet | |
929 | * been brought online, if desired. | |
930 | */ | |
931 | int proc_watchdog_cpumask(struct ctl_table *table, int write, | |
932 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
933 | { | |
934 | int err; | |
935 | ||
936 | mutex_lock(&watchdog_proc_mutex); | |
937 | err = proc_do_large_bitmap(table, write, buffer, lenp, ppos); | |
938 | if (!err && write) { | |
939 | /* Remove impossible cpus to keep sysctl output cleaner. */ | |
940 | cpumask_and(&watchdog_cpumask, &watchdog_cpumask, | |
941 | cpu_possible_mask); | |
942 | ||
943 | if (watchdog_running) { | |
944 | /* | |
945 | * Failure would be due to being unable to allocate | |
946 | * a temporary cpumask, so we are likely not in a | |
947 | * position to do much else to make things better. | |
948 | */ | |
949 | if (smpboot_update_cpumask_percpu_thread( | |
950 | &watchdog_threads, &watchdog_cpumask) != 0) | |
951 | pr_err("cpumask update failed\n"); | |
952 | } | |
953 | } | |
954 | mutex_unlock(&watchdog_proc_mutex); | |
955 | return err; | |
956 | } | |
957 | ||
58687acb DZ |
958 | #endif /* CONFIG_SYSCTL */ |
959 | ||
004417a6 | 960 | void __init lockup_detector_init(void) |
58687acb | 961 | { |
0f34c400 | 962 | set_sample_period(); |
b8900bc0 | 963 | |
fe4ba3c3 CM |
964 | #ifdef CONFIG_NO_HZ_FULL |
965 | if (tick_nohz_full_enabled()) { | |
314b08ff FW |
966 | pr_info("Disabling watchdog on nohz_full cores by default\n"); |
967 | cpumask_copy(&watchdog_cpumask, housekeeping_mask); | |
fe4ba3c3 CM |
968 | } else |
969 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
970 | #else | |
971 | cpumask_copy(&watchdog_cpumask, cpu_possible_mask); | |
972 | #endif | |
973 | ||
195daf66 | 974 | if (watchdog_enabled) |
b2f57c3a | 975 | watchdog_enable_all_cpus(); |
58687acb | 976 | } |